Guides built on test data · page 3 of 6
The SkillProof blog
Claude Skills vs OpenAI Assistants API: A Practical Comparison
A practical comparison of Claude Skills and the OpenAI Assistants API. We analyze portability, cost, and lock-in to help you choose the right tool.
codebase-memory-mcp: 127% More Tokens, Not 99% Fewer
We A/B tested codebase-memory-mcp vs plain grep. 8/8 correctness both arms, but the graph cost 172,319 tokens vs 75,817 — 127% more, not 99% fewer.
We Audited Google's zx With and Without a Review Skill
A real code audit of google/zx's core.ts, run twice: free-roaming vs. a checklist skill. Same tokens table, same P0-P3 findings, one surprising verdict.
The Skill Bench: systematic-debugging vs. baseline
We planted 3 bugs in a Snake game and tested Claude with and without the systematic-debugging skill. The baseline found a bug we never planted.
The Skill Bench, part 2: Claude TDD skill vs no skill
We ran Claude Sonnet twice on the same aiogram Telegram bot, once with the TDD skill installed and once without, and measured tokens, tests, and honesty.
We Built the Same Landing Page With and Without a Skill
The frontend-design skill vs. the naked model, same landing page brief. Real token counts, real output, honest verdict on when +98% tokens is worth it.
Claude as an SEO Translator: 4,500 Pages, 10 Languages
How we localized skillproof.dev into 10 languages with Claude agents: the six rules that mattered, real token costs, and what broke along the way.
token-discipline: an Honest Benchmark of Token Savings
9 enforceable rules for input-side token savings. Five controlled task pairs, real telemetry: −20% on multi-step work, a wash on one-shots. Free, MIT.
We Cut Claude Research Hallucinations 59% With One Skill
12 controlled research runs, every claim judged against live primary sources: 13% wrong baseline vs 5.4% with the research-discipline skill. Free, MIT.
Same Question, Twice: a Live Before/After of AI Research
We asked Claude to name the top-5 skill repos twice — bare, then with research-discipline installed. Baseline missed 3 of 5. Every number verified live.
The Claude Skill That Audits Its Own OpenAPI Spec
REST/OpenAPI design for Claude that validates its own output: 0 validator errors vs 18 baseline, 0 consistency violations vs 6, benchmarked. Free, MIT.
Changelogs You Can Trace to Real Commits — Benchmarked
We tested a changelog skill on 4 real OSS releases: 87% of lines traceable to commits vs 0%, breaking changes always first, format 30/36 vs 13/36. Free.
The Commit Message That Matches the Diff, Benchmarked
15 real diffs, two Claude agents each: diff coverage jumps 83%→97%, subjects fit 50 chars 15/15, zero fabrications. Free, MIT — with the losses shown.
Emails People Actually Reply To: One Goal, No AI Fluff
12 blind email tests, same model: 9/12 got a 'reply now' with the skill vs 3/12 without. One goal, hard word budgets, a fluff scanner. Free, MIT.
The Prompt Fixer That Diagnoses Before It Rewrites
A Claude prompt optimizer that names the failure first, changes the minimum, and ships an A/B plan — benchmarked, honest about a tiny sample. Free, MIT.