SkillProof tested Claude skills

Data report · updated Aug 10, 2026

Do Claude skills actually work?

Everyone lists Claude skills. Nobody measures whether they help. So we installed 2172 of them and ran each on a real task — once with the skill, once with plain Claude — and scored the difference. Here's where skills earn their keep, and where they're theatre.

79%deliver a real lift over baseline Claude
15%are placebo — no measurable difference
141made Claude's output worse than no skill

Where skills help — and where they don't

Share of tested skills in each category that beat plain Claude clearly (output ≥8/10). The single most useful finding: a skill's value depends heavily on the job.

Writing 74%111/150
Documents 73%113/154
Sales 71%22/31
Design 67%128/191
Crypto 61%14/23
Coding 58%392/674
Data 55%90/163
Marketing 55%54/99
Testing 53%63/118
SEO 51%18/35
Productivity 45%208/462
Efficiency 42%30/72

Take: reach for a skill on documents, writing, and data work — most deliver. For general "productivity", most are placebo; plain Claude does about as well.

Of 2172 skills, how many actually reach useful?

120 stumble on install — hardcoded paths, broken clone commands
23 never load — a vague description: so Claude ignores them
323 load fine but change nothing — placebo
1706 genuinely improve the output — the 79% worth installing

The #1 killer isn't a bad skill body — it's a mushy trigger description. You install it, nothing happens, you conclude skills are snake oil. Often the skill was fine; it just never fired.

The lift, skill by skill

Output-vs-baseline score for every tested skill. 5 = no difference. Below 5 = worse than plain Claude.

10 34
9 527
8 682
7 398
6 158
5 232
4 58
3 47
2 17
1 6
0 13

The 10 highest-lift skills we've tested

Every one measurably beat baseline Claude on a real task. Free to install.

  1. Statistical Analysis Data ★ 10.0 Enforces frame-inspect-check-assumptions-effect-size-APA-report pipeline for hypothesis tests; bundled Shapiro-Wilk/Levene's script actually runs.
  2. Pinme Auth Coding ★ 10.0 TypeScript reference for PinMe Worker auth: create/verify/list Identity Platform users.
  3. Code Health Check Coding ★ 10.0 Read-only, sub-agent-forked code health scanner with a fixed severity-graded report format.
  4. Fetch URL As Markdown Productivity ★ 10.0 Local trafilatura-first URL-to-Markdown fetcher with a clean exit-code contract for when to fall back to Exa.
  5. CocoIndex Data ★ 10.0 Turns vague 'build me a RAG pipeline' requests into working CocoIndex flow code with the real API syntax.
  6. Refresh Context Map Productivity ★ 10.0 Rebuilds the AIDEV-anchor/AGENTS.md/ADR manifest that powers claude-leverage's PreToolUse context hook.
  7. S&box Coding ★ 10.0 Blocks Unity muscle memory when writing s&box (Source 2 / Facepunch) C# components, Razor UI, and networking.
  8. Crawl4AI Data ★ 10.0 Wraps the Crawl4AI browser scraper for JS-heavy sites, batch crawls, and LLM-free schema extraction.
  9. Coding Agent Sessions Coding ★ 10.0 Finds and reconstructs real coding-agent session transcripts across Claude, Codex, OpenCode, and more.
  10. Codex GPT Image Design ★ 10.0 Generates gpt-image-2 images via local Codex OAuth, no OPENAI_API_KEY needed.

How we measured this

Each skill: fresh install following its own README, a trigger battery (3 phrasings that should fire, 2 that shouldn't), then one representative task run twice — baseline Claude vs Claude following the skill — scored install /5, trigger /5, output vs baseline /10, docs /5. The output score above is that third axis. Full protocol on the methodology page.

Honest caveats: this is 2172 skills — the top slice by stars and validator score, not a random sample (the long tail is likely worse, not better). The output axis involves judgment; we publish the baseline and skill artifacts on each page so you can check us, including the skills that lost.

Browse the 1706 that work

Every verdict, score and install command — free, no signup.

Browse all skills — free Get a tested pack — $10