We analyzed 28,721 GitHub repos' AI instruction files across @claudeai , @OpenAI (Codex) , @GitHubCopilot , @cursor_ai , and @GeminiApp.
165,063 files. 3.3 million instructions.
66% are abstract wallpaper.
The most-copied community skills are the vaguest.
frontend-design: 271 repos, 2.8% specificity.
next-best-practices: 76 repos, 92.6% specificity.
Popularity and quality are inversely correlated.
"Use consistent formatting" โ in thousands of repos.
"Format with ruff format before committing" โ barely spreads.
The second one gets followed. The first one mostly doesn't.
The median instruction file: 50 content items, 12 actual directives.
73% of your CLAUDE.md is headings, context, and examples. Only 27% is doing what you think it does.
A bare restriction is an incomplete instruction.
Directive first (name the file, the path). Context second (why - without mentioning the prohibited concept). Restriction last.
Same words, different order: 25pp difference. The largest single variable I've measured.
https://t.co/TqXYDGDUFj