I scored 492 public CLAUDE.md files against a 12-rule baseline. Median: 3/12.

Short summary

An empirical audit of 492 public CLAUDE.md files reveals the median covers only 3 of 12 critical AI-agent behavior rules; biggest gaps are scoped-editing (98% missing), tool-call transparency (94%), and error-handling visibility (91%). Adding four one-sentence rules—scoped edits, tool summaries, error verbosity, and pre-code reading—moves median compliance from 3/12 to 7/12.

• •Median CLAUDE.md covers only 25% of 12 recommended AI-agent behavior rules
• •Four one-sentence additions close majority of gaps: scoped edits, tool summaries, error visibility, code-reading discipline
• •Zero production secrets leaked across 492 scanned files; CLAUDE.md is safe for open-source

Generated with AI, which can make mistakes.