5-Minute CLAUDE.md Self-Audit (7 checks from 800-hour runs)

[yurukusa]

Most "the AI ignored my CLAUDE.md" complaints I've triaged trace back to the CLAUDE.md itself, not to the model. After 800+ hours running Claude Code autonomously, I wrote up a 7-check self-audit you can run in ~5 minutes.

• Token weight (is your CLAUDE.md paying rent every turn?)
• Vague-rule detector (careful, appropriate, as needed — all ignored)
• Redundancy scan
• Stale-reference hunt (broken paths in CLAUDE.md = silent contradictions)
• Version-pin staleness
• User vs. project source-of-truth collisions
• Testable-assertion ratio

Each check is a single command you can paste. No tool install, no dependency.

👉 Gist: https://gist.github.com/yurukusa/df29f506af33368b03b1c5aeae85f04c

Feedback welcome — if one of the checks catches something on your setup, drop the Check # in a reply and I'll fold your case into the next revision.

---

The 4-layer token-cost model that Check 1 is based on is in Chapter 1 of the Token Book (free): https://zenn.dev/yurukusa/books/claude-code-token-book