Claude MD Auditor

A bad CLAUDE.md does worse than nothing — it creates false confidence ("I set up my config") while Claude quietly ignores vague, contradictory, or impossible instructions. This skill audits the file against a five-dimension rubric (Specificity, Structure, Communication Style, Technical Accuracy, Completeness, two points each) and gives you a real 0-10 score with a one-line verdict, not a "could be improved" non-answer.

The diagnosis is line-level. Every flagged issue quotes the exact problematic text, explains what goes wrong in practice, and gives the exact replacement. No "make this more specific" — you get the rewrite. It also checks for the eight anti-patterns that consistently sink CLAUDE.md files: the Wish List (aspirational instructions that aren't testable), the Novel (over 200 lines where Claude's attention fades), the Contradiction (two rules that fight each other), the Generic Template (copy-pasted with no project-specific content), the Impossible (asking for things Claude can't do), Stale References (hardcoded paths that have drifted), Per-Person Rules in team files, and Missing Negative Instructions to override Claude's defaults.

What it deliberately won't do: manufacture issues to justify the audit. If your file genuinely scores 8 or higher, you're told that and given one or two polish suggestions instead of an invented critique. If your file is nearly empty, it doesn't pretend to audit — it tells you there's not enough there to score and switches to generation. The honesty is the point.

The output ends with a prioritised fix list, ranked by impact rather than by line order. If there are fifteen issues, you find out which three to fix first — and an estimated score after those fixes land, so you can decide if the rest is worth the effort.