Claude md audit

Audit this repository's agent rules file (CLAUDE.md, or AGENTS.md if that's what exists; the user may pass an explicit path in arguments). If neither exists, say so and suggest /claude-md-new, then stop.

Read the rules file AND spot-check it against reality: open the manifests/CI configs it references (or should reference) and verify the commands it lists actually exist. A rules file that lies is worse than none. Grading standards live in ${CLAUDE_PLUGIN_ROOT}/FIELD-GUIDE.md — read it before scoring.

Rubric — score each, then sum (0–100)

1. Commands (0–25). Exact run / test / single-test / lint / build commands present and REAL (verified against package.json scripts, Makefile, CI, etc.). Deduct for missing single-test invocation, wrong or invented commands.
2. Map of the repo (0–20). Says where things live and where new code should go — specific paths, not descriptions. Deduct for anything an agent could only learn by asking.
3. Non-obvious conventions (0–20). The rules a new hire would have to be told: invariants, footguns, "we do X here, not Y", what NOT to touch. Deduct for generic advice ("write tests", "follow best practices") — each wish-list line is negative signal.
4. Verification loop (0–15). Tells the agent how to prove a change works (run tests and read output, curl the endpoint, screenshot) — not just typecheck.
5. Terseness & altitude (0–10). Imperative bullets, no prose padding, ~30–80 lines. Deduct for paragraphs of philosophy or duplicated README content.
6. Freshness (0–10). References match the current tree (paths exist, scripts still in package.json, stack versions not stale).

Output format

• Score: N/100 with the six sub-scores on one line each.
• Worst-first fixes: numbered list; each item = the problem, why it costs agent performance, and the concrete replacement text (write the actual lines, ready to paste).
• Verified-command check: table of every command in the file → found-in / not-found.
• Offer to apply the fixes directly if the user wants.