Self-improving skill-hygiene checks

A bash block at the bottom of SKILL.md that runs a handful of grep-based hygiene checks against the data the skill operates on. Each check is a real anti-pattern with a real date. Examples from a small-team HR skill: (1) candidates are flat files, not in subfolders; (2) candidate files have cv: in frontmatter; (3) candidates with status: intro older than 14 days surface (forces a decision); (4) role cards have status: and 1-1 History sections; (5) compensation strings never leak outside the HR folder. The rule: if eval finds a violation, fix the file; if eval fails to find something it should have, add a new check.

Skills accumulate cruft. New people break the conventions. The eval block is the immune system — runs cheap, surfaces specific violations, and self-extends from real failures. A small-team HR skill caught the per-candidate-folders anti-pattern on a specific date and added a check the same day; the bug cannot recur.

1. At the bottom of SKILL.md, add a fenced bash block that sets VAULT= and SKILL_HOME= paths.
2. Write 3–5 grep checks for your top anti-patterns. Each check echoes "FAIL: <description>" on hit; exits 0 always (so checks compose).
3. Run on demand: extract the block with awk, pipe to bash:

awk '/^```bash$/{flag=1;next}/^```$/{flag=0}flag' SKILL.md | bash

4. Schedule weekly via cron or a calendar reminder.
5. On every real failure, ask: would a grep check have caught this? If yes, add it. If no, file it as a lesson.
6. Whitelist names / paths inline in the bash where needed (team members' folder names, valid statuses).

• stat -f %m is BSD/macOS; Linux needs stat -c %Y. Pick one and document it. Whitelists must update when team membership changes.

• Generalize beyond grep — wrap any skill with versioning + a self-improvement step that runs after each invocation. The same eval-loop shape applies to any skill domain where "after a change, run this audit prompt" is the right reflex.