Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Audit a skill against Pocock's write-a-skill spec and return a prioritized, diagnose-only fix report.
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-14 | ✗→✓ | ▲ Improved | 85% | 0% |
| case-15 | ✓→✗ | ▼ Worse | -71% | 0% |
| case-17 | ✓→✗ | ▼ Worse | -64% | 0% |
| case-20 | ✓→✗ | ▼ Worse | 95% | 0% |
| case-18 | ✓→✓ | = Same ✓ | 13% | 0% |
> Audits a skill against Matt Pocock's write-a-skill (the enforced Review Checklist) and > writing-great-skills (the deeper principles). Diagnose-first — the two gates below never move.
Refactoring touches only the target skill's own directory. Never other files, never user data.
SKILL.md in full, plus any linked references/.references/checklist.md — five sections: Invocation, Structure, Steps, Pruning, Hygiene.references/report-template.md (content only).Exception: user said with tags or include meta-params upfront → also fill references/meta-params-template.md in the report (propose only, no writes).
Invocation · Structure · Steps · Pruning · Hygiene. Full criteria live in references/checklist.md — score every row from that file, never from memory.
Name the smell when you find it: premature completion, duplication, sediment, sprawl, no-op, blind-guess. Tell + first fix for each: references/checklist.md § Failure-mode lookup.
Back up first: cp SKILL.md SKILL.md.bak-<date>. Then apply the report's prioritized fix list in order. Re-check the line count and grep the leading words to confirm nothing got lost in the move.
Run only on explicit request. Default timing: after content refactor. Upfront exception: with tags / include meta-params in the audit invocation → proposals in gate-1 report.
Propose via references/meta-params-template.md + references/optimize-tags.md. Nothing writes until the user consents — bulk (apply the tags) or cherry-pick (name the rows). Run each validation test before writing; skip or revert on request. Never mix tag writes into the content refactor unless the same message explicitly approves both.
Other measured skills in the registry, with their headline benchmark lift.