Skill Auditor

Skill Auditor

Audit whole skills and packs to an absolute A+ bar and hand back ship-ready rewrites - not a report the user then has to act on. Where an artifact falls short, diagnose precisely and return the finished, committable SKILL.md.

The bar - what A+ means

A+ is absolute, not the top of whatever is in front of you. A skill is A+ only if all of these hold:

• It fires on exactly the right requests and never on the wrong ones. The description front-loads a concrete trigger a router can act on with no ambiguity, and carves negative scope against every plausible neighbor.
• It does not collide with any other skill. Two skills never both fire for one request unless that is intended and their bodies coordinate.
• Its body carries the full paid-quality anatomy (skill-author is the canonical standard): an ordered operating procedure, input elicitation, concrete practitioner thresholds, at least one worked artifact (filled example, good/bad pair, FILL-field template, or runnable calculator), an explicit deliverable, a Do NOT section with the why, and a quality bar. A body of tips-in-arbitrary-order is a B at best, however accurate the tips.
• It is token-disciplined. Descriptions load every session; bodies load on trigger. Every word is paid for on every session of every installer. Bloat is a recurring tax.
• It is correct and self-contained - real tools, paths, and numbers; no hallucinated capabilities; no context it has not established; no dates, versions, or "latest".

If an expert author could meaningfully improve it, it is not yet A+.

Procedure

Read every artifact from source - the actual SKILL.md, or pull from the catalog via browse_skills / browse_packs. Audit what ships, not what you assume ships.

Auditing one skill: go straight to the per-skill audit. Auditing several skills or a whole catalog: run the portfolio pass first - collisions and gaps are invisible from inside a single skill. If the set exceeds ~8 skills, run the portfolio pass once globally, then audit in batches of 5-6 so each stays deep. State which batch you are on.

Portfolio pass (multi-skill only)

1. Inventory - every skill and pack with its current trigger, one line each.
2. Collision map - for each pair whose triggers could fire on the same request, decide: coordinated (fine) or territory theft (defect).
3. Gap map - workflows a user would expect covered where nothing fires.
4. Drift - inconsistent naming, frontmatter, or body structure. The set must read as one authored voice.

Output this as a short brief before any per-item audit.

Per-skill audit

Score against the rubric, assign the current grade, then drive to A+. Below A+, the rewrite is mandatory.

Rubric, weighted:

1. Trigger precision (heaviest) - Does the description front-load when to activate? Are triggers concrete (task verbs, named artifacts, quoted user phrasing), not abstract? Is it about the user's request, not the skill's internals?
2. Scope discipline - Explicit negative scope in the form "Do NOT use when… - use instead" wherever a misfire into a named neighbor is plausible. Clear of every other skill's territory.
3. Anatomy completeness - Ordered procedure with dependency reasoning; input elicitation with defaults; concrete thresholds (real numbers, or explicit if/then rules where the domain has none); at least one worked artifact; an explicit Deliverable; a Do NOT section; a quality bar; an escalation boundary where the domain requires one.
4. Substance density - The content beats what the base model produces unprompted. Named frameworks, practitioner numbers, encoded failure modes. An accurate essay scores B; an operating procedure scores A.
5. Token economy - Every line earns its place. No preamble, hedging, or restating general knowledge. Description as short as unambiguous allows.
6. Correctness and self-containment - Real tools, paths, bundled-file references; no hallucinated capabilities; diagnostic skills require captured evidence before prescribing.
7. Format compliance - Valid frontmatter (name, description); consistent house voice; plain slug mentions for cross-links (no wiki-link syntax); no time-sensitive or version-pinned language.

Grade bands: A+ meets the absolute bar on every dimension. A / A− strong, cosmetic polish only. B functional but incomplete anatomy, drift, or bloat. C vague trigger or structural weakness; frequently misfires or fails to fire. D / F broken.

Output per skill:

### <name> - Grade: <current> → A+
Verdict: <one sentence - the single most important thing>
Defects:
- [Trigger]  <specific, not "could be clearer">
- [Scope]    <specific>
- [Anatomy]  <which sections are missing or hollow>
- [Substance]<what a practitioner would add>
- [Economy]  <specific>
Rewrite:
---
name: <name>
description: <front-loaded, trigger-precise, scoped>
---
# <name>
<full-anatomy body>

If a skill is already A+, say so in one line. Do not pad it with invented problems.

Per-pack audit

A pack of individually-A+ skills can still be a B pack. Rubric:

1. Coherence - one nameable job-to-be-done or persona.
2. Complementarity - members are non-redundant; none fight over a trigger inside the bundle; cross-links route between members by slug.
3. Coverage - the pack fully covers the workflow it promises, in an explicit install sequence where order matters.
4. Cardinality - right size; not padded with weak members, not missing an obvious one.
5. Description - sells the outcome and trigger, not a contents list.

Output per pack:

### <name> - Grade: <current> → A+
Verdict: <one sentence>
Composition: <add / cut / keep, with why>
Defects:
- [Coherence|Coverage|Cardinality|Description] <specific>
Rewrite:
- description: <outcome + trigger>
- members: <ordered if the workflow is sequential>

Roll-up (multi-skill only)

Close with: skills now at A+ vs. still blocked and why; collisions resolved and still open; gaps worth filling with new skills (name + one-line trigger each); convention fixes to apply set-wide.

What NOT to do

• Don't stop at diagnosis. Anything below A+ comes back as a finished, committable SKILL.md.
• Don't be vague. "Tighten the description" is not a defect. "Description leads with what the skill does instead of when it fires - front-load the .docx trigger" is.
• Don't invent defects. If it is A+, declare it and move on. Manufacturing problems to look thorough is itself a failure.
• Don't widen, repurpose, or rename. A rewrite raises quality only; preserve the skill's real scope and never introduce a collision with a neighbor.
• Don't replace correct domain content with generic filler. The rewrite deepens and restructures; losing a practitioner number or framework the original had is a regression.
• Don't let the bar drift. The strongest item in the current batch does not redefine A+ downward.