Skill Audit

Skill Audit

A skill library degrades the same way a prompt library does: things pile up, two skills quietly do the same job, and descriptions get vague enough that nothing triggers reliably. This skill audits the collection and produces a ruthless, actionable cleanup plan.

Step 1 — Get the inventory

Work from the best available source, in this order:

1. The skills visible in the current environment (if you can see an installed skills list, use it).
2. Otherwise, ask the user to paste their list from Settings → Capabilities → Skills — names are enough to start; descriptions make the audit much better.
3. If the user can share the actual SKILL.md files, audit those — that enables the deepest checks.

Never invent skills the user didn't confirm having.

Step 2 — Map each skill to its job

For every skill, write one line: the job it's hired for, in the form "when [situation], produce [output]". If you cannot write that line from the skill's name and description, that is itself a finding — the skill has a triggering problem.

Step 3 — Run the five checks

1. Overlap. Two or more skills hired for the same job (e.g. a /humanizer and an /anti-ai that both strip AI tells). Overlapping skills split triggering unpredictably — Claude may pick either one. Verdict: merge into one, or sharpen each description until the boundary is unambiguous (e.g. one runs on drafts, the other runs during prompt-writing).
2. Vague triggers. Descriptions like "helps with productivity" or "improves writing" — categories, not situations. These undertrigger (Claude skips them) or overtrigger (they fire on everything). Verdict: rewrite the description to name the exact user phrases and situations that should fire it.
3. Dead weight. Skills that duplicate what Claude does well with a one-line prompt (e.g. a skill that just says "summarize this"). A skill earns its slot by encoding a procedure, format, or standard Claude wouldn't apply by default. Verdict: delete, and note the one-line prompt that replaces it.
4. Scope creep. One skill doing three jobs ("writes posts, replies to comments, and plans content"). Multi-job skills trigger at the wrong times and dilute their instructions. Verdict: split, and name the resulting skills.
5. Gaps. Look at the user's actual recurring workflow (ask one question about it if unclear) and name at most 2-3 missing skills that would compound with the existing ones. Do not pad this list — a gap is only real if the user repeats that task weekly.

Step 4 — Deliver the audit

ALWAYS use this exact structure:

## Inventory
[N] skills audited.

## Findings
| Skill | Job it's hired for | Verdict | Why |
|-------|--------------------|---------|-----|
| /name | when X, produce Y  | ✅ Keep / 🔀 Merge / 🔧 Fix description / 🗑️ Delete / ✂️ Split | one line |

## Action plan (do in this order)
1. [Highest-impact action first — usually the merge or the worst description fix]
2. ...

## Rewritten descriptions
[For every 🔧: the ready-to-paste replacement description,
naming concrete trigger situations and user phrases.]

## Gaps worth filling (max 3)
- /suggested-name — [the recurring job it would do]

Judgment calls

• Be conservative with 🗑️. Deleting a skill the user loves destroys trust in the audit. If usage is unknown, mark it "🗑️ Delete (confirm: when did you last use this?)" instead of asserting.
• A healthy library is a valid result. If the collection is clean, say so, fix the one or two weakest descriptions, and stop. An audit that always finds problems is a horoscope.
• Merges must be concrete. Never just say "merge A and B" — state the merged skill's name, its one-line job, and which instructions survive from each parent.
• Respect the library's purpose. A content creator's library and a developer's library have different "dead weight" thresholds. Judge each skill against the user's actual work, not an abstract ideal.