| name | audit-skill |
| description | Audit a SKILL.md against loading, discoverability, effectiveness, and style rules, then produce a severity-grouped report. |
| disable-model-invocation | true |
Audit Skill
Produce a severity-grouped report listing every rule violation found in the skill's SKILL.md and folder layout, with file, line, and quoted evidence for each finding.
Process
- Read the skill's SKILL.md and list its folder tree.
- Run the checklist below – every rule against the entire skill, each yielding a pass or a violation with evidence: file, line number, and the actual value or quoted text.
- Report the findings using this template.
## Blockers
-
Discoverability
Effectiveness
Quality
Summary
N blockers, N discoverability, N effectiveness, N quality.
Omit any severity heading with zero findings. If nothing is flagged, report Skill passes audit. followed by the summary line.
Checklist
Blockers
Skill won't load, won't trigger, or will malfunction.
name is 1–64 characters, lowercase a-z / digits / hyphens only, with no leading hyphen, no trailing hyphen, and no consecutive hyphens.
name does not contain claude or anthropic anywhere (reserved).
- Frontmatter is wrapped in matching
--- delimiters and contains no XML angle brackets (< or >).
description is under 1024 characters.
SKILL.md exists at the skill root with the exact case-sensitive filename.
Discoverability
Skill loads but won't trigger reliably, or triggers on wrong requests.
description names the capability in the first 80 characters.
description includes Use when... with concrete trigger verbs a user would say – model-invoked skills only; a disable-model-invocation skill strips triggers to a one-line human-facing summary.
description includes a Don't use for... exclusion only when a sibling skill could plausibly be confused for this one; omit it when there's no real overlap, rather than pad with a contrived one.
description uses third person (the user); the body is imperative or second person. Flag first person anywhere and second person in the description.
- Each
Use when... trigger names a distinct case; flag synonym triggers that restate one.
- Trigger guidance lives only in the
description; flag When to use... content in the body – it loads after the skill has already fired.
- Skill name is specific, not vague (
helper, utils, tool, agent, skill).
Effectiveness
Skill loads and triggers but steers the agent weakly.
- Every line changes behaviour versus the model's default; flag no-op instructions (
be thorough, write clean code, use best practices) and exposition the model already knows.
- Every prohibition names the replacement behaviour; flag a bare
don't X without a do Y instead.
- Alternatives are given as one default plus an escape hatch; flag a menu of interchangeable options with no stated default.
- Each rule or fact lives in exactly one place; flag the same meaning restated across sections.
- Completion criteria are checkable (
every modified file accounted for); flag vague bounds (understand the code, review thoroughly).
- A concept's definition, rules, and caveats sit under one heading; flag one concept scattered across the file.
SKILL.md inlines what every invocation path needs; material only some paths reach lives in a sibling file behind a pointer that states when to load it.
Quality
Skill works but violates style or structure rules.
SKILL.md is under 500 lines. Move detail into named sibling files if longer.
name matches the parent directory exactly.
- Folder layout: only
scripts/, references/, assets/, and agents/ appear as subdirectories, each one level deep; no README.md, CHANGELOG.md, or human-facing documentation at the skill root.
- Sibling files link directly from
SKILL.md – no reference-to-reference chains – and a reference file over 100 lines opens with a table of contents.
- Instructions are declarative: flag compliance hedging (
might want to, aim for, try to), callout boxes (> **Note:**), closing summaries, preamble sections (## Overview, ## Introduction, ## Background), and self-confirmation prompts (confirm you understand). Calibrated facts (logs are usually wrong) and imperative Consider X pass.
- Terminology is consistent: one term per concept across the skill, with no drift between synonyms.
- Sequences are explicitly ordered (numbered list or phase headings) with decision branches (
if X → Y; else → Z); rules are bulleted unless count or order is load-bearing (all three must hold).