Agent Skills: Audit Skills

Audit Skills

Audit a plugin folder for compliance with Claude Code skill and agent best practices.

Variables

PLUGIN_PATH: $ARGUMENTS

Workflow

If no PLUGIN_PATH is provided, STOP and ask the user to provide it.

Step 1: Discover Plugin Structure

1. Verify .claude-plugin/plugin.json exists and is valid JSON
2. Scan skills/ for all SKILL.md files
3. Scan agents/ for all agent definition files
4. Check for CLAUDE.md at plugin root
5. Build inventory of all components to audit

Step 2: Audit Each Skill

For every SKILL.md found, check each rule below. Track pass/fail per skill.

Frontmatter Rules

| Rule | Check | Severity | |------|-------|----------| | F1: name exists | name field present in frontmatter | ERROR | | F2: name format | kebab-case only (lowercase, hyphens, numbers), ≤64 chars | ERROR | | F3: name matches directory | name field matches the containing directory name | WARN | | F4: description exists | description field present | ERROR | | F5: description quality | ≥50 chars, starts with action verb, contains "Use when" or similar trigger phrase | WARN | | F6: allowed-tools explicit | allowed-tools field is present | WARN | | F7: allowed-tools minimal | No tools listed that the skill doesn't need (check skill body for tool usage signals) | INFO | | F8: argument-hint present | If skill uses $ARGUMENTS or $N, argument-hint should exist | WARN | | F9: context:fork has task | If context: fork, skill body contains actionable instructions (not just reference) | ERROR | | F10: context:fork has agent | If context: fork, agent field is specified | WARN | | F11: side-effects guarded | Plugin skills should have disable-model-invocation: false (true hides them from slash commands when installed). Only project-local .claude/skills/ may use true. | WARN | | F12: no unknown fields | Frontmatter only contains recognized fields | INFO |

Content Rules

| Rule | Check | Severity | |------|-------|----------| | C1: length | SKILL.md ≤500 lines. If longer, reference material should be in supporting files | WARN | | C2: supporting files referenced | If references/, templates/, scripts/ exist, they're referenced from SKILL.md | WARN | | C3: paths use CLAUDE_SKILL_DIR | References to skill-local files use ${CLAUDE_SKILL_DIR}, not relative paths | WARN | | C4: scripts executable | Any scripts in scripts/ have shebangs and are executable | ERROR |

Step 3: Audit Each Agent

For every agent definition file, check:

| Rule | Check | Severity | |------|-------|----------| | A1: name exists | name field present | ERROR | | A2: name format | kebab-case | ERROR | | A3: description exists | description field present | ERROR | | A4: description quality | Explains when Claude should delegate to this agent | WARN | | A5: tools explicit | tools field is present (principle of least privilege) | WARN | | A6: model specified | model field present (haiku/sonnet/opus) | INFO | | A7: system prompt exists | Markdown body after frontmatter has clear role + instructions | WARN |

Step 4: Audit Plugin Structure

| Rule | Check | Severity | |------|-------|----------| | P1: plugin.json valid | .claude-plugin/plugin.json exists and parses as valid JSON | ERROR | | P2: plugin.json has name | name field in plugin.json is kebab-case | ERROR | | P3: plugin.json has version | version field follows semver | WARN | | P4: plugin.json has description | description field exists | WARN | | P5: no nested components | No skills/ or agents/ inside .claude-plugin/ | ERROR | | P6: CLAUDE.md exists | Plugin root has CLAUDE.md | WARN | | P7: naming consistency | All skill/agent names use kebab-case throughout | ERROR |

Step 5: Write Analysis Report

Write the full audit report to {PLUGIN_PATH}/analysis.md. This file is the persistent artifact of the audit — it stays in the plugin folder for reference and tracking over time.

Use this structure for the file:

# Plugin Audit: {plugin-name}

**Path**: {PLUGIN_PATH}
**Date**: {today}
**Skills**: {count} | **Agents**: {count}

## Summary

| Severity | Count |
|----------|-------|
| ERROR    | {n}   |
| WARN     | {n}   |
| INFO     | {n}   |
| PASS     | {n}   |

## Plugin Structure
- P1: {PASS/FAIL} — plugin.json valid
- P2: {PASS/FAIL} — plugin.json has name
...

## Skills

### {skill-name}
- F1: {PASS/FAIL} — name exists
- F2: {PASS/FAIL} — name format
...

### {skill-name}
...

## Agents

### {agent-name}
- A1: {PASS/FAIL} — name exists
...

## Top Issues (by impact)

1. {Most impactful issue + suggested fix}
2. {Next issue + suggested fix}
3. ...

## Suggested Fixes

{For each ERROR and WARN, a specific actionable fix}

Step 6: Summarize to User

After writing the file, print a brief summary to the user:

• Total pass/fail counts
• Number of ERRORs and WARNs
• Path to the written analysis.md
• Top 3 highest-impact issues to fix first

Do NOT write fixes to skill/agent files. Only report them in analysis.md. The user decides what to apply.

Recognized Frontmatter Fields

Skills: name, description, argument-hint, allowed-tools, disable-model-invocation, user-invocable, context, agent, model, hooks

Agents: name, description, tools, disallowedTools, model, permissionMode, maxTurns, skills, mcpServers, memory, hooks, color, background, isolation

Description Quality Signals

Good signals (any of these = likely good):

• Starts with action verb: Analyze, Audit, Generate, Review, Research, Explain, Build, Deploy, Convert
• Contains trigger phrase: "Use when", "Use for", "Use after", "Use proactively"
• Mentions specific domain/context

Bad signals (any of these = likely bad):

• Under 30 characters
• Generic words only: "helper", "tool", "processor", "thing"
• No use-case context
• Just restates the name