Review Skill

Review Skill - Multi-Agent Skill Auditor

Runs 3 specialized review agents in parallel against a skill and all its associated components (hooks, agents, helpers, tests).

Arguments

• /review-skill {skill-name} — Review a specific skill (e.g., task-planner)
• /review-skill {skill-name} --structure-only — Only run Agent A (structural review)
• /review-skill {skill-name} --content-only — Only run Agent B (content quality)
• /review-skill {skill-name} --hooks-only — Only run Agent C (hooks best practices)

Workflow

1. Resolve Skill Name

Parse the argument to get the skill name. If no argument provided, ask the user which skill to review.

2. Discover All Related Files

Search these locations for files belonging to the skill:

~/.claude/skills/{skill-name}/               → SKILL.md, templates.md, etc.
~/.claude/agents/                            → agents referenced by skill
~/.claude/hooks/PreToolUse/                  → PreToolUse hooks
~/.claude/hooks/PostToolUse/                 → PostToolUse hooks
~/.claude/hooks/SubagentStart/               → SubagentStart hooks
~/.claude/hooks/SubagentStop/                → SubagentStop hooks
~/.claude/hooks/Stop/                        → Stop hooks
~/.claude/hooks/helpers/                     → helper scripts
~/.claude/tests/                             → test files
~/.claude/settings.json                      → hook registrations

Discovery strategy:

1. Read ~/.claude/skills/{skill-name}/SKILL.md — this is the primary file
2. Glob ~/.claude/skills/{skill-name}/**/* — all skill files (recursive)
3. Read ~/.claude/settings.json — find all registered hooks
4. Grep hook scripts for references to the skill name or its state files
5. Grep agent files for references to the skill name
6. Check ~/.claude/tests/ for test files mentioning the skill

Build a file inventory — a categorized list of all discovered files with full paths.

3. Spawn Review Agents in Parallel

Launch all three agents in a single message (parallel execution):

Agent A: Structural Review (plugin-dev:skill-reviewer)

Review the {skill-name} skill plugin and ALL its associated hooks, agents, and helpers for structural quality and best practices.

Files to review:
{file inventory}

Review for:
1. Plugin structure and conventions
2. Hook registration correctness
3. Agent definition quality
4. Skill frontmatter and content structure
5. Shell script quality and error handling
6. Inter-component consistency

Agent B: Content Quality (skill-content-reviewer)

Review the content quality of the {skill-name} skill and all its associated components.

Files to review:
{file inventory}

Evaluate:
1. Is the guidance comprehensive and actionable?
2. Are hook scripts doing the right things?
3. Are there gaps in coverage (missing edge cases, hooks)?
4. Is the content accurate — will it guide Claude Code correctly?
5. Are templates well-structured?
6. Do agent prompts give enough context?

Agent C: Hooks Best Practices (claude-code-guide)

I need expert guidance on Claude Code hooks best practices to evaluate the {skill-name} plugin.

This plugin uses these hooks:
{list hook events and scripts from file inventory}

Please verify:
1. Hook event usage — are the right events used for each purpose?
2. Hook registration format — is settings.json correct?
3. Input/output contracts — exit codes, stdin JSON, stderr usage
4. State management patterns — atomic writes, locking, file paths
5. Known limitations that apply to this plugin's hook usage
6. Any anti-patterns or gotchas in the hook implementations

4. Synthesize Results

After agents complete, present a unified summary. If an agent fails or returns empty output, note the failure and synthesize from the agents that succeeded — do not retry or block on the failed agent.

## Skill Review: {skill-name}

### Structural Review
{Agent A key findings — critical/major/minor counts}

### Content Quality
{Agent B key findings — gaps, strengths, accuracy}

### Hooks Compliance
{Agent C key findings — contract violations, anti-patterns}

### Priority Fixes
1. {most critical finding across all agents}
2. ...

Constraints

• MUST spawn all 3 agents in a single message (parallel)
• MUST include full file paths in each agent prompt
• MUST present unified summary after all complete
• If --structure-only, --content-only, or --hooks-only, spawn only the relevant agent (A, B, or C respectively)
• If an agent fails or returns empty, synthesize from available results and note the gap
• If the target skill has no hooks, skip Agent C (hooks review) unless explicitly requested
• If skill not found, list available skills and ask user to pick one