Marketplace Review

Arguments: [plugin-name] [--all] [--fix]. Wherever <arguments> appears below, substitute the text the user typed after the skill name.

Marketplace AI Review

Perform an AI-driven quality evaluation of plugin content across any Claude Code plugin marketplace.

What this evaluates

This review goes beyond structural validation (which /marketplace-health handles) and analyzes the semantic quality of every plugin component using AI judgment.

Procedure

Step 1: Load marketplace data

Read .claude-plugin/marketplace.json at the project root to get the full plugin registry.

Step 2: For each plugin (or the specified one), evaluate these dimensions

A. Plugin Description Quality (marketplace.json)

Evaluate the description field:

• Clarity: Is it immediately clear what the plugin does?
• Trigger coverage: Does it mention enough trigger scenarios for Claude to auto-invoke?
• Specificity: Does it name concrete actions, not just abstract categories?
• Length: Under 1024 chars, ideally under 300 for auto-load discoverability
• Keyword coverage: Keywords array contains the concrete domain terms users will search for

B. Agent Quality (for each agent in the plugin)

Read the agent .md file:

• Description directive voice: Uses "ALWAYS invoke", "You MUST", "Use PROACTIVELY" or similar (vs passive "Helps with", "Can be used for")
• TRIGGER WHEN clause: present, specific, uses concrete verbs / domain terms
• DO NOT TRIGGER WHEN clause: present only where a confusable sibling exists, and naming it (which other agent should handle this case instead). A clause that names none is context cost with no routing value: flag it for deletion
• Body structure: clear sections (ROLE, CAPABILITIES, CONVENTIONS, OUTPUT FORMAT or similar)
• Body size: 60-200 lines for simple agents, up to ~560 lines for complex; flag over 700 lines
• Tool restrictions: minimal but sufficient (e.g., a read-only reviewer should NOT have Write/Edit)
• Model choice: justified if not opus
• Color consistency: matches siblings in the same plugin where appropriate

C. Skill Quality (for each skill in the plugin)

Read the skill SKILL.md:

• Description: same directive voice + TRIGGER WHEN / DO NOT TRIGGER WHEN checks as agents
• Body size: under 500 lines; larger content should go in references/
• Progressive disclosure: uses references/ for deep content, keeps SKILL.md scannable
• Action-oriented: teaches what Claude doesn't already know; avoids restating general best practices
• Examples: 3-5 <example> tags for high-activation skills

D. Command Quality (for each command in the plugin)

Read the command .md file:

• Frontmatter shape: description and argument-hint as separate YAML keys (not a single concatenated string)
• Argument hint: accurate, shows the expected arguments
• Body: actionable procedure, not just marketing copy
• Integration: references the agent or skill that does the real work

Step 3: Cross-plugin coherence

• Detect overlapping triggers between plugins (two agents competing for the same activation)
• Detect orphan references (agent X references skill Y that doesn't exist)
• Identify missing sibling-conflict DO NOT TRIGGER clauses when two plugins cover adjacent domains

Step 4: Report

Write .marketplace-review/REPORT.md with:

# Marketplace AI Review -- <marketplace-name> -- <date>

## Summary
- Plugins evaluated: N
- Critical issues: K
- Recommendations: M

## Per-plugin findings
### <plugin-name>
- Description score: X/5 -- <reasoning>
- Agents: ...
- Skills: ...
- Commands: ...
- Recommended fixes:
  - [CRITICAL] ...
  - [WARNING] ...

## Cross-plugin observations
- ...

With --fix flag

For each recommendation, offer to apply the fix with user confirmation. Stage changes via Edit, do not auto-commit.

Notes

• Marketplace-agnostic: works on any project with a standard Claude Code plugin layout.
• Complements (does not replace) the deterministic audit run by /marketplace-ops:marketplace-health and the detailed lint in /marketplace-ops:skills-validate.