Why Claude plugin ratings feel noisy right now
Search interest in Claude Code, plugins, and skills rose sharply through 2025–2026. Marketplaces and GitHub READMEs filled quickly — which is useful for discovery and weak as a quality signal.
A plugin with thousands of stars can still dump opaque permissions on you, break after one Claude update, or wrap a thin prompt in marketing language. If you only sort by stars, you will install the loudest tools, not the best ones.
The six dimensions that actually matter
1. Job fit — Does it solve a job you repeat (SEO audit, docs, tests, research), or is it a demo?
2. Skill / command clarity — Can you name the slash commands or triggers without reading a novel?
3. Permissions honesty — What files, network, or MCP servers does it touch? Vague access is a red flag.
4. Evidence — Does output cite sources, show diffs, or invent confident nonsense?
5. Maintenance — Last commit, changelog, and whether it tracks Anthropic’s current skill format.
6. Blast radius — If it fails, does it fail loudly with a recoverable state, or silently rewrite your repo?
A simple 1–5 score you can reuse
Score each dimension 1–5, then weight job fit and permissions honesty higher than polish. A plugin that is conservative but safe beats a flashy one that needs root-equivalent trust.
Write one sentence for “when I would uninstall this.” If you cannot answer that in under twenty words, you do not understand the tool yet — wait before pinning it in your default workflow.
Turn ratings into a recurring process
Re-score plugins after major Claude Code releases. Listings drift: new sub-skills appear, APIs change, and yesterday’s “must install” becomes today’s abandoned README.
AuditHQ’s Claude plugin projects are built for this loop — track listing quality and peer comparisons on a schedule so your scorecard stays current. See the related checklist for the field-by-field pass.