Methodology
How we score Skills
Every Skill on this site has a composite 0–100 score. The number is derived from 9 verifiable inputs — no hand-curated quality dial, no hidden weights. The current rubric is version 1.0.
Inputs
The 9 inputs
| Input | Weight | Source |
|---|---|---|
| Install count | 20 | Verified installs across skills.sh, npm, and GitHub releases (log-normalized). |
| Provenance | 15 | Anthropic > verified org > community. Community skills with high adoption get a partial boost. |
| GitHub stars | 15 | Stars on the source repo (log-normalized). |
| Recency | 10 | Days since the last commit. Skills updated in the last 14 days score full marks. |
| Compatibility | 10 | Number of agents (out of 5: Claude Code, Cursor, Codex, Gemini, Antigravity) the skill is verified to work with. |
| Documentation depth | 10 | README, docs URL, "best for", and "not ideal for" all present. |
| Install ergonomics | 10 | One-command install scores full; plugin install scores partial; manual install scores low. |
| License | 5 | OSI-approved permissive licenses score full; closed but verifiable scores half. |
| Verification freshness | 5 | How recently we tested that the install command still works. |
Why
Design choices
Composite over single-axis. Other directories rank by install count alone, which rewards age over quality. We use 9 inputs so recency, provenance, and ergonomics also count.
No aggregateRating in JSON-LD. We do not collect user reviews and so do not publish a fake aggregate. The composite score is a quality signal, not a five-star average.
Provenance has weight, not a veto. Anthropic and verified-org skills get an automatic boost, but a community skill with high adoption can still outrank them — Superpowers (40k+ stars) is the canonical example.
Verification freshness matters. Five points of the score reflect when we last verified the install command works. A skill we have not retested in months loses points until we do.
Worked example
Example: frontend-design
Frontend-design earns full marks on install count (277K+), provenance (Anthropic), compatibility (5/5 agents), license (MIT), and install ergonomics (one-command). Stars and recency are at 90% (a few weeks since last commit). Documentation is at 90% (a clear README and SKILL.md, no formal "not ideal for" section). Final score: 92.