SkillMOO is the independent rating authority for AI agent skills. Every grade below is computed by static analysis, not opinion — and ratings are never for sale. This page is the full rubric.
| Safety | We read the skill and any bundled scripts and flag real risk: reading credentials, reaching the network, destructive commands, and — most importantly — combinations like credential access paired with network egress (an exfiltration path), detected from behavior, not banned keywords. Real today. |
| Token cost | We count the tokens a skill spends on every call and compare it to the portfolio median (currently 2,062). A skill that costs several times the median is a silent, permanent tax. Real today. |
| Conflicts | We compare the trigger surface of every skill against every other. Two skills that both fire on the same request shadow each other — the top cause of an agent picking the wrong skill. Real today. |
| Efficacy | Whether a skill actually improves the model's output, measured by running real tasks with and without the skill and grading deterministically. Shown only when measured, always labeled with the model and task suite. Never invented. |
| A / B | Clean and safe. Lean or near-median token cost, no dangerous capability combination. |
| C | Usable, but review first — commonly an over-broad trigger or a description that never says when to use it. |
| D / F | Serious problems — a credential-plus-network exfiltration path, a destructive capability, or a hard safety block. Do not install as-is. |
Of 42 public skills we've graded, 3 carry a high-severity safety finding, 10 are token-bloated versus the median, and we found 198 conflicting trigger pairs. These are computed facts about real, public skills — see each rated skill page for the evidence.
SkillMOO is an independent rating authority for AI agent skills. It grades the safety, token cost, and conflicts of a SKILL.md file — the instruction files that add capabilities to coding agents like Claude Code, Codex, and Cursor — so you can tell whether a skill is worth installing before you install it.
Every grade is computed by static analysis, not opinion. We read the skill text and any bundled scripts and check three things: safety (does it read credentials, reach the network, or run destructive commands?), token cost (how many tokens it spends on every call versus the median), and conflicts (does its trigger overlap with other skills so the agent fires the wrong one?). Efficacy — whether a skill actually improves model output — is only ever shown when we have measured it on a labeled model and task suite.
No. Ratings are never for sale. That independence is the entire point of SkillMOO. We monetize the fix (one-click optimization) and teams, never the score.
No. We rate a skill and link to its source; we do not rehost skill content. The optional `skillmoo scan --publish` uploads only the derived report you chose to share, with your home paths stripped.
A and B are clean and safe. C means usable but worth a review — often an over-broad trigger or missing usage guidance. D and F mean serious problems, usually a credential-plus-network exfiltration path or a hard safety block. The grade combines the safety, token, and conflict signals.
Not yet. Safety, token cost, and conflicts are computed today and are real. Efficacy (measured lift) is being added skill-by-skill and is always labeled with the exact model and task suite it was measured on. We never invent an efficacy number — "not yet tested" is an honest, valid state.
No — and this is deliberate. You can tell us, anonymously and in one click, whether a rating feels too harsh or too lenient. We watch those as an aggregate trend to decide which skills to re-analyze first. But a grade only ever changes when OUR re-measurement changes. Feedback flags what to re-check; it never moves a score directly. That is what keeps the ratings un-gameable — no amount of clicking can brigade a grade.