How the ranking works.
Every listing gets a score out of 100, and every point of it is shown. There are no votes to game, no editor's picks, and no number we made up. Across the 325 listings here the scores run from 29 to 85, averaging 57.
Safety (25) and Transparency (15), computed by reading the SKILL.md itself — destructive commands, data exfiltration, hardcoded secrets, permission footprint, obfuscation, documentation. This is the only part that belongs to the file rather than the repo, so it carries the most weight.
Full support scores 2, partial scores 1, out of a possible 10 across Claude Code, Cursor, Codex, Gemini CLI and Copilot. Inferred from the file and the repo layout, never guessed.
Contributors on the source repo, log-scaled. A star is one click; a contribution is someone sitting down and writing. A thousand contributors is not worth a hundred times ten.
Deliberately the smallest weight. Stars belong to the whole repo, so every skill inside a popular repo would otherwise inherit a score it did not earn.
Days since the last push. Under a month scores full; archived scores zero; abandoned for two years scores zero.
Two things this ranking cannot do
It does not know if a skill works. That needs running it against real cases, which we have not built. Effectiveness, Cost and Real-world use sit unscored on every page — and we would rather leave them blank than guess.
Forty of the hundred points are shared by every skill in the same repo. Contributors, stars and upkeep belong to the repository, not the file. That is why the file itself carries 60 — and why the tables show at most three entries from one repo before moving on, so a single large repo cannot take the whole board. That is a display rule; it never changes anyone's score.