How we score Skills

Two different dates appear on every entry, and they mean different things. Last update is the upstream repository’s own last push — the skill’s freshness, not ours. Verified is the date we last tested that the install command still works.

Cataloguing an entry requires all of: the repository exists and is reachable; the SKILL.md is present at the exact path we publish; the licence is confirmed by reading the licence file rather than a README badge; the provenance tier is what the evidence supports; and the upstream last-push date is recorded from the repository.

The Verified date is set only when the install command has been executed. An entry that has cleared the checks above but has not had its install run yet ships with no Verified date at all — it loses the verification points until it does. That is deliberate: a badge that means “probably fine” is worth less than an empty field.

Compatibility lists tested agents only. The SKILL.md format is read by many agents, so almost any skill plausibly works everywhere. We list an agent when the skill was actually run there, or when the publisher documents a distinct install path for it. The characteristic cross-agent failure is a skill that loads and quietly does less, which is exactly what assumed compatibility hides.

Entries are removed. An archived upstream, an install command that stops working, a licence that changes in a hostile direction, or a publisher that cannot be verified all trigger de-listing, with a permanent redirect left behind rather than a dead URL. Every removal is on the changelog with the reason. The current re-verification backlog is published on the State of Agent Skills page — including the entries whose verification has aged past 60 days.