State of Agent Skills

This is the finding that most contradicts how Skills are usually described. Of 191 catalogued Skills, 67 are published by Anthropic and 88 by an identifiable company shipping a Skill for its own product — Stripe, Resend, Supabase, Twilio, Playwright, Semgrep, and others. Only 36 (19%) are community-authored.

The practical consequence for anyone choosing a Skill: for a specific product, check whether that company publishes one before evaluating third-party alternatives. A vendor-maintained Skill tracks API changes; community advice about an API does not.

SKILL.md is one format read by many agents, which makes “write once, run anywhere” sound like a property of the spec. Verified compatibility tells a different story: 49% of catalogued Skills are verified on one independent agent or fewer, and only 9% reach four or more.

Part of that gap is testing coverage rather than capability — verification is per Skill per agent and it takes real work. But part of it is structural. The file format ports cleanly; what does not port is what the instructions assume about their host. A Skill that names one agent’s tools, expects a shell, or relies on an upload flow will load on another agent and quietly do less. That is the failure mode assumed compatibility hides.

56 of 191 catalogued Skills (29%) are action-taking: they send email, run deploys, charge cards, drive browsers, or call paid APIs. 70 are read-only and 65 are read-write within a project. That split is the single most useful thing to check before installing, and it is why capability level appears on every entry.

Licensing is healthier than most package ecosystems: 184 of 191 Skills are open source and 180 carry a permissive OSI licence. The caveat found repeatedly during verification is that a licence claimed in a README badge or in SKILL.md frontmatter is not always backed by a licence file in the repository — and no licence file means all rights reserved by default. Read the file, not the badge.

Category depth tracks where vendors have shown up. The deepest categories are the ones with a clear commercial owner; the thinnest are the ones where no company has an obvious reason to publish. 0 of 20 categories here still hold fewer than eight entries, and we publish that rather than padding them — a category filled with weak entries is worse than a short one.

1. 1.frontend-design92
2. 2.docling68
3. 3.hyperframes66
4. 4.antigravity-awesome-skills65
5. 5.remotion-best-practices64
6. 6.superpowers63
7. 7.playwright-cli62
8. 8.firebase-hosting-basics61
9. 9.browser60
10. 10.ui-test60

A “Verified” badge means nothing without a date attached, so here is ours: of 145 entries carrying a verification date, 37 were checked within the last 60 days and 108 were not. The oldest verification on record is April 29, 2026.

That backlog is real and it is ours to close — re-verification runs continuously, and an entry that fails is removed rather than quietly aged. 4 entries have gone that way so far; the reasons are on the changelog.

Read the change record · How scoring works