How Skills works

We look for public GitHub repos that hold SKILL.md files in two ways: a list of repos we review by hand, and GitHub's repo search over topics such as agent-skills, claude-skills and codex-skills. You can also send us a repo through Get listed.

We read a repo only when it is public, not a fork and not archived, has had a push in the last 12 months, and has at least 50 stars or is on our reviewed list. A repo whose file list GitHub cuts short is skipped, because a file we never saw could not be checked.

Each skill is read at one commit, the latest on its repo's default branch when we read it, and its page names that commit. We read files through GitHub's APIs; when we have to fetch a file's raw bytes, we use them only if they match the fingerprint Git records for that file.

An official collection is a repo published from a vendor's own GitHub organisation that we chose to feature. The choice is ours, not the vendor's.

We don't store the names of commit authors or committers, and we drop e-mail addresses from a skill's metadata.

A skill is a folder with a SKILL.md file in it. We list two kinds: standalone skills, and skills that a Claude Code plugin in the same repo ships, as its .claude-plugin/marketplace.json maps them.

Three kinds are counted on their repo's page but not listed: a repo's own internal skills, kept in an agent's folder such as .claude/skills for work on that repo; test fixtures and templates; and mirrors, such as vendored copies and translated docs.

A folder identical to a skill listed elsewhere is a copy: we list the original once, and the copy's address leads to it. A skill held for review isn't listed at all (see the pattern check).

The data as of 26 Oct 2026 lists 3,888 skills from 42 repos by 41 publishers. 10,489 more SKILL.md folders are counted on repo pages but not listed, 988 copies are merged into their originals and 25 skills are held for review.

A copy is a folder whose files are identical to another's, so Git gives both the same fingerprint. The same SKILL.md with different files, or the same name and description, is not a copy, and both stay listed.

The original is the one we already treated as the original; otherwise the one in an official collection, then the repo GitHub created first, then the one we saw first, then the one with more stars. We review any change of original before it applies. The copies count on a skill's page counts other owners' repos only: forks and the same owner's other repos don't add to it.

Repo stars are GitHub stars on the repo that holds the skill, not on the skill itself.

We ask search engines to index a skill's page only when its repo is an official collection or has at least 50 stars, its SKILL.md has at least 20 lines after the frontmatter, and its description has at least 40 characters. Repo pages with fewer than 2 listed skills, publisher pages with fewer than 3 unless GitHub has verified the organisation's domain, and category pages with fewer than 10 skills stay out of search too. Every page stays readable here.

We read every text file in each listed skill's folder, at the commit its page names, and match it against the 45 rules below (rules version 2026.10.26). We don't run the skill, we don't read the web pages it points to, and we can't judge everything its instructions might lead an agent to do. The same rules run in your browser on Check a repo.

A gate rule can set a skill's status, and the skill's page lists what it found. A signal never sets the status: most describe what the skill can touch, and R07, R20, R37 and R39 are kept with the check but not shown on skill pages. A text file over 384 KB, or one we couldn't fetch in full, is an item to review (R21, R46), and so is a binary, archive or document we can't look inside (R32). Images and fonts are listed on the skill's page and never set its status.

We find a skill's licence most specific first: a licence file in the skill's folder, the license field in its SKILL.md, the licence GitHub detects for the repo, then a correction we have reviewed. A licence at the skill level that we can't read stops the search there, so a repo licence never stands in for it. Cards and pages show the licence as we found it, or No licence.

We show a skill's SKILL.md text here only for the original, not a copy, only when its licence is one of MIT, MIT No Attribution, Apache 2.0, BSD 2-Clause, BSD 3-Clause, ISC, BSD Zero Clause, Unlicense, CC0 1.0 or CC BY 4.0, and never when our check flagged hidden or instruction-override text in it. The licence file, and the NOTICE file Apache 2.0 asks for, go beside it when the repo has them, and the text loads only when you ask for it. Otherwise the page links to the file on GitHub.

Each skill gets up to three categories, from keyword rules over its name, description and folder and its repo's GitHub topics. A skill no rule places goes under Other. Categories are assigned automatically and can be corrected (see disputes, corrections and removal).

Coding & code review · Testing & QA · Frontend & design · Office & PDF · Data & analytics · Cloud & DevOps · Security · AI & ML engineering · Science & research · Writing & content · Marketing & growth · Product & project management · Business & productivity · Agent tooling · Other

We rebuild the catalog from GitHub in refreshes, without a fixed date. The data on these pages is as of 26 Oct 2026. Every page says when its facts were read: repo facts carry the date we read them from the GitHub API, and each check names the commit it ran on.

A skill keeps the date we first saw it across refreshes, and its last-changed date is the last commit that touched its folder. Between refreshes a repo can change; a page still describes the commit we read.

If you wrote a skill, own the copyright in it, or are named on a page, email info@qai.io with the page's link and what you'd like changed: a finding you think is wrong, a licence or category to correct, or a page to remove. We respond to removal requests.