3.8M SKILL.md files on GitHub: audit what your agent loads

Agent skills are folders with a SKILL.md file in them — plain-language instructions an AI agent loads when it decides the skill is relevant. Anthropic introduced the format in late 2025 and it spread fast across competing agent tools. Now researchers have gone and counted them. A new dataset finds 3,797,117 SKILL.md files across 282,200 public GitHub repositories, and the numbers underneath that headline are the reason to care.

What actually happened

GitSkills, posted to arXiv on August 11 by Giuseppe Destefanis, Daniel Graziotin, Matteo Vaccargiu, and Marco Ortu, is a July 2026 snapshot of the whole public SKILL.md population. It ships as a single SQLite file with repository metadata, file paths, content hashes, parsed front matter, folder contents, and commit history for a subset.

The number that matters is the second one. Those 3.8 million files collapse into 1,877,981 distinct contents — meaning roughly half of every skill file on public GitHub is a byte-level copy of another one. Skills are being forked, vendored, and pasted between repos at scale, and the copies do not track their origin.

The authors are direct about why this is a new research problem rather than a variation on package management. Skills are written mainly in natural language, an agent picks them probabilistically at run time, and — in their words — "no compiler or type checker verifies the selection." There is no resolver, no lockfile, no semver, and no build step that fails when a skill drifts.

Why the agent skill supply chain matters for your business

If your team has installed skills into Claude Code, Cursor, or an internal agent, you have taken on dependencies with none of the machinery you would demand from an npm package.

Three things follow, and they are all cheap to fix.

Vendor your skills, don't reference them. A skill copied into your repo at a known commit is a skill you can diff. A skill pulled from a marketplace at run time is instructions that can change under you between Tuesday and Wednesday. The duplication in this dataset is evidence the ecosystem is already doing the copying — just without recording where anything came from.

Read every skill before it ships. A SKILL.md is executable in the only sense that matters: it steers an agent that has your credentials, your repo, and your database. Anthropic's own Agent Skills documentation says to use skills only from trusted sources, to audit bundled scripts and resources, and to treat installing one like installing software. Skills that fetch from external URLs get called out specifically — the fetched content can carry instructions of its own.

Keep the set small. Probabilistic selection gets worse as the candidate pool grows. Ten sharp, tested skills beat two hundred scraped ones. We said the same thing about turning SOPs into recorded skills: the value is in the process you actually run, written down precisely — not in volume.

Sources: GitSkills (arXiv 2608.10906), Anthropic Agent Skills documentation.