A dataset of 3.8 million agent-skill files, mined from 282,200 public GitHub repositories — the first large-scale record of a brand-new software artifact.
Anthropic opened the agent-skill format in October 2025. Nine months on, skill files sit in the millions across public repos — written in natural language, loaded probabilistically by a model, with no compiler or type checker to verify the selection. No existing dataset recorded this population.
Collected in July 2026 via a read-only pipeline against the GitHub code-search and REST APIs.
A skill is a folder containing SKILL.md: YAML front matter (name + description) followed by a Markdown body of instructions. The folder can also bundle scripts and reference documents.
The agent decides at run time whether to load a skill by matching the task against its description. No compiler or type checker verifies the choice — a vague description can mean a skill silently never fires.
Claude Code is the reference implementation; .claude/skills/ is treated as the standard location.
GitHub code search caps results at 1,000 per query, and its own total_count estimate proved unreliable.
One row per discovered file: repository, path, exact basename, location class, content hash, representative flag.
Owner, star count, primary language, fork status, creation date, and last-push date.
Files bundled alongside a representative skill, plus per-skill flags for whether it has scripts or reference material.
First/last commit dates and anonymised author codes, for standard-location skills plus a size-stratified sample.
Query, start/end timestamps, and result count for each collection run — full provenance.
How fast does the format spread, what do developers codify, and is skill-writing converging into a formulaic genre?
Do skills stay tied to one tool, or generalise to any agent that implements the open specification?
With no central registry, skills spread by copying folders — how concentrated is reuse, and do copies follow known clone-genealogy patterns?
Which code metrics — churn, clone coverage, readability — have a meaningful equivalent for natural-language instructions?
Skills can trigger command execution and network access — can a modified copy hide a supply-chain attack with no formal review?
Code search reported ~349,000 matches for the filename query. The pipeline actually retrieved over 3.8 million files — hence the size-partitioned discovery approach.
50.5% of collected files are verbatim duplicates. With no registry, reuse happens purely by copying folders between repositories.
Commit authors become keyed, non-reversible codes; emails and personal names are redacted. Bot logins and AI co-author trailers are kept.
Where to get GitSkills, and how to cite it
Public-repos only — treat it as a lower bound. GitHub code search skips files over 384KB, inactive repos, repos with 500,000+ files, and forks without more stars than their parent.