The rapid emergence of agent "skills" (SKILL.md folders) created a new, natural-language software artifact that standard code-mining pipelines rarely capture. GitSkills records this phenomenon at scale—millions of discovered files linked to repositories and enriched with metadata—so researchers can study how agent instructions are authored, copied, and maintained in the wild.
What Sets It Apart
- Large-scale, provenance-aware corpus: contains 3,797,117 discovered SKILL.md occurrences from 282,200 public repositories, grouped into 1,877,981 distinct contents so you can analyze both unique skill texts and their copies.
- Representative enrichment: one representative per content group includes full text, parsed YAML front matter, sibling files in the skill folder, and for a large subset, anonymized first/last commits and commit counts—enabling temporal and maintenance analyses.
- Practical data formats and tooling: distributed as a single self-contained SQLite archive and Parquet mirrors with dataset configs (artifacts, artifact_siblings, repos, mining_runs) to support pandas/Datasets/Polars/Dask workflows.
- Research-focused scope and anonymisation: collects only public GitHub data, documents collection limits (code-search coverage, file size cap), and replaces personal identifiers in commits with stable anonymized codes to preserve provenance without exposing identities.
Who It's For and Tradeoffs
Great fit if you study prompt engineering, agent/tool integration, software-engineering metrics for natural-language artifacts, or supply-chain risks in agent ecosystems. The dataset enables analyses of adoption, clone genealogy, readability/churn metrics, and whether modifications introduce risky commands.
Look elsewhere if you need complete repository indexing (the crawl targets public, indexed branches and excludes very large repos and files over the search cap) or if you require license-cleared text for redistribution—file contents retain their upstream licenses and must be checked before reuse.
Where It Fits
Use GitSkills when you want empirical, reproducible studies about how language-model agents are extended via repository-hosted skills and when you need both content-level deduplication and per-copy provenance to study reuse and maintenance patterns.
