GitSkills: A Dataset of Agent Skills on GitHub
Giuseppe Destefanis, Daniel Graziotin, Matteo Vaccargiu, Marco Ortu
cs.SE, cs.AI
2026-08-11
GitSkills collects 3,797,117 SKILL.md files from 282,200 public repos into one SQLite file; 50.5% are byte-identical copies, enabling reuse and supply-chain research.
Anthropic opened the Agent Skills format as a specification in October 2025: a folder holding a SKILL.md file, YAML front matter with a name and description, a Markdown body of instructions, and optionally bundled scripts and reference files. The model decides at run time whether the task matches the description, and loads the skill only then. Nine months later, millions of these files sit in public GitHub repositories. They resemble nothing software-engineering research usually mines: mostly natural language, selected probabilistically by a model, with no compiler or type checker verifying the selection, no central registry, and no package manager. Skills spread by copying folders. How developers write, reuse, and maintain them became an empirical question with no dataset behind it.
A read-only pipeline drove GitHub code search and the REST/GraphQL APIs. Code search returns at most 1,000 results per query, and its reported total proved unreliable: roughly 349,000 for the filename query, against more than 3.8 million files actually retrieved. The fix was to partition the search space by file size until every range could be retrieved completely. Files are grouped by content hash; one representative per group (preferring the .claude/skills/ standard location) is enriched with full text, parsed front matter, folder contents, repository metadata, and commit history for standard locations plus a size-stratified sample of the rest. Every copy is retained with its repository and path. The result ships as one self-contained SQLite file, with commit-author accounts replaced by keyed one-way codes, emails masked, and AI co-author trailers kept.
| Dimension | Count |
| SKILL.md files | 3,797,117 |
| Repositories / accounts | 282,200 / 195,841 |
| Distinct contents after dedup | 1,877,981 |
| Files that are verbatim copies | 50.5% |
| Recorded sibling files | 7,264,865 |
| Skills with commit history | 458,548 |
Half the files are byte-identical copies, a direct measurement of how a registry-less ecosystem propagates.
This is the first population-scale instrument for the artifact, collected early enough to track how a new software artifact spreads and which conventions consolidate. The sharpest proposed research question is security: a skill can instruct an agent to run commands, reach external resources, and execute bundled scripts, while copies move between repositories without formal review. A modified copy that introduces command execution or network access absent from the original is the supply-chain attack of an ecosystem with no registry. Linguistic evolution matters too: monthly cohorts converging on formulaic templates would signal an emerging genre, and shrinking diversity may track rising machine authorship.
Public repositories only, and GitHub code search indexes default branches, files under 384 KB, recently active repositories with fewer than 500,000 files, and forks only when they out-star their parents, so the dataset reads as a lower bound. Filename matching pulls in near-misses (coding-skill.md) and lowercase files predating the format; the authors kept them, flagged with location class and front-matter validity, and left stricter filtering to analysts. The representative per group is not assumed to be the original source, and commit history covers a subset of skills. This is a dataset paper: every research question it lists is unanswered, and its value depends on whether the follow-up work happens.