Reproduce “GitSkills: A Dataset of Agent Skills on GitHub” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Giuseppe Destefanis, Daniel Graziotin, Matteo Vaccargiu, Marco Ortu
GitSkills builds a dataset of agent skill files from public GitHub repositories, preserving occurrences and enriching one representative per distinct byte-identical content. Its July 2026 snapshot contains 3,797,117 occurrences from 282,200 repositories and 1,877,981 distinct contents. The authors report that 50.5% of occurrences are copies of another collected file; the dataset is a lower bound under indexing limits, not a census.
Key insights
- The dataset treats agent skills as both textual instructions and repository-distributed artifacts, enabling separate analysis of content properties and dissemination patterns.
- Exact duplication is already widespread: 3,797,117 occurrences collapse to 1,877,981 distinct contents, with 50.5% of files copied byte-for-byte from another collected file.
- A skill’s context matters: the accompanying folder may contain scripts and reference documents, while a copied file’s selected representative is not necessarily its original source and may not reflect every copy’s history or dependencies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.