3.8 million SKILL.md files mapped across GitHub repositories

Nine months after Anthropic released the SKILL.md format as an open specification, roughly 3.8 million such files are now scattered across 282,200 public repositories on GitHub. A new study collected every instance, extracted the contents, and compressed them into a single dataset named GitSkills. The work appears as a pre-print on arXiv (2608.10906) and has not yet undergone peer review.

SKILL.md files differ from ordinary source code in ways that hinder conventional data mining. They are written in natural language, selected probabilistically by an AI agent at runtime, and lack a compiler or type checker to validate the choice. No central registry or package manager exists, so the files propagate simply by copying folders between repositories. The result is an enormous, hard-to-track collection of files.

GitSkills aggregates all identified files into 1,877,981 unique contents. The dataset is distributed as a single SQLite file that includes document skeleton data, folder contents, and identity history. Researchers seeking additional AI papers can locate them in the DAIR.AI academy, an organization dedicated to AI research and education.

The SKILL.md file functions as more than technical documentation; an AI agent can consult it to decide which tools to invoke and which actions to perform. Consequently, the provenance and permission boundaries of each file become critical, especially when files are manually copied between repositories without any central oversight.