GitSkills: A Dataset of 3.8 Million Agent Skills from GitHub

The recently launched dataset, GitSkills, encompasses 3,797,117 SKILL.md files sourced from public GitHub repositories. These files denote agent skills, which were first introduced by Anthropic as an open specification in October 2025. An agent skill comprises a folder that includes a SKILL.md file detailing instructions for a language-model agent, potentially accompanied by scripts and reference materials. The agent activates the skill when it determines that a task aligns with the skill's description. In less than a year since the format's debut, millions of skill files have emerged. Unlike standard software engineering artifacts, these skills are primarily articulated in natural language, probabilistically selected during runtime, and are not subject to compiler or type checker validation. They lack a centralized registry or package manager, proliferating through folder duplication across repositories. This dataset seeks to explore how developers create, reuse, and maintain skills, filling a gap as no prior dataset has documented this demographic. The announcement was published on arXiv under the identifier 2608.10906.