Researchers release dataset of 3.8 million AI agent skills from GitHub

Researchers have assembled the first major dataset documenting the rapid adoption of agent skills across GitHub, capturing nearly 3.8 million skill files collected from over 280,000 public repositories.

What are agent skills?#

An agent skill is essentially a folder containing a configuration file that provides natural-language instructions for a language-model agent, along with optional supporting scripts and reference materials. When tasked with work, an AI agent examines available skills and probabilistically selects those it judges relevant to the job at hand. Anthropic introduced the format as an open specification in October 2025, and within nine months, the ecosystem had grown explosively.

Unlike traditional software artifacts that software engineering researchers typically study, skills operate quite differently. They rely entirely on natural language rather than code, lack any central registry or package manager system, and bypass the usual safeguards of compilation or type checking. Developers currently reuse them simply by copying folders between repositories.

The GitSkills dataset#

The new dataset, called GitSkills, documents this emerging landscape. Researchers collected 3.8 million skill files from July 2026 and organized them into 1.9 million distinct contents. For each unique skill, the team enriched representative examples with complete text, structured metadata, folder contents, repository information, and—for a subset—commit histories tracking how skills evolve over time.

The entire dataset fits within a single, portable SQLite file designed to be immediately usable for research without requiring external tools or complex setup.

Why it matters#

How developers write, reuse, and maintain these new skill artifacts had remained an open question, with no existing dataset available to study the broader population. This resource now enables empirical research into skill adoption patterns, code reuse practices, how teams structure their instructions, who creates popular skills, maintenance challenges, and potential security concerns as the ecosystem matures.