A public GitHub now holds 3.8 million agent skill files.

And finds that over half of the agent skill files on public GitHub are exact copies of another file.

Only about 1.9 million of the 3.8 million files are actually different.

That is what a format with no registry and no package manager looks like.

People copy the folder and move on. A skill is a folder with a Markdown file that tells an agent how to handle a task.

The agent reads the short description at run time and decides on its own whether to load it. No compiler checks that decision.

So the skill in your repo is frozen, and whoever wrote it cannot send you a fix.

The paper raises the obvious next question: whether edited copies of popular skills add commands or network calls the original never had.

Nobody could check that at this scale before.

The whole dataset ships as one SQLite file, so now somebody can.

– arxiv. org/abs/2608.10906

Title: "GitSkills: A Dataset of Agent Skills on GitHub"