About Skill Atlas
What it is
Skill Atlas looks at every public agent skill file (SKILL.md) in a large snapshot of GitHub. It answers one question for each skill: is this worth using?
It does four things:
- Finds copies. Many skills are copied from repo to repo. We group copies together and pick the original.
- Spots slop-skills. A slop-skill is one of many look-alike skills mass-produced by one account from a template.
- Checks for danger. We look for code that reads passwords or keys and sends them somewhere, for hidden instructions, and for skills that come from collections of deliberately harmful test skills.
- Gives a verdict. Each skill gets
use,cautionorignore, with the reasons.
A near-copy is a skill whose text is almost the same as another skill, with small edits.
Where the data comes from
The skills come from GitSkills (mvaccargiu/gitskills), shared under CC BY 4.0. Please cite: Destefanis et al., GitSkills, arXiv 2608.10906. Our verdicts are built on that data, so the same credit applies. Each skill keeps its own license. Check it on GitHub before you reuse a skill.
What to keep in mind
- The verdicts come from simple rules, not from a person reading each skill. They catch a lot, but not everything.
- A high risk score means "read this before you run it". It does not prove the skill is harmful.
- The data is a snapshot. Some repos have changed or gone since.
- The atlas can't tell whether a skill only works inside one project. That was the most common problem our reviewer found.
How we measured this (technical detail)
Known errors. 3 of 50 hand-checked embedding pairs were wrong: letter-suffixed template siblings, shared banners before the frontmatter, and same-author rewrites. The translation tag is a cheap script/stop-word test. Origin dates in the snapshot are unreliable, so ranking puts stars first.