Skip to main content
A skill is a folder with a SKILL.md and, optionally, scripts, references and templates. Claude sees only each skill’s name and description until a task matches; then it reads the whole SKILL.md and follows it. That keeps hundreds of skills cheap to have installed.
Written to .claude/skills/systematic-debugging/, every file of the folder included.

​
Anatomy

.claude/skills/systematic-debugging/SKILL.md
The description decides when the skill is used, so good skills say both what they do and when to use them. The format follows the open Agent Skills specification.

​
Skills from their author’s repository

Some skills are not copied into this repository. Their catalog folder holds only a source.json that points at the author’s repo, pinned to a full commit:
source.json
You install them the same way (--skill git/babysit-pr). The CLI downloads the files at that exact commit, checks each one against its git blob hash, and installs the skills listed in requires too. An upstream change reaches you only after a pull request here moves the pinned commit, so what you get is what was reviewed and scanned. How to add one: External skills.

​
Security

Skills can ship scripts that Claude will run. Every skill in the catalog is scanned by NVIDIA SkillSpector, and a pull request that adds a high-risk skill is blocked. Read the files on the component page before installing. Browse all skills at aitmpl.com/skills.