Skills
A skill is a folder with a SKILL.md. Its frontmatter description is the
trigger — it is the only part loaded until the skill fires, so it decides
whether the skill is ever used.
---
name: my-skill
description: What it does. Use when <situation>. Do NOT use when <situation>.
---
# My skill
...instructions...
Mine vs everyone else's
| Where | Updated by | |
|---|---|---|
| Mine | claude-code/skills/<category>/ in this repo | me, in git |
| Theirs | ~/.claude/skills/ via the CLI | npx skills update |
Never edit an installed third-party skill in place — the next update reverts it silently. Fork it into the repo instead.
Categories
| Folder | Contents |
|---|---|
communication/ | caveman, caveman-commit, caveman-review, ponytail, abrege-frere |
delivery/ | spec-workflow, implementation-planning, completion-plan, progress-audit, harvest-questions, ai-time-estimates |
quality/ | verify-change, verify-module, verify-quality, verify-security, debug-loop, pr-workflow, pr-review-governance |
architecture/ | diagram-to-adr, diagram-to-likec4, likec4-to-openspec |
design/ | interface-cheatsheet |
tools/ | pdf-inspector, reverse-engineer-api, memory-snapshot-perf |
Folders are for browsing. Install flattens them — see what lands where.
The skills CLI
npx -y skills@latest add <owner>/<repo> # install
npx -y skills@latest add <owner>/<repo> -l # list first, install nothing
npx -y skills@latest list # what's installed, from where
npx -y skills@latest update -y # update everything
npx -y skills@latest find "<topic>" # search
npx -y skills@latest use <repo>@<skill> # one-shot, no install
npx -y skills@latest init <name> # scaffold a new one
Browse at skills.sh. Full recommendations in
skills/THIRD-PARTY.md.
Writing one that actually fires
- The description is the trigger. Say when to use it and when not to.
pdf-inspectorsays "not for small PDFs, native Read handles those" — that line is what stops it shadowing a built-in. - One skill, one job. Two jobs means two descriptions competing.
- Record
source:in the frontmatter — upstream, route, license, date. - Ship scripts when prompt-only will not do. The
verify-*skills carry 8–13 KB Node analyzers; they are not instructions pretending to be tools.
Security
A skill is a prompt that runs with your tool permissions. Read SKILL.md
before installing, exactly as you would read a shell script before piping it
to bash. Look at allowed-tools, and at whether it tells the agent to run
network commands or write outside the project.