A Skill is a folder, not just a Markdown file. Three-level progressive disclosure makes it nearly free to keep around: ~100 tokens of metadata at startup, the body on task match, scripts and references on demand.
A Skill is a folder, not just a Markdown file. Three-level progressive disclosure makes it nearly free to keep around: startup reads only ~100 tokens of metadata, the body loads on task match, and scripts and references are read on demand.
---
name: pr-review # required; also the /pr-review command name
description: > # required; the ONLY signal for when to trigger:
Review PRs against team standards: # say what it does + when to use it +
security, performance, test coverage. # the words users might say
Use when the user says "review /
look at this PR".
allowed-tools: Read, Grep, Glob # optional: restrict tools (read-only review)
context: fork # optional: run in a forked context
agent: Explore # optional: delegate to a built-in/custom subagent
user-invocable: false # optional: Claude-triggered only, no slash menu
---
# Review process
1. Use Grep to locate the touched modules and their callers
2. Check item by item against references/checklist.md
3. Output a three-tier report: blockers / suggestions / ignorable
# Gotchas (the most valuable section: real scars)
- Any change to signature verification under webhooks/ is a blockerThe folder can also hold references/ (docs), scripts/ (executables), and assets/ (templates) — mention them in the body and Claude will read them at the right moment. Custom slash commands have merged into Skills: .claude/skills/deploy/SKILL.md automatically becomes /deploy.
Skill directories hot-reload: adding or editing a skill takes effect immediately in the session, no restart needed (except creating the top-level skills directory itself).