Chapter 06 · Power tools
Skills & subagents
Two ways to stop repeating yourself: skills package instructions you reuse, subagents are helpers Claude sends off to do side-work without cluttering your conversation.
Skills: your own slash commands
A skill is a folder with one file, SKILL.md, containing instructions. Put it in .claude/skills/ in your project and it becomes a command: /monthly-summary, /weekly-report, whatever you name it. Claude can also invoke a skill on its own when your request matches its description.
---
name: monthly-summary
description: Build my end-of-month portfolio summary
---
1. Read the newest export in data/
2. Work out each position's gain or loss for the month
3. Update summary.html: the table, then a short plain-
English recap (net of fees) with the top 3 movers
4. Show me a preview before calling it done
Now /monthly-summary runs that whole checklist, every time, the same way. The easiest way to make one:
Skill or CLAUDE.md? CLAUDE.md is read every session, so keep it for always-true rules. Skills load only when needed, so they're the home for step-by-step workflows and occasional knowledge. Added a skill mid-session? /reload-skills picks it up.
Subagents: delegate the side-quests
Some jobs generate a mess: researching across dozens of files, digging through logs, comparing libraries. A subagent does that work in its own separate conversation and reports back just the answer, so your main chat stays clean and focused.
You rarely need to set anything up; Claude delegates on its own when it helps. You can also define specialists in .claude/agents/, one markdown file each:
---
name: proofreader
description: Check my writing for typos and awkward phrasing
---
You are a careful proofreader. Point out problems;
never rewrite without being asked. British spelling.
Manage them with /agents. A specialist can use a faster model, see only certain tools, and keep its own memory.
Which one, when?
| You want… | Reach for |
|---|---|
| A rule that applies to everything, always | CLAUDE.md |
| A repeatable multi-step workflow you trigger | A skill |
| Messy research done off to the side | A subagent |
| Something to happen automatically, no exceptions | A hook (next chapter) |
make me a skill called /tidy that finds unused files in this
project and asks me one by one whether to delete them