Skills

One Agent, a Stack of Playbooks: Turn Your Repeat Tasks Into Claude Skills

7 minute readUpdated October 2026Explore more

TL;DR

You don't need a separate agent for every job. Anthropic's own engineers describe skills as a way to specialize a general-purpose agent instead of building fragmented, custom agents for each use case. The play: Claude interviews you about your work and lists the tasks you repeat. Then each one becomes a playbook, saved as a skill in ~/.claude/skills. Next time you just name the task, and Claude loads that playbook. Below are the interview prompt, the playbook-to-skill prompt, a blank template and a finished example.

It's tempting to build a new agent for every job: one for email, one for reports, one for research. Each needs its own setup, and most never get finished. The simpler setup is one agent that knows a stack of playbooks. In Claude Code, a playbook is a skill: a folder with a SKILL.md file that holds the steps for one task you repeat.

What Anthropic actually says about skills

  • Anthropic describes Agent Skills as "organized folders of instructions, scripts, and resources that agents can discover and load dynamically to perform better at specific tasks."
  • From the same post: "Instead of building fragmented, custom-designed agents for each use case, anyone can now specialize their agents with composable capabilities by capturing and sharing their procedural knowledge."
  • They compare building a skill to "putting together an onboarding guide for a new hire."
  • The Claude Code docs say to create a skill "when you keep pasting the same instructions, checklist, or multi-step procedure into chat."
  • Only each skill's name and description sit in context until it's needed. The full SKILL.md loads when you or Claude invoke it, so a big stack of playbooks stays cheap.
Equipping agents for the real world with Agent Skills (Anthropic)

Anthropic's engineering post introducing Agent Skills, published October 16, 2025.

Extend Claude with skills (Claude Code docs)

The official reference: where skills live, the frontmatter fields, and how Claude decides when to load one.

Step 1: Let Claude interview you about your job

Anthropic's Claude Code best practices recommend having Claude interview you with the AskUserQuestion tool before a big piece of work. They use it for feature specs. Here it's pointed at your job, so Claude finds the playbooks for you.

promptI want to turn the work I repeat into Claude Code skills. Interview me about my job using the AskUserQuestion tool.

Go one topic at a time:
- My role, who I do the work for, and what a normal week looks like.
- Every task I do more than once a month: what kicks it off, what I start with, the steps, the tools and files I use, and what "done" looks like.
- Where each task goes wrong, the judgment calls I make, and the rules I never break.
- Where my best past examples of each task live, so you can learn from them.

Don't ask obvious questions. Dig into the steps I do on autopilot and would forget to mention.

When we've covered everything, write PLAYBOOKS.md with a table: task, how often, time per run, trigger, inputs, output, and whether it's a good skill candidate (repeatable steps and a clear finish line). Rank the candidates by hours saved per month. Don't build anything yet.

Step 2: Turn one task into a playbook skill

Start a fresh session (type /clear) so the build has clean context, then paste this. Do one task at a time.

promptTake the top task in PLAYBOOKS.md and turn it into a Claude Code skill.

1. Pick a short kebab-case name I'd actually type, like weekly-report.
2. Create ~/.claude/skills/<name>/SKILL.md with these sections: frontmatter (name, description), When to use, Inputs, Steps, Rules, Done when, Output.
3. Start the description with what the skill does, then the exact phrases I'd use to ask for it.
4. Write the steps as numbered instructions with my real tools, folders and rules from the interview. No generic advice.
5. Keep SKILL.md under 500 lines. Put long reference material, examples or templates in separate files in the same folder and link to them from SKILL.md.
6. If the task has side effects (sends, posts, deletes, deploys or spends money), add disable-model-invocation: true so it only runs when I type the slash command.
7. Show me the file before you save it. Then run it on one real example and fix whatever went wrong.

The playbook template

Every playbook follows the same shape. The description matters most: it's what Claude reads to decide when to load the skill, so put the job first and then the words you'd actually say.

markdown---
name: task-name
description: What this does in one line. Use when I say "...", "...", or ask you to ....
---

# Task name

## When to use
- Trigger: what kicks this task off
- Not for: similar jobs this skill should not handle

## Inputs
- What I give you: ...
- Where to find the rest: folders, files, links

## Steps
1. ...
2. ...
3. ...

## Rules
- What must always be true
- What you must never do

## Done when
- A finish line you can check (file saved at ..., every row filled, totals match the source)

## Output
- Format, length, where to save it, and who it is for

Example: a finished playbook

Saved as ~/.claude/skills/client-update/SKILL.md.

markdown---
name: client-update
description: Drafts the weekly client update email from this week's project notes. Use when I say "client update", "weekly update", or "write the Friday email" for a client.
---

# Client update

## When to use
- Trigger: end of the week, or when I ask for a client update
- Not for: invoices or proposals

## Inputs
- The client name. Ask if I didn't give one.
- notes/<client>/ for this week's notes

## Steps
1. Read every file in notes/<client>/ changed in the last 7 days.
2. List what shipped, what is in progress, and what is blocked, with dates.
3. Draft the email in 3 short sections: Done, Next, Need from you.
4. Take every number and date from the notes. Never estimate.

## Rules
- Plain language. No jargon, no hype.
- If something slipped, say so in the first section with the new date.
- Draft only. Never send anything.

## Done when
- Every item traces back to a note file.
- The email is under 200 words.

## Output
- Save to drafts/<client>-update-<date>.md and show it to me.

Step 3: Just name the task

  • Type the slash command with anything it needs after it, like /client-update Acme. Claude Code passes the text after the name to the skill.
  • Or ask in plain words: "write the Friday email for Acme." Claude matches your request to the description and loads the playbook.
  • Not triggering? Ask Claude "What skills are available?" to check it loaded, then add the words you actually use to the description.

Keep the playbooks sharp

Anthropic's advice is to iterate with Claude: when a run goes off track, ask it to capture what went wrong and what worked into the skill. After a messy run, paste this:

promptThat run of /client-update had problems. Update ~/.claude/skills/client-update/SKILL.md so the next run gets it right:
- Add a step or rule for each mistake you made this session.
- Keep what worked.
- Show me the diff before you save it.

Now you just say the word, and one agent does the work. Inside the Claude Code Club we share the skills, prompts and workflows we actually run every day, and help each other set them up. It's nine dollars a month at https://www.skool.com/claudecodeclub/about. Everything on this page works without it.

Common questions

  • What's the difference between an agent and a skill?

    An agent is the thing doing the work, like Claude Code itself. A skill is a folder of instructions and optional scripts that teaches that agent one specific job. Anthropic's pitch for skills is that you specialize one general agent with many skills instead of building a separate agent for each job.

  • Where do Claude Code skills live?

    Personal skills go in ~/.claude/skills/<skill-name>/SKILL.md and load in all your projects on that computer. Project skills go in .claude/skills/<skill-name>/SKILL.md inside a repo, so everyone working in it gets them.

  • Do lots of skills slow Claude down?

    Much less than pasting the same instructions every time. Only each skill's name and description stay in context, and the full instructions load when the skill is used. If you install a very large number, Claude Code may drop some descriptions from its skill listing to fit the budget, starting with the skills you use least.

  • Can a skill run by itself without me asking?

    Yes, by default Claude can load a skill when your request matches its description. For anything with side effects, like sending, posting or deploying, add disable-model-invocation: true so it only runs when you type the slash command.

  • How many playbooks should I start with?

    One. Pick the task from your interview that eats the most hours each month, build it, use it for a week, and fix it after each messy run. Then do the next one.

Want help turning your job into playbooks?

Get the other 8 in the skills stack, plus 8,000+ members - $9/mo, cancel anytime.

Join the Club