AI Guides › Step-by-step guides
By Nigel Guy · 8 min read
The usual pattern goes like this: you find a job worth automating, you build an "agent" for it — a custom GPT, a new project, a fresh tool with its own login — and a month later you have fifteen of them, half-finished, each holding a slightly different version of how you like things done. It feels like progress because each one worked in the demo. It fails because every new job means starting the setup from nothing, and nobody, including you, remembers which agent knows what.
The rule: keep one agent and give it a written playbook for each job. The playbook is the asset; the agent is just the thing that reads it.
In Claude, that playbook has a name: a skill. A skill is a folder holding a SKILL.md file — a short description of when to use it, followed by the steps, standards and examples for one job. Claude reads only the description until a request matches it, then loads the rest. One Claude, many skills, and it picks the right one itself.
| You need | Detail at time of writing |
|---|---|
| A Claude account | Skills are available on Free, Pro, Max, Team and Enterprise plans, so you can do this for £0. Pro is listed at $20 a month in the US and is billed in local currency; check the £ figure at claude.ai/upgrade before you subscribe. |
| Code execution switched on | Skills depend on it. Individual plans: Settings > Capabilities > Code execution and file creation. |
| One job you do at least weekly | Something with steps and a recognisable "done", such as a client report, a weekly newsletter draft or a supplier invoice check. |
| About an hour | Most of it is the interview in Step 2. |
On Team and Enterprise, an organisation owner has to allow skills first, under Organization settings > Plugins & skills, on the "Policy" tab. Team plans have it on by default.
Write the job down in one line, including its trigger and its output: "Every Monday, turn last week's sales export into a one-page summary for the team." If you can't write that line, the job isn't ready for a playbook yet; it is still a vague wish.
Pick the job you most want to stop explaining. The rest wait until this one has survived a few real runs.
The playbook is only as good as what you get out of your own head, and you will skip things you do on autopilot. So make Claude ask. Open a new chat and paste this:
You are helping me write a reusable playbook (a Claude skill) for one recurring job, so that you can do it later without me supervising.
The job: [ONE-LINE DESCRIPTION OF THE JOB]
How often I do it: [FREQUENCY]
What I start with: [INPUTS, e.g. "a CSV export from Xero" or "my rough notes"]
Who the output is for: [AUDIENCE]
Steps:
1. Interview me one question at a time. Cover: the trigger, the inputs, every step in order, the decisions I make along the way and how I make them, the standards the output must meet, common mistakes, and what I check before I call it finished.
2. When an answer is vague, ask a follow-up instead of filling the gap yourself. Never invent a standard, a figure or a preference I haven't given you.
3. Stop interviewing only when you could do the job without asking me anything. Tell me when you reach that point and summarise what you've understood in five bullets so I can correct it.
4. Then write the playbook as a SKILL.md file:
- YAML frontmatter with `name` (lowercase, hyphens, under 64 characters) and `description` (under 200 characters, saying what the skill does and exactly when to use it).
- Sections: Purpose, Inputs, Steps (numbered), Standards, Common mistakes, Done checklist (tick-box items specific to this job), and one short example of good output.
5. Before you show me the file, check: every step is something you could act on; the checklist items can each be answered yes or no; nothing in it came from you rather than me. List anything you were unsure about underneath the file.
Fill in the four bracketed lines. Mention the bits that feel too obvious; those are what a stranger gets wrong.
Alternatively, Anthropic's public skills repository on GitHub includes a skill-creator skill that runs a similar interview and can set up test prompts.
Read the SKILL.md Claude produced before you install anything. Two fields matter most:
weekly-sales-summary. The folder must have exactly the same name.Then package it. The ZIP must contain the folder, not loose files:
weekly-sales-summary.zip
└── weekly-sales-summary/
└── SKILL.md
The help article writes the filename as skill.md; Anthropic's example repository and the open Agent Skills specification use SKILL.md. Capitals are the safer choice.
If you use Claude Code instead, there is no upload: put the folder at ~/.claude/skills/weekly-sales-summary/SKILL.md for yourself, or in .claude/skills/ inside a project to share it through the repository. You can then call it directly with /weekly-sales-summary.
This is the part people skip, and the whole point. A playbook written from memory is a first draft until real runs correct it.
To make step 3 quick, paste this after a run:
You just completed [JOB NAME] using the [SKILL NAME] skill. Here are my corrections to your output: [WHAT I CHANGED AND WHY].
1. For each correction, decide whether it is a one-off (specific to this week's inputs) or a rule (would apply every time). Explain your reasoning in one line each.
2. For each rule, propose the exact wording to add to or change in SKILL.md, and say which section it goes in.
3. If a correction contradicts something already in the skill, flag the conflict and ask me which wins. Do not resolve it yourself.
4. Do not rewrite sections I haven't corrected.
Output a short table: correction, one-off or rule, proposed edit.
Replace the three bracketed fields. When you're iterating on a skill in a claude.ai chat, the help centre says you can highlight text in the skill file beside the conversation, click "Edit with Claude" and request the change there. Otherwise, edit the file and re-upload.
Once runs stop producing rules, leave the skill alone until the job itself changes.
SKILL.md; use a proper connector for access to outside services.disable-model-invocation: true in the frontmatter stops Claude running a skill unless you call it by name.