AI Guides › Playbooks
By Nigel Guy · 7 min read
You have a folder of bookmarked guides, a notes app full of clipped threads and a few PDFs you meant to read properly. Saving feels like progress because the information is now "somewhere", but nothing in that pile ever runs. The next time you need the thing, you search, skim, half-remember and improvise.
The fix is not a better filing system. It is converting the useful part of a resource into an instruction Claude can follow whenever you ask.
The rule: a saved resource only counts once it has been turned into a named workflow with a trigger, inputs and an output, and you have run it once on real work.
Every guide, note or article that survives your cull gets one card with six fields. Two steps produce it: extract, then install.
| Field | What goes in it |
|---|---|
| Name | A short kebab-case label you will actually type, such as weekly-client-update |
| Trigger | The situation and the phrases that should start it |
| Inputs | What you hand over each time (the variable parts) |
| Steps | The ordered procedure, stripped of the author's anecdotes |
| Output | Format, length and what "done" looks like |
| Stop rules | What Claude must ask about or refuse to guess |
Most resources are one procedure wrapped in stories, screenshots and sales copy. Paste the resource into Claude (or attach the file) and ask it to pull the procedure out and fill the card. Use a prompt like this.
You are a workflow editor. I will give you a guide, set of notes or article. Your job is to turn the useful procedure inside it into a reusable workflow that an AI assistant can run on demand.
Source material:
[PASTE_OR_ATTACH_SOURCE]
What I want to use it for:
[MY_REAL_TASK_AND_WHEN_IT_COMES_UP]
Do this in order:
1. Say in two sentences what the source is actually teaching, and whether it contains a repeatable procedure at all. If it does not, say so and stop.
2. List the steps as a numbered sequence. Drop anecdotes, marketing and anything I will not need for my task.
3. Separate what stays fixed every time from what I will supply each time. Show the variable parts as [SQUARE_BRACKET_PLACEHOLDERS].
4. Fill in this card: Name (kebab-case), Trigger (situation plus three phrases I might say), Inputs, Steps, Output (format, length, a definition of done), Stop rules.
5. Mark any step that depends on a claim, price, setting or product detail in the source that may be out of date, so I can check it before relying on it.
Rules:
- Do not add steps the source does not support. If you think a step is missing, list it separately under "Suggested additions".
- If an input I need to supply is unclear, ask me before writing the card rather than guessing.
- Keep the card under 400 words.
Before you answer, check that every step is something I can act on without reading the original again, and that nothing in the card contradicts the source.
Fill in the source and your real task. The "real task" line matters: it stops Claude summarising the whole resource when you only need one thread of it.
A card in a notes file is just another saved resource. Install it in one of three places, depending on how you use Claude.
| Where | How it works | Best for | Catch |
|---|---|---|---|
| A Skill in Claude (claude.ai) | A folder containing a skill.md file, zipped and uploaded; Claude reads the description to decide when to use it |
Procedures you want available in any chat | Needs code execution enabled; availability and limits can change |
| A Skill in Claude Code | A SKILL.md in ~/.claude/skills/<name>/ (personal) or .claude/skills/<name>/ (project); run it with /name |
Developers and anyone working in a terminal or project folder | Only loads where Claude Code runs |
| A Project | Add the card to the project's custom instructions or knowledge | Work tied to one client, topic or ongoing piece of work | Only applies to chats inside that project |
Installing as a Skill in claude.ai. At time of writing, Anthropic's help centre says skills are available on Free, Pro, Max, Team and Enterprise plans and require code execution. On Free, Pro and Max, go to Settings, then Capabilities, and turn on "Code execution and file creation". Then go to Customize, then Skills, click the plus button and choose "Upload a skill". On Team and Enterprise, an owner has to enable both code execution and Skills under Organization settings, Plugins and skills. Menu names move, so if yours differ, check the current help page linked below.
The file needs frontmatter at the top. Per the help centre, name is up to 64 characters and description up to 200:
---
name: weekly-client-update
description: Drafts a weekly client update from my notes. Use when I say "client update", "weekly update" or paste raw project notes.
---
Steps, output format and stop rules go here, taken from your card.
The folder name should match the skill name, and the zip should contain that folder as its root. The description does the real work, because Claude uses it to decide when to apply the skill. Put the phrases you naturally say in it.
Installing in Claude Code. Create the folder and SKILL.md as above. According to the Claude Code docs, you can run it by typing /weekly-client-update, and adding disable-model-invocation: true stops Claude from starting it on its own. Arguments can be passed with $ARGUMENTS, so /weekly-client-update acme can feed the client name in. Check the Claude Code skills page for the current list of frontmatter fields before you rely on any.
Installing in a Project. Paste the card into the project's instructions. Anthropic's help pages say Projects are available to all users, with a limit of five on the free plan at time of writing, and paid plans get retrieval over larger knowledge bases. Use this route when the workflow only makes sense next to the project's files.
Say you run a small bakery and saved a long thread on writing a weekly supplier-order email. The thread has twelve tips and a story about a failed order.
Step 1 extracts a four-step procedure: check last week's waste, check next week's bookings, set quantities, write the email in a fixed layout. Inputs: [LAST_WEEK_WASTE_NOTES], [NEXT_WEEK_BOOKINGS], [SUPPLIER_NAME]. Stop rule: if bookings are missing, ask rather than estimate.
Step 2 installs it as a skill called supplier-order, with the description "Drafts the weekly supplier order email. Use when I say 'supplier order' or paste waste and booking notes."
On Monday you type: "supplier order, here are the notes", paste two lists, and get a draft in the fixed layout. That is the thirty-second version of the thread you used to reread.
Run the workflow on a real, current task, not a test case. Then check three things. Did it fire on your natural phrasing, or only when you named it exactly? Did it ask for missing inputs instead of inventing them? Could you act on the output without opening the original resource? If any answer is no, edit the card, not the habit. Triggering failures usually mean the description is too vague or uses the author's vocabulary instead of yours.