How to write a WordPress playbook an AI agent can follow

To write a WordPress playbook an AI agent can follow, turn each line of your checklist into a numbered step that names the tool to use, says how much to read, and says when to stop or ask. A checklist works for a colleague because they fill the gaps from experience. An agent fills them with a guess, so the playbook has to close the gaps in writing.

This post is about the writing, not the setup. For what a playbook is and how to add one, read give an AI agent step-by-step instructions for WordPress. For running one on a limited connection, read run WordPress playbooks in ChatGPT or Claude on limited access.

Takeaways:

  • Rewrite each checklist line as one action, with the tool named by its label, a number where you mean one, and a stopping point.
  • Put “only read” in the first line of a read-only playbook, and make every write wait for a yes to that exact change.
  • Write the fallback for a missing tool into the step. The playbook cannot make a tool available.
  • Test with a dry run before you trust it. A playbook is guidance an assistant may not follow exactly, and it grants no access.
  • Advice marked Recommendation is ours, not plugin behaviour. The plugin facts were checked against its code.

How do I turn my checklist into a playbook?

Rewrite each checklist line as a numbered step with one action. Name the tool by its label, give numbers (how many, how far back), and say what to do when a tool is missing or a step needs a decision. Make any change wait for a yes to that exact change, and fix the shape of the reply.

What changes when an agent reads your checklist

An agent loads your playbook word for word, with no memory of how you do it. Anthropic’s prompting guide puts it this way: “Think of Claude as a brilliant but new employee who lacks context on your norms and workflows” (Claude Docs, Prompting best practices). Its test is to show the prompt to “a colleague with minimal context on the task” and see whether they would be confused.

OpenAI’s guide makes the same point about GPT models, which “benefit from more explicit instructions around how to accomplish tasks” (OpenAI, Prompt engineering). A playbook is that kind of prompt, stored on your site.

A checklist, before and after

Here is a checklist an editor might keep, written the way people write them.

Friday newsletter prep
- Grab this week's posts
- Check they all have an excerpt
- Put the best three at the top
- Send me the list

Every line leaves something open that a person would settle without thinking.

Line What it leaves open
Grab this week’s posts Which dates, which tool, how many, and when to stop
Check they all have an excerpt Whether to write a missing one, and where
Put the best three at the top Best by what, and who decides
Send me the list Sent where, and in what shape

Here is a rewrite for an agent. It uses one read tool and changes nothing.

Use this when the user asks for the Friday newsletter list. Only read. Change nothing.
1. Date range. Use the last 7 days up to today. If you do not know today's date, ask the user before you start.
2. Posts. List published posts with the Get posts tool, 50 per page. They come newest first. Keep the ones dated inside the range and stop paging once you reach an older one.
3. Excerpts. Note each post whose excerpt is empty. Do not write or save an excerpt. List those posts under the table so the user can add one.
4. Top three. Suggest the three posts you would lead with, with one line on why for each. Say they are suggestions. The user decides.
5. If a tool is off or fails, or paging stops before you reach 7 days back, say exactly what is missing. Never fill a gap with a guess.
6. Reply with one table with a row per post: title, link, publish date, and "Excerpt: yes" or "Excerpt: missing". Put the three suggested posts first and mark them.

To use it, tick Get posts under “Abilities it uses”. Call it “Friday newsletter list”, and give it this description: “Lists the last 7 days of published posts for the Friday newsletter, flags missing excerpts and suggests a lead three. Read-only.”

We wrote this pair for this post. It is not one of the shipped starters, and we have not run it on a site, so test it before you rely on it.

Name the tool in the step

The steps name tools by the labels on your Abilities tab. When an agent loads a playbook, Get playbook returns, for each ability you ticked, its name, its tool name, its label and whether this connection can use it, so the agent can match a label in your steps to the tool to call. Recommendation: write “the Get posts tool”, not “grab the posts”.

The “Abilities it uses” picker on the Playbooks tab lists this plugin’s abilities by subject, and you can tick one that is off.

Vague Specific
Look at recent posts List published posts with the Get posts tool, 50 per page
Check the images Page through the Get media tool and note every image with empty alt text
See what is waiting Page through the Get pending comments tool and keep the comments dated inside the range

The specific wording is adapted from the shipped Weekly content report and Content health check starters.

Recommendation: if two tools could do the job, say which one and when. In our 1.10.2 demo run the agents used Get page where a starter said Get post, because the item was a page (the run write-up). A named tool anchors the agent, but it did not stop one choosing its neighbour.

Give numbers and a stopping point

“All the posts” is not a number an agent can act on. The Get posts tool returns 10 items a page unless asked for more, and it clamps a request to 50. A step that says “list the posts” and names no page size reads 10.

The starters say what they mean. Weekly content report says “stop paging once you reach an older one”, and Content health check says “Check the 100 most recent published posts unless the user asks for a different number.”

Dates need the same care. “This week” means nothing until the agent knows today’s date, which is why the Weekly content report starter says “If you do not know today’s date, ask the user before you start.”

Recommendation: for every step that reads a list, write how many, how far back, and what counts as finished.

One job per playbook

The agent sees your playbooks first as an index in the server instructions, the text a server hands a client when a session opens. Each Active playbook gets one line, its slug and its description. The index appears only when both playbook tools are on and the connection can discover both, so the description is all the agent has to match your request against.

Recommendation: one job per playbook, and put when to use it and whether it changes anything in the description. The starters do. Publish checklist reads “Check one draft before it goes live and list what is missing or worth fixing. Changes nothing unless you say yes. Never publishes.” Weekly content report ends “Read-only.” A playbook that does three jobs gets a description that fits none of them.

The limits on titles, descriptions, steps and the number of playbooks are in the guide’s write your own section.

Stop and ask before a write

If a playbook can change the site, name the exact moment it must stop. Publish checklist does it twice. Its Rules block says “Change nothing until the user says yes to that exact change”, and its category step says to add terms “only after the user says yes to those exact terms. Do not create new terms.”

Draft a new post sets a ceiling instead: “Only ever save as a draft. Never publish or schedule, and never change any other post.”

Recommendation: put “Only read. Change nothing.” in the first line of a read-only playbook, as the report starters do. For a playbook that writes, name the write tool in the one step where it is allowed, say what the user must approve, and repeat the limit in a short Rules block at the end.

Asking first matches what the standards say. The MCP specification says “there SHOULD always be a human in the loop with the ability to deny tool invocations” (Model Context Protocol, Tools). OWASP’s advice on excessive agency is to “require a human to approve high-impact actions before they are taken” (OWASP, LLM06:2025 Excessive Agency).

A line in a playbook is still only a request. The switches on the Abilities tab and the connection’s access decide what runs, and the guide’s what a playbook cannot do section covers it.

Plan for a tool the connection cannot reach

Get playbook reports which of the abilities you ticked this connection cannot use. A tool counts as unavailable when it is off, its plugin is not active, the connection’s server left it out, or it is hidden from the connection’s user. Why the agent says a tool is unavailable explains the field.

That report only covers what you ticked, and it tells the agent a tool is missing. It does not tell the agent what to do next, so write that into the step. The starters all do:

  • Weekly content report: “If paging stops before you reach 7 days back, or a tool is off or fails, say exactly what is missing. Never fill a gap with a guess.”
  • Publish checklist: “If a tool is missing or returns an error, say which check you could not do and finish the others.”
  • Draft a new post: “If saving fails or the tool is off, paste the full post into the chat so nothing is lost, and say why it was not saved.”

The three differ on purpose. A report should admit the gap, a checklist should finish what it can, and a writing job should save the work somewhere else.

Recommendation: for each tool a step depends on, decide which of those three you want, and say so in that step.

Fix the shape of the reply

Recommendation: say what the answer looks like. Weekly content report’s last step reads “Reply with one table with three rows: Published, New drafts, Comments waiting.” Two runs then give you something you can compare.

Anthropic’s guide agrees: “Be specific about the output format and constraints you want.” It also recommends “sequential steps using numbered lists or bullet points when the order or completeness of steps matters”, which is the format every starter uses. The Steps box is plain text, so do not rely on markdown to render.

Test it with a dry run

Recommendation: test a new playbook in three passes, and change one thing between passes. First, read the playbook as a stranger: does each step name its tool, a number and a way out? Second, ask the agent for a dry run, in which it describes each step without running it.

Load the friday-newsletter-list playbook. Do not call any tool except List playbooks and Get playbook. For each step, tell me which tool you would call, with what inputs, and what you would say if that tool were missing. Change nothing.

We have not tested that wording across clients, and an agent may still call a tool you did not expect. Treat the answer as a way to spot a vague step, not as proof the playbook works. Compare the tools the agent names with the abilities you ticked.

Third, run it for real on a connection set to the Read only preset, as the guide suggests. A playbook with a write step will have that write refused, which tells you whether the steps and the access match. The run a playbook one step at a time recipe makes the agent report after each step.

After an edit, Get playbook returns your saved steps the next time it is called, so a new chat picks up a changed step. List playbooks and Get playbook read what you saved each time they are called. The index in the server instructions is different: it is built when a session starts and, per the plugin’s FAQ, refreshes when the client reconnects, so a new playbook or a changed description may not show there until then. The Activity Log shows which tools a run called, and how to find what an AI agent did on your WordPress site walks through reading it.

What a playbook cannot do

A playbook is guidance the agent may follow loosely, it grants no access, and the agent reads its plain text word for word. That is why the advice above leans on tests, fallbacks and a yes before any write. The guide’s what a playbook cannot do section covers all three.

Agent Abilities for MCP is a free WordPress plugin that keeps your playbooks on your own site and lets Claude or ChatGPT load them through two read tools. It is best for site owners who already have a routine and want an agent to follow it without pasting the steps into every chat.

Questions people ask

How long should a playbook be?

As short as closes the gaps. The limit is 8,000 characters for the steps, and the five starters use roughly 1,100 to 1,550. A long playbook is a sign it holds more than one job.

Can I paste in the checklist I already have?

You can, since the Steps box takes plain text, but the agent reads it as written. Rewrite each line as an action with a tool, a number and a stopping point first, as in the before and after above.

Do I tick the tools or just name them in the steps?

Do both. The steps name tools in plain words, and the “Abilities it uses” list is what Get playbook uses to tell the agent which tools this connection can use.

Can the agent write the playbook for me?

It can draft one, and you paste the result in, because agents cannot save playbooks. The turn a job you just ran into a playbook recipe does this from the Activity Log.

Where can I see more written playbooks?

The Playbooks library has the five starters and more written the same way, and the Playbooks docs explain every field.