A written method for one recurring video job — the order to make the calls in, which parts of the answer are measurements, and which parts are judgements that belong to the person reading. An agent fetches one with get_playbook before starting the work.

Why are playbooks not just part of the tool descriptions?

Because tool descriptions are paid for on every connection, and a playbook is paid for only by whoever asks. Everything permanent is about 2,600 tokens; the playbooks together are far more than that, and loading them into every session would tax every user for a method most of them will not use.

Can I use a playbook without an agent?

Yes. They are readable methods, and each one is published here in full. Nothing in them depends on being executed by a model — they are how a competent person would do the job, written down.

Playbooks

A playbook is a method for one recurring job — the call order, what is measurable, and where the judgement belongs to the reader rather than to the tool. The agent fetches them with get_playbook, and clients that support MCP prompts show them as commands before anyone has described their problem.

Last updated 6 August 2026

The playbooks

Each one is published here in full. The when column is the sentence an agent matches a request against, which is why it describes a situation rather than a document.

Available playbooks
Playbook	What it is for	When it applies
Take an ad apart	Break a performing ad into its beats and claims so the reusable part can be named	an ad or piece of marketing is to be taken apart, or called a winner and copied from

Take an ad apartBreak a performing ad into its beats and claims so the reusable part can be named1,285 words

Using one from an agent

Clients that support MCP prompts show these as commands, so a playbook can be picked before anybody has described their problem.

Fetch a method, then do the worktext

get_playbook(task: "ad-teardown")
watch_video(url: "https://www.facebook.com/ads/library/?id=1234567890")

Long playbooks are split into sections, fetched by name and only when asked for — get_playbook(task, section). That keeps the first response small enough to be worth reading in full.

What a playbook will not do

The constraint that makes them worth following rather than just long.

The same rule runs through the engine itself: there is no is_music field and no hook_ends_at. Those are judgements, and the evidence is returned instead — along with the frame the number belongs to.

Questions

What is a playbook?: A written method for one recurring video job — the order to make the calls in, which parts of the answer are measurements, and which parts are judgements that belong to the person reading. An agent fetches one with get_playbook before starting the work.
Why are playbooks not just part of the tool descriptions?: Because tool descriptions are paid for on every connection, and a playbook is paid for only by whoever asks. Everything permanent is about 2,600 tokens; the playbooks together are far more than that, and loading them into every session would tax every user for a method most of them will not use.
Can I use a playbook without an agent?: Yes. They are readable methods, and each one is published here in full. Nothing in them depends on being executed by a model — they are how a competent person would do the job, written down.