How to write a Playbook: goal, role, boundaries and flow

A Playbook is the written set of instructions your Digital Mind follows in every conversation. It has four parts: Goal, Role & style, Boundaries and Flow. Write each in plain, specific sentences, tag the people, Specialists and Data Collections it needs with @, and test it with real questions before you rely on it.

What does a Playbook do that Knowledge doesn't?

Knowledge holds facts: class times, policies, the answers you have already written. The Playbook holds the rules for using them: how your Mind sounds, what it asks, what it refuses and who it hands over to. A fact copied into the Playbook has to be maintained in two places, so keep facts in Knowledge and let a Playbook step say when to use them. For the wider picture, see what a Digital Mind is.

Specificity matters because your Mind can only follow what you have written down. Anthropic's prompting guide (opens in a new tab)^[1] tells developers to "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows." Its test: give the prompt to a colleague who has little context and ask them to follow it. A confused colleague means a confused model. OpenAI's prompt engineering guide (opens in a new tab)^[2] says GPT models "benefit from very precise instructions". Both guides are written for developers, not for Plandora's editor, but the principle carries over.

What belongs in each of the four parts?

Goal

Name the visitor and the outcome in two or three sentences. "Help new customers choose a class" is a goal; "be helpful" is a mood. Prefer one main outcome, such as answering an enquiry, qualifying a need or helping someone book.

Role & style

Describe how a good colleague would sound, then add what you can measure: reply length, questions per reply, whether to use the visitor's first name. "Ask one question at a time" can be checked in a transcript. "Sound professional" cannot.

Boundaries

Write what your Mind must never do and what it should do instead. A bare "never" leaves a gap when it matters. Add the exit: never promise a refund, and pass the request to @Sam. Anthropic's guidance on reducing hallucinations (opens in a new tab)^[3] recommends explicitly allowing the model to say it does not know, which "can drastically reduce false information", though the page notes such techniques reduce errors without eliminating them. AI replies can be wrong, so put financial, legal, medical and other consequential decisions behind a handoff to a person.

Flow

Write numbered steps with one job each: ask, check, consult, collect, hand off. Tag a Specialist where a step needs focused help, a person or team where a handoff goes, and a Data Collection field where a detail should be asked for. Scheduled follow-ups belong to Automations, which work best when they follow your judgment.

How do you tag Specialists, people and Data Collections?

Type @ in a Flow step, or use Add info, and pick from the list. Typing a name as plain text does not create a tag, so "Sam" written by hand points at nobody.

Can you start from a template?

Yes. An empty Playbook offers Start from a template, or Write with AI, which drafts a Playbook from a short description of what your Mind should do. Two templates to know:

• Receptionist ("Greet, understand intent, route to the right team"): it sounds warm, keeps replies to two short sentences, routes by intent to @Support, @Sales or @Team, and never quotes delivery dates or policies from memory.
• Retail advisor ("Guide shopping and product questions from a store, collection, or product"): it is curious first and sales second, offers at most three options at once, and never invents stock or ingredients.

Three more sit in the picker: Lead qualifier, Founder mentor and Support first. Confirming a template replaces every section, so apply one before you write your own text. Then swap placeholder names such as @Support for the real person, team or Specialist, and edit the rest for your business.

What is the difference between a weak and a specific instruction?

A specific instruction names an action, a limit or a trigger. A weak one names a quality.

Anthropic's guide adds one more habit: give the reason. Its example swaps a bare ban on ellipses for the reason (a text-to-speech engine will read the reply aloud and cannot pronounce them), because the model can generalise from an explanation. Do the same in a Playbook: say why a rule exists.

What does a finished Playbook look like?

This is an example, not a customer: Harbour Lane Studio is a made-up pottery school. In the editor, each @ name below would be picked from the list so that it becomes a real tag: @Class Finder is a Specialist, @Booking details is a Data Collection and @Sam is a person.

Goal. You're speaking to people who want to try pottery or book a class at Harbour Lane Studio, a small pottery school. Most are beginners. Help each person find a suitable class and book it.

Role & style. Warm and plain-spoken, like a teacher at the front desk. Keep replies to two or three sentences and ask one question at a time. Avoid pottery jargon unless the visitor uses it first.

Boundaries. State class details only if they appear in the knowledge sources. If a date or policy isn't there, say so and offer to pass the question to @Sam. Never promise a refund or a place on a full class. If the visitor is upset or asks for a person, hand off to @Sam with a one-line summary.

Flow.

1. Ask what they would like to make and whether they have used clay before.
2. Consult @Class Finder with their answers and suggest up to two classes, saying briefly why each fits.
3. Ask for @Booking details one field at a time, starting with preferred date.
4. Confirm what you have collected in one line, then hand the booking to @Sam and say when to expect a reply.

Class times and the refund policy are absent on purpose: they live in Knowledge.

What are the most common mistakes?

1. Storing facts in the Playbook. Class times and policies belong in Knowledge, where they are updated once.
2. Describing a mood instead of a behaviour. "Warm" cannot be tested. Two short sentences and one question at a time can.
3. Writing a "never" with no exit. Without an alternative, your Mind improvises. Pair each prohibition with a handoff or a plain refusal.
4. Assuming a name or a Specialist is wired up. A hand-typed @Sam is not a tag, a Specialist that is Off or untagged is never called, and a field nobody tags is never asked for.
5. Writing the whole Flow before testing any of it. Start with three short steps, test them, then add more.

How long can a Playbook be, and when do changes go live?

Each section allows 4,000 characters, the four together allow 12,000, and a Playbook can hold up to 50 mentions. These are characters, not words, as of October 2026. An edit that breaks a limit is listed as an issue to fix, shown in the header as "N to fix", and does not go live until you fix it.

Ordinary edits need no Publish step. Valid edits save and go live automatically, and the header shows Live. If it shows Draft with a Go live button instead, the saved changes have not been published yet, so select Go live. If it shows Changed elsewhere · Reload, the Playbook was changed in another window: reload before saving again, and copy out any unsaved edits first, because reloading discards them. Because edits go live as they save, change one thing at a time and test straight after.

How do you test a Playbook before you rely on it?

Use the chat preview and rerun the same short list of questions after every change.

1. Send a normal request and check reply length, tone and one question at a time.
2. Try something that should reach a tagged Specialist, then check its Runs for the consult.
3. Raise something that should be handed off. Check the Inbox (availability depends on your plan or rollout) and select Take over to reply yourself.
4. Test a question your Knowledge does not cover and confirm your Mind says so instead of guessing.
5. Give the Playbook to a colleague who does not know the business and ask what they would do if a visitor demanded a refund. Rewrite the lines where they hesitate.

Anthropic's guide to building evaluations (opens in a new tab)^[4] says tests should mirror your real task and include edge cases, and OpenAI's guide recommends tests and evaluation suites for prompt behaviour. For a Playbook, that means a saved list of real visitor questions, awkward ones included. When a reply goes wrong, why AI replies go wrong helps you tell a missing fact from a missing rule.

How does Improve suggest changes?

Improve is a button in the Playbook header, shown once the Playbook has content or there are suggestions waiting. Its panel is headed "From real chats in the last 7 days" and lists suggestion cards, each labelled by kind, such as Unanswered questions, Weak answers or Hand-offs. Each card has a fix button and a Dismiss button.

Choose the fix on a suggestion and Plandora drafts edits to the relevant sections, shown in place with the old text struck through and the new text highlighted. Accept or reject each change, or select Accept all or Discard. Accepted changes save and go live like any edit. A change is skipped if you edited that section after it was drafted: Accept is switched off for it and Accept all leaves it out. Some suggestions open the Knowledge FAQ with the questions filled in instead, since a missing fact belongs there. You are editing live instructions, so read each change and rerun your test questions.