The one page that stops your team's AI from inventing a retailer policy

What does a finished one-page playbook look like?

Here is the output for a made-up brand, Harbor Oat Co. Every detail is illustrative, not real data.

| Part | What Harbor Oat Co.'s playbook says |

|---|---|

| Tasks AI does | Drafts replies to customer emails. Writes the Monday sales summary. Turns broker notes into a to-do list. |

| Files it reads | The current price sheet. The retailer terms folder. The list of claims the packaging may make. |

| Needs a person's approval | Any reply that offers a refund. Anything sent to a retailer buyer. Any change to a price. |

| Never touches | Bank and payroll files. Contracts. Anything that states a health benefit. |

That is the whole document. Anyone on the team can read it in a minute, and the AI reads it before it starts work.

What goes in a one-page AI playbook, and what stays out?

Four parts go in, because each answers a question the team would otherwise settle by guessing. Which tasks can I hand over? What should the AI read first? When must I check before it goes out? What is off limits?

What stays out is policy language. A page of principles about responsible AI use gives the AI nothing to act on. "Be careful with customer data" is a worry. "Never open the payroll folder" is a rule. If a line cannot be followed or broken, cut it.

Keep it to one page on purpose. A longer document gets skimmed by people and ignored in parts by the AI. If a section runs long, it is usually two jobs, and each job can get its own short playbook.

What prompt writes the playbook for you?

Before AI, you would book a half day with your operations lead, interview each person on the team about what they do, and type the notes into a document that nobody opens twice. That is a day of two people's time (illustrative, yours will differ), and the result is out of date within the quarter.

Today you let the AI do the interviewing. Paste the text below into Claude or ChatGPT, which are AI assistants you type to, like a chat window. (A prompt is just the instructions you type in.) It asks you questions one at a time and writes the page at the end.

Interview me to write a one-page AI playbook for my consumer packaged goods company.
Ask one question at a time and wait for my answer.
Cover, in this order: what we sell and where; who is on the team and what each person does in a week; which tasks the team already hands to AI or wants to; which files, sheets and folders those tasks need; which outputs a person must approve before they go anywhere; and what AI must never see or do.
Ask a follow-up whenever an answer is vague.
When you have enough, write the playbook with exactly four headings: Tasks AI does, Files it reads, Needs a person's approval, Never touches.
Write short, plain sentences that an AI can follow without guessing.
Keep it under one page, then ask me what is wrong or missing and fix it.

You do the talking and the AI does the writing. If nobody on your team uses AI yet, the playbook is still the right first page: the tasks row becomes the list of what you want to try first, and the approval row stops the first mistake from reaching a buyer.

How does the team's AI learn the playbook?

It reads the page every time it starts work. Paste the playbook at the top of a task. If your AI tool lets you save instructions it follows every time, the playbook can live there instead. The point is that nobody has to remember to tell the AI the rules.

The files in the "Files it reads" row do the same job from the other side. They give the AI your actual prices, terms and claims, so it stops inventing a retailer policy or a price from a channel you do not sell into. Support replies that learn shows one task where the AI improves from what your team sends, and teaching AI your business covers why the AI needs your company's own files.

Who owns updating it when a task changes?

One named person, with a name on the page. A playbook owned by the team is owned by nobody. In a small CPG company that is usually whoever runs operations, because that person sees tasks change first. The upkeep is a few minutes per change, but it is one person's job, so decide who before you decide anything else. The AI CPG Workshop sets up the company files and rules this page points to.

Update it when a task moves, not on a calendar. A new retailer, a new file format or a new task handed to AI each change a row. Tell the AI what changed and ask it to rewrite the page:

Here is our current AI playbook, and here is what changed this week. Update the playbook, keep it to one page, and list the lines you changed so I can check them.

How do you tell if the team is using it?

Ask the AI to check. Take a handful of this week's outputs, such as replies sent, summaries written and files touched, and paste them in under the playbook with these instructions:

Compare these outputs to the playbook above. List anything that was sent without the approval the playbook requires, any file used that is not on the list, and any task done that the playbook does not cover.

A list of gaps tells you one of two things. Either people are skipping the playbook, or the playbook is missing a task the team really does. Both are useful. For where this sits among the stages of using AI, see seven levels of AI for CPG.

Sources

No outside facts are stated in this post.