AI agent brief template for product owners
A seven-line AI agent brief template for product owners: goal, scope, checks and stopping points, with an illustrative example and a table of common mistakes.
Written by Phan Lam Duy Khoi
An AI agent brief states the goal, the context, what done means, what not to touch, how to check, when to stop and ask, and what to report back. With those seven parts, the agent guesses less and the PO can check the result.
What is an AI agent, and why does it need a brief?
An AI agent is an AI system that plans its own steps and uses tools (reading files, running commands, editing code) to finish a task. Unlike a chatbot that answers one message at a time, an agent acts on your behalf, so it can change the right thing or the wrong thing.
A brief is the handoff you write before the agent starts. For a product owner this is a familiar skill: it is like writing a good ticket for a developer, except the person receiving it is not sitting next to you to ask questions. The agent fills every gap with a guess. The clearer the brief, the less it has to guess.
Anthropic, the company behind Claude, describes agents as systems where the model decides how to do the work and which tools to use. It also recommends stopping conditions and checkpoints for human feedback. The template below turns that advice into seven lines a PO can fill in within a few minutes.
What should an AI agent brief include?
I use seven fields. Each one answers a question the agent would otherwise have to guess.
- Goal: what this is for, and for whom. The agent needs the reason to pick well when it sees two ways of doing something.
- Context: the product, page or screen involved and the docs worth reading. The agent does not know what only lives in your head.
- Done means: a result you can see. Without this line, the agent decides for itself when it is finished.
- Do not touch: files, data and environments that must stay as they are. This is the line that protects you from surprise changes.
- How to check: a test command, steps to click through, screenshots. Agents do better work when they can check their own result.
- Stop and ask before: anything that is hard to undo or that customers will see, such as merging, releasing or deleting data.
- Report back: what changed, what was checked and what was not. The "not checked" part is the one to read most carefully.
The AI agent brief template (copy and use)
Replace the [brackets] with real information. If a line does not apply, write "none" rather than deleting it. A line that says "none" still tells the agent you thought about it.
Goal: [one sentence: what this is for, and for whom] Context: [the product, page or screen involved; why now; links to docs] Done means: [a result I can see, as 2–4 bullet points] Do not touch: [files, data, settings, environments that must stay as they are] How to check: [test command, steps to click through, screenshots you need] Stop and ask before: [merging, releasing, deleting data, changing anything customers see, or when you hit something the brief does not cover] Report back: [what you changed and where; what you checked; what you did not check; open questions]
If your team already keeps standing rules for agents (for example, an instructions file in the repo), the brief only needs what is specific to this task. Do not paste all the standing rules into every brief.
Example: a brief for a small website copy change
The illustrative example below is a very small task: changing the label on the Contact page submit button. Small tasks still need a brief, because button labels are often shared. Changing one place can quietly change another page too.
Goal: The submit button on the Contact page says what the visitor gets, so people know they are booking a consultation. Context: The company's marketing site. Page /contact. The button currently says "Send". The team wants "Book a consultation". The page has a Vietnamese and an English version. Done means: - The /contact button reads "Book a consultation" in English and "Đặt lịch tư vấn" in Vietnamese. - The label does not wrap on a 375px-wide screen. - The screen reader label (aria-label, if there is one) matches the new text. Do not touch: the form submission logic, the email address that receives the form, other pages, deploy configuration. How to check: run the site locally, open /contact in both languages, take screenshots at 375px and 1280px, run the existing tests. Stop and ask before: merging, deploying, changing any other copy, or if the "Send" label is shared with other pages. Report back: files changed, screenshots at both sizes, test results, and what you did not check.
Look at the "Stop and ask before" line. I do not know whether the "Send" label is shared. Instead of guessing, I tell the agent to stop and ask if it is. The scope is also one button only. The agent may spot other copy it thinks should change, but that call belongs to the PO, not the agent.
What are the common brief mistakes, and how do you fix them?
| Mistake | What happens | How to fix it |
|---|---|---|
| The task without the goal | The agent follows the words but misses the point | Add one sentence: what this is for, and for whom |
| A vague "done", such as "make it look better" | The agent decides when it is finished | Write 2–4 visible results, with screen sizes or exact copy |
| No "do not touch" list | The agent edits files or settings you did not expect | List the files, data and environments that must not change |
| No way to check | The agent says it is done and you have nothing to compare against | Give the test command, the steps to click through and the screenshots you need |
| No stopping point | The agent merges, deploys or deletes data on its own | Name the actions that need your approval, especially the ones that are hard to undo |
| Several tasks in one brief | Results get mixed up and are hard to check or roll back | One brief, one outcome; split bigger work into several briefs |
| Skipping the agent's report | Mistakes reach users | Read the "not checked" part first, then try the result yourself |
Most of these mistakes come from the same place: the person writing the brief knows exactly what they want, so they forget the agent does not. A quick test: hand the brief to a colleague who knows nothing about the task. If they need to ask you something, the agent will have to guess it.
When should you not hand a task to an AI agent?
Not every task should be delegated. Anthropic also recommends starting with the simplest approach and adding agents only when they are needed. I keep these tasks for myself:
- Product decisions: setting priorities, cutting scope, saying no to a request. An agent can draft the options; the PO makes the call.
- Tasks where you cannot yet say what "done" looks like. If you cannot write the "Done means" line, clarify the request before you delegate it.
- Anything touching real customer data, money, access rights or production systems, when there is no test environment and no reviewer.
- Tasks so small that doing them is faster than writing the brief. A typo fix does not need seven lines.
- Tasks nobody can check. If no one can verify the result, do not let an agent do it and simply trust it.
My rule: AI does the work, the PO checks, the PO decides. A good brief makes the first part faster. The other two are still yours.
Frequently asked questions
How is an AI agent brief different from a normal prompt?
A prompt asks for an answer. A brief hands over a task the agent will carry out over several steps, so it also needs boundaries: what not to touch, how to check, and when to stop and ask.
How long should a brief be?
Long enough that a new teammate could do the task without asking you anything. For a small task, the seven lines of the template are usually enough. If the brief runs past a page, the task probably needs splitting.
Does this template work with Claude Code, Codex or Cursor?
Yes. It does not depend on a tool. Each tool has its own place for standing rules, but the goal, scope and checks for each task still need to come from you.
The agent says it is done. Do I still need to check?
Yes. Read what the agent says it did not check first, then open the result yourself in the real environment. Passing tests do not prove that users see the right thing.
Sources to check
Related articles
- 5 AI prompts I reuse every sprint as a PO: Five copy-ready prompts for research, user stories, acceptance criteria, edge cases and AI agent briefs, each with the check a PO should never skip.
- Using Claude as a PO: where it helps, what you still do: AI drafts fast. This article shows four mistakes to check and a checklist to use before a document reaches the team.
- Can AI replace a Product Owner? What still needs a PO: Can AI replace a Product Owner? The short answer is not yet. This guide separates what AI can draft from what a PO must decide, plus a self-check for new POs.
- What data should you keep out of AI tools at work?: Six kinds of data a PO should keep out of AI tools, how to filter each piece before pasting, an anonymised complaint example and a ready-to-use checklist.