1. Prepare inputs that justify the requirements
Before using an AI PRD writer, gather the user problem, the decision already made, evidence supporting it, constraints, and unresolved questions. Label observations separately from hypotheses. Do not ask a model to invent customer research to make the document look complete.
For an illustrative appointment-reminder feature, useful inputs might include the existing reminder process, a verified cancellation rule, feedback from authorized research participants, and a decision to test one reminder channel. “Build the best reminder system” leaves all of those choices unresolved.
- Source notes: assign identifiers such as E1 and E2 so claims can be checked.
- Decision: state what has been approved and what remains a proposal.
- Boundaries: include non-goals, affected users, and known constraints.
- Outcome: describe the behavior or result that would make the work useful.
2. Use a prompt that preserves uncertainty
Copy the following prompt into an assistant you are permitted to use. Replace the input block with approved material; exclude private information that is not needed for the task.
You are drafting a product requirements document for human review.
Use only the supplied context. Do not invent research, approvals,
metrics, integrations, or technical constraints.
Inputs:
- User and problem: [insert]
- Current behavior: [insert]
- Evidence with source IDs: [insert]
- Accepted decision and rationale: [insert]
- Desired outcome and proposed measurement: [insert]
- Constraints and explicit non-goals: [insert]
- Unresolved questions: [insert]
First list the questions that could materially change scope.
If those questions are blocking, stop before writing requirements.
Otherwise draft: problem, evidence, goals, scope, non-goals,
numbered requirements, acceptance scenarios, risks, measurement,
rollout considerations, and open questions.
Link each requirement to the decision or evidence that supports it.
Label assumptions and proposed targets explicitly. Separate product
requirements from suggested implementation choices. Do not mark
this draft approved.A missing constraint is not permission to choose one silently. If the assistant cannot establish who may send a reminder, that question should appear before an implementation-ready requirement.
3. Review a generated requirement, not just the prose
Illustrative weak output: “Automatically remind all customers about upcoming appointments.” This does not identify the recipient, timing, permissions, or cancellation behavior.
Reviewed draft: “For appointments opted into the pilot, send one reminder to the verified contact using the approved schedule. Do not send it after cancellation or after the contact opts out. Leave the schedule value unresolved until the product owner confirms it.”
The improvement is a clearer behavioral boundary, not a promise that the AI has validated the policy. Add acceptance scenarios for cancellation, missing contact details, retries, and duplicate delivery. Keep policy-sensitive unknowns blocking rather than hiding them in a footnote.
Use the PRD template if the generated structure omits a necessary section.
4. Run an evidence and scope review
- Open each cited source and verify that it supports the actual claim.
- Check every proposed requirement against the accepted decision. Remove attractive additions that have no authorization.
- Separate desired outcomes from unmeasured claims about current performance.
- Test whether an engineer could distinguish expected behavior from an implementation suggestion.
- Ask a reviewer to describe how each requirement would fail and how that failure would be detected.
- Record approval, remaining questions, and the version approved for handoff.
Reject fabricated quotations, unsupported percentages, and phrases such as “integrates seamlessly” that conceal work the team has not investigated. A draft is useful when it exposes decisions still needed.
5. Turn the approved draft into a controlled handoff
A generated PRD is not automatically an instruction to code. The product owner accepts the behavior and scope; an engineer reviews technical feasibility and adds repository-specific constraints where appropriate. Then prepare a PRD for AI coding agents with explicit acceptance checks and stop conditions.
Return changes that alter the agreed product direction to a decision log. In PM Harness, generation supports the Specify stage; it does not replace Research, Decide, or Verify.
For background on the document's purpose, see Atlassian's explanation of product requirements documents. The prompt and review sequence above are an original working method, not a vendor integration.