Practical guide
How to create a product requirements document with AI
An AI-assisted product requirements document should make a product decision testable, not make an uncertain idea sound settled. Use AI to organize controlled evidence, identify gaps, compare statements, and challenge ambiguity. Keep people accountable for the problem definition, priorities, trade-offs, acceptance criteria, and release decision.
An AI-assisted product requirements document should make a product decision testable, not make an uncertain idea sound settled. Use AI to organize controlled evidence, identify gaps, compare statements, and challenge ambiguity. Keep people accountable for the problem definition, priorities, trade-offs, acceptance criteria, and release decision.
Start with an evidence packet, not a blank prompt
Before drafting, assemble a versioned packet with the product decision, user research, service or business evidence, current workflow, constraints, dependencies, applicable policies, known risks, open questions, and named owners. Label each item as observed evidence, stakeholder input, approved constraint, assumption, or hypothesis.
The GOV.UK Service Standard says teams should understand users and the whole problem rather than begin with a preselected solution. The 2026 UK Government Functional Standard for Digital says user needs are independent of a proposed solution and defines two-way traceability between a requirement and its source. NASA's Systems Engineering Handbook likewise treats requirements as a basis for acceptance and calls for necessary, precise, consistent, testable, and bidirectionally traceable requirements. These sources support disciplined requirements work; they do not prescribe one universal PRD format. GOV.UK: Understand users and their needs GovS 005: Digital NASA Systems Engineering Handbook
Sources: [1] [2] [3]
Use the PROVE structure
PROVE keeps the document tied to five decisions: problem, result, options, verification, and evolution.
### P — Problem and people
State who experiences the problem, in what context, what they are trying to achieve, and what evidence shows the problem exists. Separate users from stakeholders and purchasers. Include affected people who may not directly operate the product.
Record:
AI can cluster research notes or draft a problem statement from approved sources. A human reviewer must reopen decision-driving evidence and check whether minority, contradictory, or inaccessible experiences were erased during synthesis.
### R — Result and decision measure
Describe the outcome in user and operational terms before specifying features. Define the decision that the first release should enable and the conditions that would show the approach is unsuitable.
For each result, record a baseline if one exists, a measure definition, population, time window, source, owner, and interpretation limit. A metric is not an acceptance criterion unless the team can observe it within the release decision. Distinguish product behavior, service outcome, adoption, quality, and longer-term impact.
### O — Options, scope, and trade-offs
Document at least the current approach, a non-product or process option, and the proposed product increment. Explain why the chosen scope is the smallest coherent test of the problem, not merely the smallest build.
Create three lists:
Record dependencies, interfaces, data flows, operating support, migration, failure recovery, and exit. AI can compare source statements and propose missing categories, but it must not silently turn a brainstormed feature into an approved requirement.
### V — Verifiable requirements and acceptance
Give every requirement a stable ID. Write one observable need per row and trace it backward to evidence or an approved constraint and forward to a verification method.
| ID | Requirement | Source and rationale | Priority | Acceptance evidence | Owner | Status | | --- | --- | --- | --- | --- | --- | --- | | PR-01 | | | | | | |
A useful requirement states the actor or product behavior, conditions, and observable result without prescribing unnecessary implementation. Acceptance evidence might be a usability task, accessibility inspection, security test, calculation, interface test, or operational rehearsal. Include normal, boundary, failure, and recovery cases where they matter.
Do not write “works accurately” or “is intuitive.” State the relevant population, scenario, measure, threshold, review method, and critical failure. Keep policy approval, user acceptance, technical verification, and business authorization separate.
### E — Evolution, ownership, and release record
Record requirement changes, affected evidence, alternatives considered, approval, and implications for tests or prior decisions. Define telemetry, user research, support feedback, incidents, and review cadence for the released increment. Name who can accept, reject, pause, or rollback it.
The release record should state the PRD version, source snapshot, unresolved items, test evidence, limitations, approvers, release conditions, and next decision date. A PRD is a living decision record, but frequent editing should not erase what the team believed at the time of an approval.
- the current journey or workflow and its failure points;
- source IDs for user observations, service data, support evidence, or research;
- frequency and consequence without turning anecdotes into prevalence estimates;
- accessibility, privacy, security, policy, legal, and operational constraints; and
- uncertainties that research must resolve.
- in scope now — capabilities required for the stated decision;
- explicitly out of scope — useful ideas deferred with a reason; and
- unresolved — decisions with an owner and due date.
AI drafting protocol
Use AI in bounded passes rather than one request to “write the PRD.”
Do not put personal, confidential, regulated, or proprietary information into an unapproved tool. Retain prompts or outputs only when the approved information-handling process calls for them.
- Inventory: ask it to list supplied sources, dates, authors, evidence types, and missing items without synthesizing.
- Contrast: ask it to identify agreements, contradictions, unsupported requests, and terms used inconsistently.
- Draft: generate one PROVE section at a time with source IDs beside material statements.
- Challenge: ask for solution assumptions, affected users, failure cases, implementation details disguised as needs, and untestable language.
- Verify: reopen original evidence, run independent requirement and acceptance reviews, and record accepted and rejected AI suggestions.
Worked PRD: source-linked maintenance request triage
A fictional facilities team wants a desktop tool that helps coordinators triage building-maintenance requests. The packet contains 24 anonymized request examples, six coordinator interviews, current routing rules, accessibility findings, service-level data, and security constraints.
The first AI draft proposes automatic priority scoring and direct assignment. PROVE exposes two problems. The user evidence supports faster identification of missing details, but not automated priority. Interviews also disagree about whether water leaks always outrank access failures. The team narrows the first increment to extracting stated location, asset, impact, and missing fields; a coordinator retains priority and routing decisions.
The result is defined as a reviewable intake summary that reduces avoidable follow-up without hiding uncertainty. PR-01 requires every extracted field to link to the exact request text. PR-02 requires the tool to display “not stated” rather than infer a location. PR-03 requires full keyboard operation. PR-04 requires the coordinator to confirm the summary before it enters the work-order system. A fifth requirement covers degraded mode when extraction is unavailable.
Acceptance uses 18 unseen synthetic requests: ordinary cases, conflicting descriptions, an image without alternative text, a request containing unnecessary personal information, and a service outage. The first test fails because “near the east entrance” becomes a precise building zone not present in the request. The team changes PR-02, adds a non-inference test, and returns the build for repair. It does not average the critical traceability failure into an overall usability score.
The release record preserves the deferred automated-priority proposal, names the evidence required to revisit it, and schedules a four-week review of missing-field corrections, coordinator overrides, accessibility findings, and support issues. The worked example demonstrates a requirement decision; it does not establish that the fictional tool improves live service outcomes.
PRD release checklist
[ ] The problem and affected users are supported by identifiable evidence.
- [ ] User needs are not written as a predetermined solution.
- [ ] Outcomes, product behavior, adoption, quality, and impact are not collapsed.
- [ ] Scope includes alternatives, exclusions, dependencies, failure, recovery, and exit.
- [ ] Every material requirement has a stable ID and backward and forward trace.
- [ ] Acceptance methods use representative conditions and identify critical failures.
- [ ] Assumptions, contradictions, missing evidence, and unresolved decisions are visible.
- [ ] AI contributions and material human decisions can be distinguished.
- [ ] Accessibility, privacy, security, legal, policy, and operating reviews have owners.
- [ ] Version, change history, approvers, limitations, and next decision are recorded.
Two Prune boundary
Two Prune can assess whether a participant frames a product problem, navigates supplied evidence, uses AI thoughtfully, writes testable requirements, verifies claims, and communicates trade-offs in a realistic simulation. It reports observed assessment performance. It does not approve a product strategy, certify a PRD, replace user research or specialist review, predict workplace performance, or make employment decisions.
Sources
GOV.UK Service Standard: Understand users and their needs
- UK Government Functional Standard GovS 005: Digital
- NASA Systems Engineering Handbook
Sources: [1] [2] [3]
Sources
- 1.GOV.UK: Understand users and their needs · UK Government
- 2.GovS 005: Digital · UK Government
- 3.NASA Systems Engineering Handbook · www.nasa.gov