Practical guide
AI product requirements document verification checklist
A product requirements document is ready for a decision only when a reviewer can trace the problem, understand the chosen scope, test the requirements, and see what remains uncertain. This checklist is for a human review of an AI-assisted PRD. It does not turn a document into an approved product plan by counting ticks. Use it after the author has assembled the source packet and before the team.
A product requirements document is ready for a decision only when a reviewer can trace the problem, understand the chosen scope, test the requirements, and see what remains uncertain. This checklist is for a human review of an AI-assisted PRD. It does not turn a document into an approved product plan by counting ticks. Use it after the author has assembled the source packet and before the team commits to build or release. Record pass, repair, not applicable, or not evidenced beside every item. A critical failure sends the PRD back for repair even when most other rows pass.
Evidence basis and limits
The GOV.UK Service Manual says user stories should identify the actor, need, and goal; acceptance criteria should describe outcomes and link to supporting evidence. Its user-research guidance also recommends retaining the connection between a user need and the stories derived from it. The UK Government Functional Standard for Digital defines traceability as the ability to follow a requirement back to its source and forward through delivery. NASA's Systems Engineering Handbook treats good requirements as necessary, clear, consistent, feasible, verifiable, and traceable. These sources support the review method below; they do not prescribe a single PRD format or approve a particular product choice. GOV.UK: Writing user stories GOV.UK: Learning about users and their needs GovS 005: Digital NASA Systems Engineering Handbook
Sources: [1] [2] [3] [4]
Review header
Complete this block before reviewing the body.
| Field | Entry | | --- | --- | | PRD title and version | | | Decision the PRD supports | | | Source-packet version | | | Author | | | Independent reviewer | | | Product and specialist approvers | | | Review date | | | Intended release or experiment | |
If the decision, source snapshot, or accountable owners are missing, stop. The reviewer cannot determine whether the document is current or who may accept the residual risk.
1. Problem and user evidence
Critical failure: the proposed solution is treated as proof of the problem, or a decision-driving claim cannot be reopened in the source packet.
- [ ] The problem is stated as an observed user or service condition, not a preferred feature.
- [ ] The affected users, non-user stakeholders, and operating context are named.
- [ ] Material statements link to source IDs, dates, and evidence types.
- [ ] Anecdotes are not presented as prevalence estimates.
- [ ] Conflicting or minority evidence is visible rather than averaged away.
- [ ] Accessibility, privacy, security, policy, legal, and operational constraints have named owners.
- [ ] Assumptions and hypotheses are labelled separately from observations.
- [ ] The cost of doing nothing and the current workaround are described.
2. Outcome and decision measure
Critical failure: success is defined only as shipping the feature, or the proposed measure cannot be reproduced from identified data.
- [ ] The desired outcome is distinct from the feature or interface being proposed.
- [ ] The next decision is explicit: continue, stop, narrow, expand, or investigate.
- [ ] Each measure has a population, denominator, time window, source, owner, and interpretation limit.
- [ ] Baseline values are dated, or their absence is recorded.
- [ ] Product behavior, use, artifact quality, service outcome, and longer-term impact are not collapsed into one number.
- [ ] The PRD states which evidence would count against the preferred option.
- [ ] Thresholds are justified for the decision rather than copied from another context.
3. Scope, options, and dependencies
Critical failure: a required dependency or operating responsibility is absent, making acceptance impossible to test in the intended environment.
- [ ] At least the current approach, a process alternative, and the proposed product increment were considered.
- [ ] In-scope, out-of-scope, and unresolved items are separate lists.
- [ ] Every unresolved decision has an owner and due date.
- [ ] Dependencies, interfaces, data flows, migration, support, failure recovery, and exit are addressed where relevant.
- [ ] The document explains why this is the smallest coherent test of the problem.
- [ ] A brainstormed request is not silently promoted into a requirement.
- [ ] Trade-offs identify who receives the benefit and who bears the burden.
4. Requirement quality and traceability
Review every material requirement in a traceability table.
| ID | Requirement | Source or constraint | Acceptance evidence | Downstream test or control | Owner | Result | | --- | --- | --- | --- | --- | --- | --- | | PR-01 | | | | | | |
For each row, check:
Critical failure: a safety-, rights-, security-, privacy-, or decision-critical requirement has no verification method.
- [ ] The requirement has one stable identifier and one observable need or behavior.
- [ ] Conditions and expected results are precise enough for two reviewers to interpret consistently.
- [ ] Necessary implementation constraints are distinguished from solution preference.
- [ ] The requirement is consistent with other rows and uses defined terms.
- [ ] Backward trace identifies evidence or an approved constraint.
- [ ] Forward trace identifies the test, inspection, calculation, rehearsal, or approval that will verify it.
- [ ] Normal, boundary, failure, recovery, and misuse cases are included when material.
- [ ] Changes preserve the previous decision and explain the reason, approver, and affected tests.
5. Acceptance and release evidence
[ ] Acceptance criteria describe observable outcomes rather than words such as “intuitive” or “accurate.”
- [ ] Representative cases are selected before results are known.
- [ ] Critical failures are named and are not diluted into an average score.
- [ ] User acceptance, technical verification, specialist review, and business authorization are separate decisions.
- [ ] The reviewer can see test versions, inputs, results, exceptions, and repairs.
- [ ] Accessibility and degraded-mode evidence are included where relevant.
- [ ] Residual uncertainty and intended-use limits appear in the release record.
- [ ] Telemetry, support feedback, incident review, and the next decision date have owners.
6. AI-assistance audit
[ ] The approved source boundary for AI use is documented.
- [ ] Personal, confidential, regulated, or proprietary material was handled under the applicable policy.
- [ ] Material AI-generated statements link to sources that a reviewer reopened.
- [ ] The reviewer checked quotations, calculations, citations, names, dates, and inferred requirements.
- [ ] Rejected AI suggestions and important human changes remain visible in the decision record.
- [ ] AI did not invent user evidence, approval, a threshold, or a requirement owner.
- [ ] A qualified human remains accountable for prioritization, acceptance, and release.
Worked review: a fictional maintenance-triage PRD
A facilities team proposes an AI-assisted intake summary for maintenance requests. The source packet contains anonymized requests, coordinator interviews, accessibility findings, routing rules, and security constraints. The PRD has 28 checklist passes and three repairs. That count is not the decision.
The problem evidence passes: coordinators repeatedly need to find missing location, asset, and impact fields. The outcome section needs repair because “80% automation” has no denominator and confuses extraction with final routing. The team replaces it with two measures: the proportion of eligible requests with all stated fields surfaced, and the rate at which reviewers correct an extracted field. Each measure now has an eligible population, source, review window, and owner.
Requirement PR-04 fails critically. It says the tool “assigns an appropriate priority,” but the interviews disagree about how access failures compare with water leaks, and no approved rule set exists. The reviewer does not accept a fluent AI-generated priority rubric. The team narrows the increment: the tool extracts stated facts, marks absent fields as “not stated,” and a coordinator owns priority and routing.
The acceptance set includes ordinary requests, conflicting descriptions, an inaccessible attachment, unnecessary personal information, and an extraction outage. One case converts “near the east entrance” into a precise building zone. That unsupported inference triggers repair even though the other cases pass. The PRD is accepted only after the requirement, prompt, test, and release limitation all reflect non-inference.
Disposition record
End the review with one disposition:
Record the reasons and evidence, not just the label. A later reviewer should be able to reconstruct why the team proceeded.
- Accept: no critical failure remains; repairs are closed; named approvers accept the recorded limits.
- Accept with condition: the PRD supports only a bounded experiment, with a condition, owner, and date recorded.
- Repair: a resolvable evidence, clarity, traceability, or test gap remains.
- Hold: the decision lacks essential evidence, authority, or an acceptable risk treatment.
- Retire: the problem, option, or timing is no longer current.
Two Prune boundary
Two Prune can assess whether a participant frames a product problem, navigates supplied evidence, uses AI thoughtfully, writes testable requirements, catches unsupported 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 Manual: Writing user stories
- GOV.UK Service Manual: Learning about users and their needs
- UK Government Functional Standard GovS 005: Digital
- NASA Systems Engineering Handbook
Sources: [1] [2] [3] [4]
Sources
- 1.GOV.UK: Writing user stories · UK Government
- 2.GOV.UK: Learning about users and their needs · UK Government
- 3.GovS 005: Digital · UK Government
- 4.NASA Systems Engineering Handbook · www.nasa.gov