Turn authorized customer evidence and a bounded product problem into a decision brief, a small requirements document and testable acceptance criteria. Use AI to organize competing interpretations while keeping prioritization, delivery estimates and approval with your team.
A traceable product packet connecting the user problem, evidence, proposed scope, excluded work, acceptance criteria and open decisions. It is a proposal for discussion, not a committed roadmap or automatically created engineering backlog.
Before you start
A specific problem and evidence
Supply a short problem statement, dated research excerpts and relevant observed behavior. Give each excerpt an ID. Remove customer identifiers and confidential details before sharing. State how the evidence was collected so a support anecdote is not treated as a representative survey.
Goals and non-goals
Identify the intended user outcome, business constraint and smallest acceptable scope. Include what must not change, known dependencies and the decision owner. If success metrics have not been agreed, label them proposed instead of letting a model invent a target.
A reviewable output contract
Ask for a one-page brief or a small set of stories rather than an unlimited backlog. Decide which fields you need: evidence ID, proposed behavior, exception, acceptance check, owner and unresolved question. Supply existing terminology to avoid introducing conflicting feature names.
Practical guidance
From input to reviewed output
1
Separate needs from requested solutions
Ask chat to summarize what people are trying to accomplish, what blocks them and which excerpts support each observation. Keep requested features in a separate column so a suggested solution does not become the problem definition.
Checkpoint: Trace each theme to evidence. A frequently repeated quotation is not necessarily a frequently occurring user problem.
2
Compare bounded options
Develop two possible responses, including a smaller experiment where useful. Compare user benefit, unknowns, dependencies and reversibility. Supply any effort estimates from the delivery team; do not ask AI confidence to stand in for engineering estimation.
Checkpoint: List what would change your decision and which assumption deserves investigation first.
3
Draft a narrow requirement set
Choose one direction and write the intended behavior, exclusions and major states. Convert it into stories with observable acceptance criteria. Include cancellation, missing data or permission failures where relevant rather than documenting only the happy path.
Checkpoint: A tester should understand the expected result without interpreting words such as seamless, intuitive or fast.
4
Review with design and engineering
Share the proposal in your team's normal tools. Return authorized review notes to chat for a targeted revision. Keep each decision linked to its rationale and avoid regenerating approved sections when only one dependency changed.
Checkpoint: Unresolved architecture, access, data and scheduling questions should remain visible, not be silently answered by the model.
5
Prepare the decision and learning record
Record what was approved, deferred or rejected, who decided and the next evidence-gathering step. Draft a short stakeholder update that distinguishes planned work from released behavior. Implementation and measurement happen through the team's existing process.
Checkpoint: Do not call an experiment successful until someone records its actual method, observations and limitations.
Technical details
What the workflow does—and does not do
GOV.UK service-design guidance
Needs require evidence, not synthetic consensus
GOV.UK advises grounding user needs in research and describing the problem rather than assuming a solution. Apply that distinction to your supplied material: an AI-generated persona or imagined quotation is a hypothesis, not a new research participant. Preserve dissenting evidence instead of smoothing it into consensus.
The user-story guidance separates the actor, desired action and goal from acceptance outcomes used to check completion. Use this as a drafting structure, then have your team verify edge cases and testability. A correctly formatted story can still describe the wrong need or an infeasible change.
This workflow does not establish Jira, Linear, analytics or research-repository connections. Supply authorized context and move reviewed output into your existing systems yourself. Story IDs, owners and statuses in a response are text until someone creates or updates the actual records.
Editorial review practice
Preserve uncertainty in scores and dates
If your team uses a prioritization formula, provide the definitions, inputs and units. Recalculate it outside chat and retain sensitivity to uncertain inputs. Do not present generated effort estimates, adoption forecasts or dates as delivery commitments; a missing dependency can invalidate the whole comparison.
Technical details
Choose by task, not by a universal ranking
These are editorial starting points, not measured comparisons. Check the exact model guide and current app access before using a model; linked catalogs may also contain historical entries.
Requirements and tradeoff drafting
Evaluate a chat model on the same evidence packet and ask for a concise decision brief with explicit gaps. Check whether it preserves non-goals and separates a customer observation from its own proposed explanation.
Tradeoff: A persuasive rationale is not customer validation, and more detailed output can make unsupported assumptions harder to spot.
Use another available chat model to review acceptance criteria and identify ambiguous states or hidden dependencies. Give it the approved scope instead of inviting a new product strategy.
Tradeoff: Two models agreeing does not confirm feasibility. Design, engineering and research owners still need to evaluate the proposal.
Check the exact model's input support before evaluating more context. Organize excerpts by ID and date, and request an evidence ledger alongside any summary.
Tradeoff: Provider context limits do not prove complete recall, document-upload support or automatic access to your research repository.
Replace every bracketed placeholder with authorized context. These are original templates, not generated-output demonstrations or generation-tested recipes. Copying does not submit anything or incur a generation charge.
Create an evidence-led product brief
Where to use it: Paste into AI Chat with sanitized, labeled research excerpts.
Create a one-page product brief for [USER PROBLEM]. Evidence: [DATED EXCERPTS WITH IDS]. Goals: [OUTCOMES]. Constraints and non-goals: [BOUNDARIES]. Separate observations, interpretations and assumptions. Propose one small experiment and one larger option, with dependencies and unanswered questions. Cite excerpt IDs for factual claims. Do not invent customer quotes, usage counts, effort estimates or a delivery date.
Check: Read the original excerpts and confirm the brief preserves uncertainty and contrary evidence.
Refine the selected scope into testable stories
Where to use it: Paste into AI Chat after the team chooses an approach.
Refine [APPROVED APPROACH] using [TEAM REVIEW NOTES]. Produce at most [NUMBER] stories with actor, need, goal, evidence ID and observable acceptance criteria. Include [REQUIRED FAILURE OR EMPTY STATES]. Preserve [NON-GOALS]. Label missing technical decisions and estimates as unresolved. Do not create tickets, assign people or change approved priorities. Finish with only the questions that block a useful handoff.
Check: Ask a teammate to describe how each criterion would be tested without further clarification.
Review the decision packet
Where to use it: Paste into AI Chat with the current brief, stories and decision notes.
Review [PRODUCT PACKET] against [SOURCE EVIDENCE] and [APPROVED SCOPE]. Find contradictions, unsupported demand claims, untestable criteria and hidden dependencies. For each issue, quote the affected passage and identify the relevant evidence ID or missing fact. Separate blockers from optional improvements. Summarize approved decisions, assumptions and next checks without claiming implementation, launch or experiment success.
Check: Confirm the review did not quietly reintroduce work that the team explicitly excluded.
Illustrative example
What a useful handoff could contain
An editorial example of structure, not a generated result, customer case study or measured outcome.
Illustrative problem record
For a fictional workshop-booking service, participants cannot tell whether a booking change was saved. Evidence fields reference supplied session notes; frequency remains unknown. The proposed goal is clearer confirmation, not a redesign of the booking system.
Bounded story and check
A participant changes an existing booking and receives an explicit result. Proposed checks cover a successful save, unavailable slot and failed request. Exact interface behavior and technical feasibility remain team-review items, not claims about an implemented feature.
Decision handoff
Record the chosen approach, rejected alternative, unresolved dependency and research follow-up. Keep proposed success measures in a separate field from observed results. The packet should be understandable without access to the original chat session.
Technical details
Check the result before relying on it
Quality checklist
Every claimed user problem links to supplied evidence, with sample limitations retained.
Acceptance criteria describe observable results, including relevant failure states.
Priority scores and estimates retain their inputs, units, assumptions and human owner.
The handoff distinguishes approved scope, unresolved decisions, planned work and actual releases.
When the result misses the brief
A list of feature requests becomes a roadmap
Return to the underlying user problem and ask which evidence supports each proposal. Add team-owned prioritization criteria before choosing work; a generated ordering is not approval.
The brief keeps expanding
Fix the maximum scope, restate non-goals and ask for a revision of only the affected section. Move useful but unrelated ideas into a clearly deferred list.
Acceptance criteria repeat the headline
Ask what a reviewer would see in a normal, empty and failure state. Replace subjective adjectives with an observable behavior the team can actually test.
Provider maximums, picker estimates and actual customer charges are different things. Read the pricing explanation.
Common questions
Before you start the workflow
Can AI decide our roadmap priorities?
It can organize supplied criteria and challenge assumptions, but this guide keeps prioritization and commitments with the product team. It has not validated demand, engineering effort or business impact.
Will these stories appear in our project tool?
No project-management integration is established here. Review the text, then create or update records through your normal workflow. Check actual IDs and statuses after the handoff.
Can we use generated personas as research?
Use them only as labeled discussion aids or hypotheses. They are not interviews, observations or evidence of what real customers need. Validate important assumptions through appropriate research.
The linked technical references were reviewed on . Documentation and app availability can change.
The workflow, templates and fictional example are editorial guidance. Official sources support the attributed research and story-writing principles, not EZ Ai Assist integrations, product decisions or measured delivery outcomes. No customer research or model benchmark was performed for this page.
Opening EZ Ai Assist does not prefill these templates or automatically run the workflow. Choose the relevant mode and review your input before submitting it.