How to Turn User Interviews into an Evidence-Linked PRD

Download Hemory
On this page

You have run six user interviews and written a one-page summary. A week later, an engineering lead asks where requirement 7 came from, and a designer asks whether anyone said that in their own words. The summary cannot answer either question. The wording is gone, the context is gone, and you are left defending a requirement from memory.

This guide shows how to build a product requirements document (PRD) in which each requirement keeps its evidence: the quote, the interview it came from, the situation around it, and the questions that are still open.

TL;DR

  • Give every requirement an evidence reference: interview ID, timestamp, the participant’s words, and what they were doing when they said it.
  • Keep three layers apart: what the participant said or did, what you infer from it, and what you propose to build.
  • Write acceptance criteria from the observed need, and label the requirement “Proposed” until stakeholders approve it.
  • Keep open questions next to the requirement they affect, including single-source claims and conflicts between interviews.
  • If your interviews were captured in Hemory, its built-in Q&A or a connected AI agent can retrieve the relevant passages and draft from them. You still supply the current goal, add anything new, and review every line before it goes to stakeholders.

What an evidence-linked PRD contains

A normal PRD entry says what to build. An evidence-linked entry also says why you believe it and how far that belief goes. Each requirement carries these fields:

FieldWhat it holdsWhy it matters
RequirementOne statement of the need or behaviorThe thing stakeholders approve
EvidenceInterview ID, timestamp, short quote, participant roleLets anyone check the claim
ContextWhat the participant was doing or responding toPrevents a quote from being read out of place
ObservedWhat the participant said or didSeparates fact from interpretation
InterpretationWhat you conclude from itMakes your inference visible and challengeable
Acceptance criteriaTestable conditions for “done”Turns the need into something verifiable
Open questionsWhat the interviews did not settleStops a guess from reading as a finding
StatusProposed, accepted, deferred, or rejectedKeeps the draft honest about approval

A generated or hand-written PRD is a draft. It becomes a decision only when the people with authority accept it.

Step 1: Gather the interviews and decide how you will read them

Start by listing the interviews and giving each a short ID, such as I1 to I6, with the date, the participant’s role, and what the interview was about. Use roles rather than names in the PRD where possible. That keeps the document shareable and still lets you find the original.

Then decide where the words come from, because that changes the next step:

  • The interview was captured in Hemory. Hemory produces a timestamped transcript with speaker labels, and its Q&A can answer questions about your captured conversations. A supported MCP-connected agent can also retrieve them. Name the speakers in each transcript first, so “participant” and “interviewer” are not left anonymous. Where Hemory shows a suspected match such as “possibly” a named person, treat it as a hint and confirm it.
  • The interview was recorded or noted elsewhere. Use that transcript or your notes. Hemory cannot reconstruct a conversation it never captured, and the method below works the same way with any transcript you can read and cite.
  • Some interviews fall into each group. Use the best available source for each, and record which one in the interview list.

For future interviews, capturing the conversation lets you stay in it: ask the follow-up, watch the screen, and listen instead of writing down every detail you might need later. Get the participant’s consent before recording any interview.

Step 2: Pull out statements, problems, and context

The goal here is a set of attributed statements, not a summary. For each relevant moment, you want the participant’s words, the interview and timestamp, and the situation around it.

If you work directly from transcripts, read each one and copy short passages into a table. If your interviews are in Hemory, you do not have to build that table by hand first. You can ask for it.

For a single interview, use Hemory’s Q&A, which answers from the transcript of your own captured conversations:

In my interview on Tuesday with the operations lead, what problems did they describe with weekly reporting? For each one, give the nearest transcript wording, roughly where it occurred, and what they were doing or responding to.

For several interviews at once, use a supported connected agent. The agent retrieves from Hemory and does the drafting itself. Give it the goal and the clues you have, not a finished brief:

I'm drafting a PRD for the task-reporting feature. Using my captured interviews about reporting from the last three weeks, list the problems participants described. For each, give the participant's role, the interview and timestamp if available, the wording they used, and the situation they were in. Say where you could not find supporting wording, and do not fill gaps.

Two cautions apply to either route. A search that returns nothing does not prove that nobody said it, and a transcript can contain mistakes, so judgment about what matters stays with you. Also, do not assume the answer includes exact timestamps or verbatim lines. If it does not, find the passage in the transcript yourself before you quote it.

The result should look like this:

IDInterviewTimeParticipant’s wordsSituation
S1I2, operations lead14:32”I export to a spreadsheet every Friday because I can’t filter by owner.”Describing how the weekly report is built
S2I4, team lead08:10”I just need to see my own people’s tasks.”Asked what they check first on Monday
S3I5, project coordinator21:45”Sometimes I send the export to the client directly.”Describing who sees the report

Step 3: Separate what was observed from what you interpret

This is the step that makes the PRD defensible. For each statement, write down three things separately:

  • Observed: what the participant said or did, in plain terms.
  • Interpretation: what you think it means.
  • Proposal: what you might build in response.

Take S1 and S2 above. The observed facts are that two participants work around the report’s lack of an owner filter, one by exporting weekly. A reasonable interpretation is that filtering by owner would remove the export for those two. That is not yet established, because S3 suggests the export may also be how reports reach clients. The proposal, an owner filter, may solve one need and miss another.

Two rules keep this honest:

  • A shared keyword or topic does not mean two participants had the same need. Check that the situation and goal match before you group statements.
  • Repeated mentions are not independent demand. Three participants raising the same issue is useful, but note whether they work in the same team or came from the same recruiting source.

Step 4: Turn grouped needs into requirements with acceptance criteria

Group statements by need, not by interview. Then write each requirement as an entry that carries its evidence. Here is a worked example built from the statements above:

R3 (Proposed): Users can filter the weekly task report by owner.
Proposed priority: High, pending review.

Evidence:
- I2, operations lead, 14:32: "I export to a spreadsheet every Friday because I can't filter by owner."
- I4, team lead, 08:10: "I just need to see my own people's tasks."

Observed: Two participants work around the missing owner filter, one by exporting to a spreadsheet each week.
Interpretation: An owner filter may remove the need to export for these users. It may not cover sharing with clients (see I5, 21:45).

Acceptance criteria:
- Given a report containing tasks from several owners, when the user selects one owner, then only that owner's tasks appear.
- The totals shown on the report update to match the selected owner.
- The selected filter is still applied when the user reopens the report in the same session.

Open questions:
- Is a filter enough, or do users also need to share a filtered view? (I5 suggests the export may serve sharing.)
- Does I4's need apply to team leads with more than one team?

The acceptance criteria describe behavior someone can test. They do not mention a particular design, because the interviews did not establish one.

If you are using a connected agent, you can ask for this draft directly after the retrieval, in the same conversation. Supply what the interviews could not contain: the current goal, the scope of the release, constraints, and any decision made since the interviews.

Draft requirements from these problems in the PRD entry format I use: requirement, evidence with interview and timestamp, observed, interpretation, acceptance criteria, open questions. Mark every requirement as Proposed. Use only wording you retrieved. This release is limited to the web report, and the export feature is out of scope.

The agent drafts. Hemory supplies the captured conversation, and it does not write the PRD or update your project tools. Paste the result into the document where your team keeps requirements.

Step 5: Keep open questions attached to the requirement

A separate list of open questions at the end of the PRD is easy to skip. Put each question under the requirement it affects, and add a short label for its type, because the type tells you what to do next:

TypeExampleNext action
Single sourceOnly one participant mentioned client sharingLook for confirmation in other interviews or run another one
ConflictI2 wants a filter, I5 wants a shareable exportDecide whether these are one requirement or two
Missing contextUnclear whether I4 manages one team or severalCheck the transcript, or ask the participant
Interpretation to testFiltering may replace the exportTest with a prototype or a follow-up question
Not askedNobody was asked about mobile useAdd it to the next interview guide

When the transcript is silent, say so. “Not discussed” and “not needed” are different statements, and a PRD that blurs them will be challenged later.

Step 6: Review before anyone approves it

Before the draft goes to stakeholders, check it against the sources. You do not need to read every transcript again. Check the places where an error would change a decision:

  • Quotes. Open the transcript at the timestamp and confirm the words. In Hemory, you can replay a transcript line on the device that listened, which helps when a line looks wrong or ambiguous.
  • Attribution. Confirm the speaker is the participant, not the interviewer, and that the role is right.
  • Context. Read the sentences before and after a quote. A complaint about a workaround can look like a request for a feature.
  • Interpretations. Make sure each one is labeled as yours and that nothing in the requirement is stated more strongly than the evidence supports.
  • Acceptance criteria. Check that each one can be tested and that none sneaks in a design decision.
  • Status. Keep everything Proposed until the people who decide have accepted it. A revised draft is not an approved one.

Then share a self-contained document. Stakeholders should be able to read the quotes and questions without needing access to your interview archive. A reference to a source is not the same as giving someone access to it.

Where Hemory fits, and where it does not

The method above works with any transcripts. Hemory reduces the work around it:

WorkWithout a capture toolWith captured interviews in Hemory
Recording the interviewYou take notes while trying to listenThe conversation is preserved, so you can focus on the participant
Finding the passageYou search several documents by memoryQ&A or a connected agent retrieves candidate passages from the transcripts
Assembling across interviewsYou copy and compare by handA connected agent can gather and draft from several interviews in one request

There are limits to keep in mind. Retrieval may miss relevant statements, and a transcript is not guaranteed to be complete or error-free. The agent’s draft is a draft. Hemory does not decide which statements matter, does not approve requirements, and does not write into your project tools.

Next step

Pick one interview you have already run and take a single problem from it through the six steps. Write one requirement entry with its quote, interpretation, acceptance criteria, and open questions. If a stakeholder can read that entry and trace it back without asking you, the rest of the PRD can follow the same pattern.

FAQ

Can AI write a PRD from user interviews?

An AI agent can draft one from interview material it can access, but the result is a draft. It can miss statements, merge distinct needs, or phrase an interpretation as a finding. Treat the agent’s output as a first pass and review quotes, attribution, and status before sharing it.

How is this different from an interview summary?

A summary compresses what was said and usually drops the original wording and the context. An evidence-linked PRD keeps the wording, the source, and the situation attached to each requirement, and it separates what was observed from what you infer.

How many quotes does a requirement need?

There is no fixed number. One clear quote can support a small requirement if you label it as single-source. A larger requirement needs evidence from more than one situation. What matters is that the evidence matches the scope of what you are asking stakeholders to approve.

Do I have to summarize every interview before using an agent?

No. If the interviews are captured in Hemory, you can give a connected agent the goal and a few clues, let it retrieve and draft, then review the gaps and conflicts it reports. If an interview was not captured there, give the agent that interview’s transcript or notes directly.

Related reading

Start building your memory palace

Download Hemory — from today on, every real conversation becomes material you can generate from tomorrow.

Get Hemory for iOS Get Hemory for Android Download APK

macOS, Windows, Linux, Web, and Self-host coming soon