Discovery Interviewing

Discovery interviewing

Core principle

An interview collects stories about specific past episodes. It does not collect opinions, predictions, feature requests, or approval. Everything valuable comes from "tell me about the last time", and everything worthless comes from "would you".

The interviewer's job is to stay quiet, follow a real episode in detail, and resist the urge to explain. Most bad interviews are bad because the interviewer talked.

Three modes, kept strictly separate

Never blend them. A rehearsal transcript that loses its stamp becomes indistinguishable from a real one within a week, and there is no way to recover the distinction downstream.

The interview types

Guide structure

Learning goals first, at most three. A guide with six goals produces six shallow answers.

1. Consent and framing (2 min). Who you are, what this is for, recording permission, no right answers, they can stop or skip anything.
2. Warm-up (3 min). Their role and context. Real questions, not small talk.
3. The story (20-25 min). One specific recent episode, walked through in order. This is the interview. Everything else is scaffolding.
4. Probing (inside the story). Depth over breadth.
5. Solution exposure (10 min, optional, always last). Two or more options, counterbalanced. Never before the story.
6. Close (3 min). Anything I should have asked. Who else should I talk to.

references/question-bank.md has the openers, probes and closers, plus the banned forms.

Probing: the only technique that matters

When you get a general statement, get the specific episode underneath it.

"Reconciliation is a nightmare." → "When did you last do it?" → "Walk me through that evening, from when you started." → "What did you have open in front of you?" → "What happened when the numbers did not match?" → "What did you do then?" → "How long did that take?" → "Who else was involved?" → "What would you have done if you had not caught it?"

Five levels down is where the finding is. Most interviewers stop at one.

The four probes that do most of the work: "Tell me about the last time." "What did you do next?" "What did you do before this?" And silence, held for three full seconds.

Sample size

Saturation is measured, not felt. discovery-synthesis has the saturation curve method.

Recruiting

Recruiting is the constraint, not the interview. Teams that interview weekly have solved recruiting once; teams that interview rarely re-solve it every time.

Full guidance in references/recruiting-and-sampling.md, including screener design, the automated in-product recruiting pattern, incentives, and the sampling-frame statement that must appear in every readout.

AI-moderated interviews with real participants

Permitted under the protocol in references/live-interview-copilot.md. Disclosure and consent up front, fixed protocol, full transcript reviewed by a human before synthesis, and the method's limits stated in the readout. Good for structured collection at volume and for languages the team does not speak. Poor for exploratory or sensitive work.

Rehearsal mode

The model plays a participant so the interviewer can practise. Useful, and safe only under the stamp.

Protocol: the model takes a described persona, answers in character, includes realistic vagueness and reluctance, and does not volunteer perfectly structured insight. After the session, it switches out of character and critiques the interviewer: leading questions, missed probes, pitching, rescuing, questions about the future, talk-time ratio.

Marking is per unit, not per file. The stamp at top and bottom does not survive someone selecting three paragraphs and pasting them into a PRD, and the paragraph is what gets pasted. So: every participant is named SYNTHETIC-P01, SYNTHETIC-P02; every fabricated line opens with an inline [SYNTHETIC]; the filename starts SYNTHETIC_. Test it by pasting three random lines into a blank document. A reader who has never seen the original must still be able to tell.

Keep it deliberately dull. A vivid, specific, quotable synthetic line is the one that gets lifted into a real document. Rehearsal material should be hedged and free of the numbers, names and turns of phrase that make a quote attractive. If a line would look good on a slide, rewrite it duller. The point is to practise the interview, not to write prose.

After the interview

Within 24 hours, while it is still in your head:

1. Interview snapshot (../product-discovery/templates/interview-snapshot.md)
2. Verbatim quotes with timestamps, marked [inaudible] where unclear, never filled in
3. Update the experience map and the opportunity solution tree
4. Note contradictions with previous interviews explicitly
5. Note what to ask next time

Then discovery-synthesis for coding across the corpus.

Output

Terse by default. One line per point, not a paragraph per point. Prose only where a number would be misread without a sentence of context.

No preamble, no restating the request, no closing summary of what you just said. Drop any section that would be empty rather than filling it.

Ask when a missing answer changes the recommendation, not otherwise. Use the AskUserQuestion tool where the runtime has it and the choice is a real fork between two to four known options, recommendation first. Ask in plain numbered text for open-ended facts only the user holds. Ask nothing where a sensible default exists: state the assumption inline and move on.

Full contract and word ceilings: ../product-discovery/SKILL.md, Output contract.

Read next

Red flags

• The guide contains your solution before the story section
• More than three learning goals
• Any question starting "would you", "do you think", or "how important is"
• The interviewer spoke more than a quarter of the time
• The participant agreed with everything
• No specific date, number, or name in the whole transcript
• You finished feeling validated