The struggle

You did the hard part. You recruited, you scheduled, you ran eight real conversations with real users. Then the quarter caught fire, and the recordings went where recordings go to die. Now the PRD is due, its “Problem” section is running on vibes and memory, and the sharpest person in the review knows it — “how many customers actually said this?” — and you nod along, because the honest answer is somewhere in seven hours of video I’ll never re-watch.

The failure isn’t laziness. Re-watching is a seven-hour tax nobody pays twice, so the knowledge quietly evaporates and every PRD after that gets written from the same three anecdotes. This playbook removes the tax: capture becomes passive, synthesis becomes a 20-minute prompt with a built-in lie detector, and the output drops straight into the PRD section that was running on vibes.

The chain

Step 1 — Capture without babysitting · Granola

Connect Granola to your calendar once (setup guide). It captures your calls from system audio — no bot joining the meeting to make interviewees self-conscious — and merges your typed half-notes with the transcript. From here on, every interview lands as clean text automatically. That’s the whole point: the pipeline starts passive, so it survives busy quarters.

Honest trade-off: if you need a generous free tier or bulletproof speaker labels on messy multi-voice calls, Otter.ai is the better capture layer — at the cost of a visible bot in the room. And if your interviews are already recorded (the exact rot this playbook targets), skip capture entirely: export whatever transcripts your Zoom/Meet plan gives you and go straight to step 2.

Step 2 — Synthesize with a lie detector · Claude

Any frontier model runs this step. ChatGPT and Gemini do this job as well as Claude does; the prompt is plain text in, text out. We name Claude because that’s what we field-tested with (see last field-tested above) — not because it’s required. Use whichever you already pay for.

Label each transcript P1–P8 (attribution survives synthesis this way) and paste them into one Claude conversation — its long context holds all eight comfortably; past ~10, batch in fives and merge. Then run:

You are a user researcher synthesizing interviews. Below are 8 transcripts, labeled P1–P8.

1. Identify recurring themes ACROSS participants (not per-interview summaries).
2. Rank themes by how many distinct participants raised them — show the count.
3. For each theme: one-line description, the participants who raised it, and
   2–3 VERBATIM quotes, each tagged with its participant.
4. Separately list surprises — single-participant signals that feel important.

Hard rule: every quote must be word-for-word from the transcripts. If you can't
find a real quote for a theme, say so instead of paraphrasing. Do not soften or
merge quotes.

Transcripts:
"""
<paste labeled transcripts>
"""

The verbatim-quote rule is the load-bearing part. It converts “trust the AI” into “check the AI”: every claim arrives with a checkable citation.

Step 3 — Spot-check before you ship it (30 minutes, not optional)

Pick the top three themes and grep each quote against the source transcript. Word-for-word matches → trust the pattern and move on. Paraphrased or “improved” quotes → re-run with the hard rule stated more forcefully, and distrust that theme until verified. This half hour is what makes step 4 defensible in front of the room — skip it and you’ve replaced “vibes” with “confident vibes.”

Step 4 — Drop the evidence into the PRD

Paste your verified themes into the PRD’s Problem section with counts and quotes attached: “6 of 8 participants described losing trial users during workspace setup — ‘I honestly couldn’t tell if it was broken or loading’ (P3).” Next review, “how many customers actually said this?” has a number for an answer, with a named human behind it.

Optional — Make it durable · Dovetail

Solo PM, occasional interviews? Stop at step 4 — a research platform is overhead you don’t need. But if interviews arrive weekly, import transcripts and verified highlights into Dovetail while they’re fresh. One caveat as of 2026-08-03: Dovetail’s free plan is capped at one channel and one project, and the self-serve paid tier is gone, so “several PMs share discovery” now means an Enterprise sales call rather than a $15 seat. Six months later, “pricing objections” is a searchable query returning tagged, attributed quotes across every study — instead of a seven-hour tax nobody pays.

Where it breaks

Skipping step 3. The single biggest failure mode. LLMs paraphrase convincingly; an unverified quote in an exec review costs you more credibility than having no quote at all.
Unlabeled transcripts. Without P1–P8 labels, attribution dies in synthesis and every quote becomes “someone said.” Label before pasting, not after.
Too much context. Past ~10 transcripts in one shot, theme quality degrades quietly. Batch in fives, then run a merge pass on the two theme lists.
Outsourcing your judgment. Torres and Wille’s warning applies: AI compresses what was said, not what it felt like to hear it. Use the machine for recall and coverage; keep interpretation — what actually matters for the product — yours.
Consent and privacy. Recording tools and LLMs are processors of your users’ words. Record with consent, strip PII you didn’t need, and check your company’s data policy before pasting transcripts into any cloud tool.

“Eight discovery interviews are sitting unwatched in your drive — and the PRD you're writing this week cites none of them.”