Turning interview transcripts into research themes
After a round of 8-12 user interviews you are staring at hundreds of pages of transcripts with a readout due Friday. NN/g's research agenda names extracting themes from qualitative data as one of the clearest places AI helps, and Figma's 2025 report found designers already reach for it in discovery. AI can cluster observations into candidate themes faster than sticky notes on a wall — as a first pass you then check against the raw quotes.
You are a UX research analyst helping me synthesize a round of user interviews. My research goal: {{research_goal}}. Product context: {{product_context}}. I will paste de-identified transcripts below, with each participant labeled by number: {{transcripts}} Produce an affinity-style synthesis: 1. 5-8 candidate themes, each with a one-line description. 2. For every theme, 2-3 supporting quotes copied VERBATIM from the transcripts, each tagged with the participant number I used. 3. A rough frequency per theme (how many participants it appears in) and a confidence note (strong pattern / early signal / single voice). 4. A "contradictions and outliers" section — quotes that cut against the main themes. Do not discard these. 5. Open questions the data does not answer. Hard constraints: use ONLY the transcripts I provide. Do not invent, paraphrase, or embellish quotes — if you cannot find a verbatim quote to support a theme, write [NEEDS QUOTE] and lower the confidence. Do not infer demographics, emotions, or intent that participants did not state. Never reconstruct names. This is a first-pass synthesis I will verify against the raw transcripts before any readout.
Fill in your details and the prompt updates live — then copy.
**Theme 1 — Onboarding stalls at team setup (strong pattern, 7/9).** "I got to the part where it asked for my team's info and I just... closed it. I didn't have that yet." (P3) "Setup felt like it was for someone who'd already decided." (P7) **Theme 4 — Power users want fewer confirmations (early signal, 3/9).** [NEEDS QUOTE] for two of the three — only P2 stated this explicitly. **Contradictions:** P5 praised the same setup flow Theme 1 flags: "Honestly the guardrails saved me." Worth probing before we redesign.
The full workflow
- De-identify transcripts first: remove names, employers, emails, and anything that could re-identify a participant.
- Paste the transcripts and run the synthesis prompt.
- Verify every theme against the raw quotes and delete any tagged [NEEDS QUOTE].
- Pull the contradictions and outliers yourself — they usually matter most.
- Build the readout from the verified themes, not the AI draft.
Watch out for
Participant data is personal data: names, recordings, and identifiable transcripts fall under GDPR and CCPA/CPRA, and your consent form must cover AI processing. De-identify before pasting, or use an approved enterprise or zero-retention deployment — never paste raw research into a consumer tool that may train on it.
AI flattens outliers and states shaky patterns with confidence; the quote that contradicts the theme is often the real insight. Treat the synthesis as a first pass and check it against the source before it drives any decision.
Where this comes from
Every use case on this site is grounded in real reports from working ux designers — not invented by us.