Synthesize Interviews into Themes with Quote-Level Citations

Workflow area: Discovery & Research · Tier: Foundation · Hands-on time: ~30 minutes

Six interviews used to mean an afternoon with sticky notes. A model does the first clustering pass in minutes — but a model asked for themes will produce confident themes whether or not the transcripts support them, complete with plausible-sounding quotes no one ever said. The citation-and-audit discipline in this playbook is not optional polish; it is the difference between synthesis and fiction.

When to use — and when not to

• Use it after any interview batchfour to twelve transcripts on a related question is the sweet spot; below four, just read them.
• Use it when synthesis feeds decisionsroadmap bets, PRD problem statements, opportunity trees: anywhere a wrong theme becomes an expensive quarter.
• Use it to merge interviewerswhen three PMs each ran four calls, AI clustering across all twelve finds patterns no single interviewer saw.
• Skip it as a substitute for readingread every transcript once yourself first. Synthesis you can't sanity-check is synthesis you can't defend.
• Skip it for a single contentious callone angry customer is an anecdote; clustering tools make anecdotes look like patterns.

The recipe

1. 1.

   Assemble the reviewed notes or full transcripts with stable IDs. Label each source (INT-01, INT-02...) with persona and date. Stable IDs are what make citations checkable later.

2. 2.

   Run the synthesis prompt with citations mandatory. In Claude (large context handles a full batch) or ChatGPT:

   Copy

   You are synthesizing user research for a product team.
   Research question: [WHAT WE WANTED TO LEARN — e.g. "why do mid-market
   customers stall in onboarding?"]
   
   Below are [N] interview transcripts labeled INT-01 through INT-[N].
   
   Produce:
   1. THEMES — 4-7 themes, each with:
      - a one-line claim stated as a finding, not a recommendation
      - how many of the [N] interviews support it (count + IDs)
      - 2-3 verbatim quotes as evidence, each cited as
        (INT-ID, timestamp)
   2. CONTRADICTIONS — where interviewees disagree with each other,
      with cited quotes on both sides.
   3. OUTLIERS — strong signals from only one interview.
   4. NOT FOUND — questions from the research question the transcripts
      do not answer.
   
   Rules: every quote must appear verbatim in a transcript — never
   compose, merge, or clean up quotes. A theme supported by fewer than
   2 interviews goes to OUTLIERS. If evidence is thin, say so.
   
   [PASTE ALL TRANSCRIPTS, EACH PRECEDED BY ITS ID LINE]
   

3. 3.

   Read the output for shape before trusting content. Good output looks like:

   Copy

   THEME 3 — Admins fear bulk actions they can't undo (4/6: INT-01,
   INT-03, INT-04, INT-06)
   - "I exported everything to a spreadsheet first, just in case your
     tool wrecked it" (INT-03, 18:40)
   - "There's no undo, right? So I do ten records at a time" (INT-06,
     09:12)
   
   CONTRADICTIONS
   - Pricing sensitivity: INT-02 calls the price "trivial" (22:10)
     while INT-05 says it "nearly killed the renewal" (31:44).
   

   A contradictions section with content and a non-empty NOT FOUND list are health signals. A synthesis where six humans agree on everything is a synthesis the model smoothed.

4. 4.

   Run the citation audit — this is the step that earns your signature. For every theme, pick at least one cited quote and find it in the source transcript at the given timestamp. Log each check in a small table: quote, source, found-verbatim / found-paraphrased / NOT FOUND. One fabricated quote means you audit every quote in the doc, not just samples.

5. 5.

   Demote or delete anything that fails. A theme whose evidence was paraphrased gets its real quotes substituted. A theme whose evidence doesn't exist dies, however plausible it sounds — plausible-but-unsupported is precisely what hallucination looks like.

6. 6.

   Hand-write the "so what." The implications section — what this means for the roadmap — is judgment, and it is yours. AI clusters evidence; it does not know your strategy, your constraints, or your appetite for risk.

Make it yours

• Claude Projects as a living study: keep the whole study — question, transcripts, synthesis drafts — in one Project and interrogate it: "which interviews mention pricing before I raised it?" Follow-up questioning against sources beats one-shot synthesis.
• ChatGPT for a second clustering opinion: run the same prompt in a second model and diff the theme lists. Themes both models find independently are sturdy; themes only one finds get extra audit scrutiny.
• Dovetail at team scale: if your reviewed notes from pb-01 live in Dovetail, its AI clustering works over tagged highlights with links back to source video — the citation audit becomes clicking through to the clip. Keep the same rule: no theme ships without a human having watched at least one supporting clip.

Quality check

• Every quote passed the auditfound verbatim, at the cited timestamp, in the cited interview. This is the non-negotiable.
• Counts are real"4 of 6 interviews" means you can name the four; re-count one theme's support yourself.
• No hallucinated usersevery INT-ID in the output corresponds to a transcript you actually provided; models will invent INT-07 when it helps a theme.
• Contradictions survivedcheck that a disagreement you remember from reading the transcripts appears in the output rather than being averaged away.
• Findings and recommendations are separatedthemes state what users said; the recommendations section is in your voice, and you can defend every line of it without the AI.