Synthesize Interviews with Quote-Level Citations

Workflow area: Research & Synthesis · Tier: Core · Hands-on time: ~35 minutes

Manual synthesis is the most honest work in research and the reason it doesn't happen: six transcripts is two days of highlighting, and under deadline the team "remembers" the interviews instead. AI collapses two days to an afternoon — but a synthesis model's failure mode is uniquely poisonous for research. It doesn't fail loudly; it produces a clean quote, in the participant's voice, supporting a plausible theme, that was never said. A fabricated quote in a themes report is worse than no research at all, because it carries research's authority. The citation-and-verification discipline is not overhead on this recipe; it is the recipe.

When to use — and when not to

• Use it for any multi-transcript synthesis3+ interviews, usability sessions, or the mixed pile from pb-01's calls; the discipline is identical, the gain scales with volume.
• Use it when stakeholders will act on the outputroadmap-influencing findings need the citation trail precisely because they'll be challenged.
• Use it to reprocess old studieslast year's transcripts, synthesized with today's discipline, often answer this quarter's questions for free.
• Skip it below three transcriptsread two interviews yourself; the model's leverage is breadth, and at n=2 you are the better instrument.
• Skip it when transcripts can't leave your tenantsensitive-domain research (health, finance, minors) runs only in approved tools, or by hand.

The recipe

1. 1.

   Prepare transcripts with stable timestamps and IDs. Each transcript needs participant ID (anonymized — P1, P2…), and timestamps throughout (your notetaker's export has them). The timestamp is what makes every claim checkable in seconds instead of minutes — it's the load-bearing part of the whole discipline.

2. 2.

   Run the synthesis with citations made mandatory. In Claude (long transcripts favor large context windows):

   Copy

   You are synthesizing user interviews into themes. The citation
   rules below are absolute.
   
   Research question: [e.g. "How do finance approvers decide what to
   scrutinize vs. rubber-stamp?"]
   Transcripts: [PASTE ALL, EACH LABELED P1, P2, ... WITH TIMESTAMPS]
   
   Produce themes, for each:
   - THEME — one sentence, in plain language
   - STRENGTH — how many participants evidence it (list IDs)
   - EVIDENCE — 2-4 VERBATIM quotes, each with participant ID and
     timestamp, exactly as spoken. Never paraphrase inside quotation
     marks. Never merge statements from different moments into one
     quote.
   - COUNTER-EVIDENCE — quotes that cut against the theme, same
     citation rules. If none, write "none found".
   Then: SURPRISES (contradicts our assumptions), and TENSIONS
   (participants who disagree with each other).
   Rules: a theme with fewer than 2 quote citations does not appear.
   If you are not certain a quote is verbatim, mark it [CHECK].
   

3. 3.

   Spot-verify every theme against source — the hallucination hunt. For each theme, take at least one cited quote per participant, open the transcript at the timestamp, and compare word-for-word. You're hunting four specific frauds:

   Copy

   VERIFICATION LOG
   Theme 1 / P3 @ 14:22 — VERBATIM ✓
   Theme 2 / P1 @ 08:15 — CLEANED: model dropped "I guess maybe" —
   hedge matters, restored original
   Theme 2 / P5 @ 21:40 — MERGED: two statements 4 min apart joined
   into one "quote" — split and re-cited
   Theme 4 / P2 @ 11:03 — FABRICATED: nothing like this at 11:03 or
   anywhere in P2 — quote deleted, theme re-evaluated (now n=1,
   theme demoted to open question)
   

   A fabricated or unfindable quote kills itself and puts its theme on trial. Every [CHECK] flag gets resolved. Keep the verification log — it's the report's warrant.

4. 4.

   Weigh the themes yourself. Counts are not importance: a theme with two quotes from your exact target segment can outweigh one with six from adjacent users. Re-order the verified themes by decision relevance, and write one line per theme on what it means for the product — the model synthesized what was said; deciding what it implies is the researcher's half of the job.

5. 5.

   Publish with citations intact and link the sources. The report keeps every quote's ID and timestamp; file it in Dovetail (or your repo) with links from each theme to its source transcripts — this is what pb-03's persona sweeps and pb-18's repository queries will consume. A stakeholder who can click from claim to quote to recording extends trust; one who can't, shouldn't.

Make it yours

• Dovetail-native version: if transcripts already live in Dovetail with highlights, run its AI synthesis for the first pass and apply steps 3–5 unchanged — platform AI hallucinates too; the verification discipline transfers whole.
• Notably for the visual-canvas pass: teams that synthesize spatially can auto-cluster highlights in Notably, then enforce the same rule — no cluster graduates to "theme" until its quotes pass the timestamp check.
• Claude Projects for longitudinal synthesis: accumulate verified themes reports (not raw transcripts) in a Project per product area and ask cross-study questions — "what has changed about approval anxiety since Q1?" — with citations inherited from the underlying verified reports.

Quality check

• Every quote traced or diedthe verification log covers at least one quote per participant per theme; anything unfindable at its timestamp was deleted, not shrugged at.
• Hedges and mess survivedparticipants say "I guess" and contradict themselves; over-clean quotes mean the model polished the humans out of the data.
• Counter-evidence sections have content sometimesa synthesis where every theme is unopposed is a synthesis that wasn't looking; real data argues with itself.
• Theme strength matches citation realitya theme claiming five participants cites five distinct IDs; check the IDs, not the claim.
• Surprises made it to the topthe finding that contradicts the roadmap belongs in the summary, not paragraph nine; coherence bias buries exactly what you paid for.
• The report links to sourcesclaim → quote → transcript → recording, every link live, or the citation discipline was cosplay.