Verified against ChatGPT · 2026-08-09
Synthesize a stack of interview transcripts into themes without losing which participant actually said what
Turns raw or lightly-cleaned interview transcripts into a theme map with direct quotes attributed to specific participants, so the synthesis stays traceable back to the source instead of reading like one anonymous composite voice.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are synthesizing multiple interview transcripts into a theme map. I need the output to stay traceable — every theme should be backed by attributed quotes, not a paraphrased blend that could have come from anyone. STUDY GOAL Understanding why new sales reps disengage from the CRM within their first 60 days. TRANSCRIPTS (paste each with a participant label) Participant A (Rep, 3 months tenure): 'I stopped logging calls because nobody ever looked at the data anyway.' Participant B (Rep, 5 months): 'The fields ask for stuff that doesn't map to how we actually sell.' PARTICIPANT CONTEXT All five participants are reps hired in the same quarter under the same manager; none have prior CRM experience from a previous job. MINIMUM QUOTE THRESHOLD at least 3 of the 5 participants STEP 1 — OPEN CODING Read through each transcript and pull out short, direct phrases that capture a distinct idea, attributed to the participant who said it. Do not merge similar-sounding phrases from different participants yet. STEP 2 — THEME GROUPING Group the coded phrases into themes only when a phrase from at least 3 of the 5 participants confirms the pattern is shared, not idiosyncratic to one person. For each theme, name it in the participants' own language where possible rather than abstract academic terminology, and list which participants contributed to it and which did not. STEP 3 — TENSION CHECK Flag any theme where participants seem to agree on the surface but mean something different underneath — for example two people both saying "communication was hard" while one means volume of messages and the other means unclear ownership. Do not resolve these into one meaning; state both readings. STEP 4 — OUTLIERS List anything a single participant said that contradicts the majority theme, labeled as an outlier rather than folded silently into the nearest theme. WHAT NOT TO DO Never produce a quote that isn't a close match to something actually in the transcripts I gave you — if a theme feels true but you can't point to the specific line supporting it, say the theme is your inference, not participant-sourced. OUTPUT FORMAT For each theme: name, participant count, 2-3 attributed quotes, and one line noting any tension or outlier tied to it. End with a short list of themes with only single-participant support, marked as tentative.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
The step-by-step structure — open coding before theme grouping — exists specifically to stop GPT-5.1 from doing what it does by default when handed raw transcripts in one pass: jumping straight to plausible-sounding themes and retrofitting quotes to match them, rather than letting the themes emerge from what's actually attributed to specific speakers. Requiring a minimum-participant threshold before something counts as a theme prevents a single vivid quote from one participant getting inflated into a headline finding, which is a common and misleading pattern in synthesized qualitative writeups where the most quotable line wins regardless of how many people actually said something like it. The tension-check step targets a specific failure mode of language-model synthesis: the model tends to treat surface-similar phrasing as semantic agreement, collapsing "communication was hard" into one meaning when the underlying complaints might be structurally different — naming this explicitly as something to check, rather than resolve, keeps the nuance visible instead of smoothing it away for a cleaner-sounding output. Explicitly forbidding fabricated or approximate quotes matters because a model asked to synthesize themes under a fixed format will, absent this constraint, sometimes generate a quote that reads as representative of a theme even if no participant said anything that close to it — flagging that as the single thing never to do gives it a hard boundary rather than a vague accuracy aspiration, and the requirement to mark theories as "my inference" when unsupported keeps analytical leaps distinguishable from participant-sourced findings.
What you get back
Theme: 'Data entry feels like busywork' (4/5 participants). Quotes: A — 'nobody ever looked at the data anyway'; C — 'I fill it in because I have to, not because it helps me.' Tension: B's version is about fields not mapping to their sales motion, which is a design complaint rather than a motivation complaint — worth treating as adjacent, not identical. Outlier: participant E actually finds the CRM useful for tracking their own follow-ups, contradicting the majority pattern.
Verified against
ChatGPT GPT-5.1 · 2026-08-09
Changelog
- 2026-08-09 — Initial publish, verified against ChatGPT GPT-5.1.
Need this built into your business?
If a prompt isn't enough — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
