Verified against ChatGPT · 2026-08-13
Synthesize raw usability-test notes into themes ranked by how many participants actually hit each one
Synthesizes raw usability-test or interview notes into an affinity-mapped set of themes, each one weighted by how many distinct participants actually surfaced it, rather than a list of interesting quotes presented as if equally representative.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Synthesize the raw research notes below into themes, the way an affinity-mapping session would, but with an explicit weight on each theme showing how many distinct participants actually raised it — not a curated list of the most quotable moments presented as if they're equally representative of what most people said. RAW NOTES 7 usability test session notes for a new onboarding flow: P1 got stuck on the permissions screen; P2 also got stuck there and gave up; P3 breezed through but mentioned confusion about what 'workspace' meant; P4 got stuck on permissions too; P5 completed it with no issues; P6 accidentally granted a permission they didn't mean to and was upset when they noticed later; P7 also confused by 'workspace' terminology. NUMBER OF PARTICIPANTS 7 RESEARCH QUESTION THIS WAS MEANT TO ANSWER Can new users complete onboarding without assistance in under 3 minutes? STEP 1 — EXTRACT DISCRETE OBSERVATIONS Pull out every discrete observation from the notes, tagged with which participant it came from. Do not merge two different participants' points into one bullet even if they're similar — keep the count honest at this stage so it can be aggregated correctly next. STEP 2 — GROUP INTO THEMES Group the tagged observations into themes based on what they're actually about, not based on which ones make the best pull-quote. For each theme, state how many distinct participants it draws from out of the total, and note if a theme is actually held by only one person — a single-participant point can still be worth including if it's a serious risk, but it must be labeled as a one-person signal, not implied to be a shared finding. STEP 3 — RANK BY PREVALENCE, FLAG BY SEVERITY Rank themes primarily by how many participants raised them, but call out separately any low-prevalence theme that represents a severe risk (a safety issue, a blocking bug, a legal or trust concern) even though it wouldn't rank highly by count alone — prevalence and severity are different axes and both matter, so don't let a ranking by count alone bury a rare but serious finding. STEP 4 — TIE BACK TO THE RESEARCH QUESTION For each major theme, state explicitly whether it answers the original research question, contradicts an assumption behind it, or is a finding adjacent to the question but not actually what was asked. WHAT NOT TO DO Do not present a theme's supporting quote as more representative than it is by omitting the participant count. Do not invent a theme that isn't actually traceable back to specific observations in the notes. OUTPUT FORMAT A numbered list of themes ranked by prevalence, each with: theme name, participant count (e.g., '4 of 7'), one representative observation, and how it relates to the research question. Follow with a separate short 'severity flags' section for any low-prevalence but high-risk findings.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Research synthesis prompted without an explicit prevalence count tends to produce GPT-5.1's most narratively satisfying summary — themes built around the most vivid or quotable observations in the notes, since a compelling quote is easier to foreground than a dry tally, which systematically overweights whichever participant happened to phrase their frustration most memorably regardless of how many other participants actually shared the underlying problem. Forcing a discrete, participant-tagged extraction as a separate first step before any grouping happens is what keeps the later count honest: if grouping and counting happen in the same pass, it's easy for the model to silently merge two similar-sounding but distinct participant observations into one theme and then report a prevalence number that's actually an artifact of how the grouping was done rather than a real count of who said what. Explicitly separating prevalence-ranking from severity-flagging addresses a real and common analytical error in research synthesis — a genuinely dangerous or trust-breaking finding (like a user accidentally granting a permission they didn't intend to) can come from just one participant out of seven, and a synthesis that ranks purely by count would bury it below three people casually mentioning unclear terminology, when the accidental-permission issue is very plausibly the more consequential finding regardless of how many people hit it. Tying every theme back explicitly to the original research question, and naming when a finding is adjacent to but not actually answering that question, prevents scope creep in the synthesis where interesting-but-tangential observations get treated as if they answer the question that was actually asked, which is a common way research readouts end up misdirecting a design decision toward a problem the study wasn't actually measuring.
What you get back
Theme 1 — Confusion at the permissions screen (4 of 7 participants: P1, P2, P4, and indirectly P6). Representative observation: P2 said 'I didn't understand what I was being asked to allow, so I just gave up.' Relation to research question: directly contradicts the assumption that onboarding is completable unassisted — this is the primary blocker to the 3-minute goal. Severity flags: P6's accidental permission grant (1 of 7) is low-prevalence but represents a trust/safety risk distinct from the completion-time question and should be escalated separately regardless of its low count.
Verified against
ChatGPT GPT-5.1 · 2026-08-13
Changelog
- 2026-08-13 — Initial publish, verified against ChatGPT GPT-5.1.
Need this built into your business?
If a prompt isn't enough — branding & design, built and maintained for you — that's Scult's day job.
EXPLORE BRANDING & DESIGN
