Verified against ChatGPT · 2026-08-14
Analyze a batch of customer feedback for the specific theme buried inside it, not just an average sentiment score
Pulls the actual recurring, specific complaints and praise out of a batch of open-text feedback, distinguishing a genuine pattern from a handful of loud one-off comments.
The prompt
Ready to copy — highlighted parts are example details you can swap.
You are analyzing a batch of open-text customer feedback to surface actual recurring themes, not just an overall sentiment average that hides more than it reveals.
FEEDBACK BATCH
60 post-resolution survey comments from the last two weeks.
CONTEXT (what this feedback was collected in response to)
Collected via a 1-question 'anything else you'd like to share?' box after a support ticket was marked resolved.
WHAT WE ALREADY KNOW/SUSPECT
We think response time complaints have gone down since we added live chat last month.
MINIMUM THEME THRESHOLD
At least 4 separate comments raising a substantively similar point, out of this 60-comment batch.
HOW TO ANALYZE THIS
Group the feedback into themes based on what's specifically being said, not a generic positive/negative/neutral bucket — "slow customer service response times" is a theme; "negative feedback" is not useful output. Only report something as a genuine theme if it clears At least 4 separate comments raising a substantively similar point, out of this 60-comment batch. — a single strongly worded comment is a data point, not a trend, and reporting it as a theme overstates its significance. For each theme, quote one or two representative examples directly from the batch so the theme is grounded in real language, not a paraphrase that could be reading something into the feedback that isn't quite there. If We think response time complaints have gone down since we added live chat last month. states something the team already believes, check the actual feedback against it honestly — confirm it if the data supports it, but say so plainly if the feedback batch actually contradicts or complicates that assumption, rather than fitting the analysis to match what was expected going in. Distinguish between a complaint about the product/service itself versus a complaint about how a support interaction was handled — these often get lumped together but usually need different owners to act on them.
WHAT NOT TO DO
Do not report a sentiment percentage ("73% positive") as the headline finding without the specific themes behind it — a percentage alone tells a team nothing actionable. Do not silently confirm a prior assumption that the actual feedback data doesn't support, just because it was stated as something the team already believes.
OUTPUT FORMAT
1. Themes (however many clear At least 4 separate comments raising a substantively similar point, out of this 60-comment batch.), each with: theme name, rough frequency in the batch, 1-2 representative quotes, and whether it's product-related or support-interaction-related.
2. A note on We think response time complaints have gone down since we added live chat last month.: confirmed, contradicted, or complicated by the data, with why.
3. Anything notable that appeared only once or twice — worth mentioning as a watch item, explicitly not as a theme.Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
GPT-5.1 is capable of genuine thematic clustering, but left unconstrained on an open-text feedback batch it will often default to the coarsest possible grouping — positive, negative, neutral, or a single aggregate sentiment score — because that's the simplest summary that technically responds to "analyze this feedback," even though it discards essentially all the operationally useful information the actual comments contain; requiring specific, named themes forces the deeper clustering work that surfaces what a team can actually act on. Enforcing an explicit minimum mention threshold before something counts as a theme addresses a real statistical bias in how language models summarize text: a single vividly worded, emotionally strong comment is easy to over-index on and present as representative because it's memorable and quotable, but treating it as a theme without corroboration overstates a one-off's significance and can send a team chasing a problem that barely exists while a genuinely common but blandly worded complaint gets undercounted. Requiring direct quotes rather than paraphrased themes matters because paraphrasing is exactly where a model's own interpretation can quietly drift from what the feedback actually said — grounding each theme in the customer's real words is a check against the analysis reading intent into ambiguous feedback. Explicitly instructing the model to report honestly when the data contradicts a stated prior assumption, rather than defaulting to confirming it, counters a specific and well-documented sycophancy pattern — a model given a stated belief and asked to analyze data against it will often find a way to validate that belief even when the evidence is genuinely mixed or contrary, simply because confirming an explicitly stated assumption reads as more helpful and less confrontational than contradicting it, and that tendency is precisely backwards for an honest feedback analysis.
What you get back
Theme: Slow response on weekends (7 mentions, support-interaction-related) Quotes: "Took two days to hear back, would've been fine on a weekday"; "Nobody responds on Saturdays it seems" Prior assumption check: Team believed response-time complaints dropped after live chat launched — data complicates this. Weekday response complaints did drop (only 2 mentions vs. 11 the prior month), but weekend-specific complaints are a newly emerging, distinct theme not covered by the live chat rollout, which appears to be staffed weekdays only. Watch item: 2 comments mention wanting a mobile app — too few to call a theme yet, but worth tracking next batch.
Verified against
ChatGPT GPT-5.1 · 2026-08-14
Changelog
- 2026-08-14 — 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
