Verified against ChatGPT · 2026-08-13
Pull real themes out of a batch of customer feedback instead of a bucket of restated complaints
Groups raw customer feedback into themes defined by what would actually change a decision, checked against your existing taxonomy, so the report answers a real business question instead of producing a generic word cloud in prose form.
The prompt
Ready to copy — highlighted parts are example details you can swap.
Analyze this batch of raw customer feedback and produce a voice-of-customer theme report aimed at answering a specific business question, not a generic summary of what people said. RAW FEEDBACK BATCH 42 post-resolution survey comments from the past two weeks, ranging from one-word ratings to multi-sentence complaints about response time and being transferred between agents. TIME PERIOD Two weeks, covering the rollout of a new ticket-routing system. BUSINESS QUESTION Did the new ticket-routing system make the transfer-between-agents complaint better or worse compared to before? EXISTING THEME TAXONOMY Existing categories: Response Time, Agent Knowledge, Transfer/Handoff Friction, Resolution Clarity, Pricing Perception. STAKEHOLDER AUDIENCE The support operations director, who will use this to decide whether to roll back the new routing system. Group the feedback into themes defined by what would actually change a decision if it were true at scale, not by surface topic similarity — two comments that use similar words but point to different underlying concerns (one about price, one about perceived value for that price) belong in different themes even if a naive keyword grouping would merge them, and two comments using different words for the same underlying concern belong together. Check new themes against the existing taxonomy given rather than inventing fresh category names for things that already have a name — a report that renames "onboarding friction" as "early-experience challenges" every time it's rerun makes historical comparison impossible, which defeats the purpose of tracking themes at all. Only propose a genuinely new theme if it doesn't fit any existing category. Answer the specific business question directly, using the themes as evidence, rather than presenting the themes and leaving the reader to work out the implication themselves. State theme volume relative to the total batch size, not just raw counts, so a theme that appears frequently only because the batch itself is small doesn't get over-weighted. WHAT NOT TO DO Do not report a theme based on a single comment as if it represents a pattern — state clearly when something is an outlier worth watching rather than an established theme. Do not soften a theme that reflects badly on the product to make the report read more positively for the stakeholder audience; report what the feedback actually says. OUTPUT FORMAT 1. Themes with volume (count and % of batch) and one representative quote each. 2. Direct answer to the stated business question. 3. Any single-comment outlier worth flagging separately. 4. New theme proposals, if any, and why they don't fit the existing taxonomy.
Customize
Optional — swap in your own details for the highlighted parts above.
Why this works
Defining themes by what would change a decision rather than by surface keyword similarity is the mechanism that prevents the most common failure of AI-generated voice-of-customer reports: a model asked to theme feedback will naturally group by shared vocabulary, which merges genuinely different concerns that happen to use similar words (a price complaint and a value-perception complaint both mention 'expensive') and splits genuinely identical concerns phrased differently, producing a theme list that looks organized but doesn't actually map onto anything a stakeholder could act on. Requiring new themes to be checked against an existing taxonomy before being proposed fresh addresses a specific and quietly damaging pattern in repeated AI analysis: each run, given no memory of prior categorization, will invent its own fresh-sounding theme names for the same underlying concerns, which breaks period-over-period comparison — the entire value of a recurring voice-of-customer report depends on themes staying stable enough to track a trend, and a model with no instruction to preserve that continuity will silently erode it every time. The instruction to state volume as a share of the batch, not a raw count, and to flag single-comment outliers as outliers rather than themes, counters GPT-5.1's tendency to treat any repeated pattern in the input, however small the sample, as worth naming as a finding — without this check, a report on a 42-comment batch can present a theme built from two comments with the same confident framing as one built from twenty, which misleads a stakeholder deciding whether to roll back a whole system based on the report.
What you get back
Transfer/Handoff Friction: 9 of 42 comments (21%) — e.g. 'I had to explain my issue to three different people.' Response Time: 14 of 42 (33%) — largest theme this period. Business question answer: transfer-friction comments dropped from an estimated 35% of feedback pre-routing-change (per prior report) to 21% this period, suggesting the new system is helping, though response-time complaints have risen and may be a new side effect worth watching rather than confirmation of overall improvement. Outlier: one comment describing a billing issue unrelated to routing — noted, not counted as a theme. New theme proposal: none needed, all feedback fit the existing taxonomy.
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 — what Scult builds, built and maintained for you — that's Scult's day job.
EXPLORE WHAT SCULT BUILDS
