Our NPS was 34. It had been 34 the previous quarter too. Every review meeting turned into the same three-way standoff: engineering was certain it was a performance problem, sales was certain it was pricing, and our CEO was certain it was onboarding. The funny part is that we had 200 NPS free-text comments sitting in a spreadsheet that, I am fairly sure, no single human had read all the way through. We had each sampled about thirty and, surprise, each of us had found the quotes that confirmed our existing theory. It was a confirmation-bias buffet.
The afternoon I stopped guessing
I exported all 200 comments to a plain text list, stripped the names, and dropped the whole thing into the feedback_batch field of the Cluster Raw User Feedback into Ranked Themes with Verbatim Quotes prompt. I set the product name and told it the source was an NPS survey. I picked GPT 5.2 because I wanted the careful read, not the fast one, and I had 200 messy human comments full of typos and run-ons.
Fifteen seconds later I had six ranked themes. The number one theme, by a mile, was onboarding confusion: 68 of the 200 comments touched it, sentiment Negative, and the recommended action was the unglamorous but correct 'add an interactive first-run checklist that marks setup steps complete.' Theme two was a specific, repeatable feature request for bulk export. Crucially, it was not the vague word 'performance' that engineering had been bracing for. Actual performance complaints were theme four, and they were mostly about one slow report, not the whole app.
Why the verbatim quotes mattered more than the counts
The counts got attention, but the exact quotes ended the war. When you paraphrase customer feedback, everyone re-paraphrases it into their own theory. When the readout has the literal sentence 'I signed up, saw an empty dashboard, and had no idea what to do next,' there is nothing left to spin. That is not 'performance.' That is not 'pricing.' That is a person staring at an empty room. Having three of those sitting under the onboarding theme did more to align the room than any chart I have ever made.
I think the deeper reason this worked is that raw counts are arguable but specific human sentences are not. Anyone can dispute 'onboarding is the biggest theme' by claiming the clustering is wrong. Nobody can dispute a customer who literally wrote, in their own words, that they did not know what to do after signing up. The quote is evidence in a way a bar chart never is. Our previous attempts at this kind of analysis always died in the 'well, it depends how you categorize it' swamp. The verbatim requirement drains that swamp, because the categories are just folders and the quotes are the actual proof.
- We had our first roadmap meeting in months where nobody disputed the priority order, because the theme data dissolved the anecdote wars.
- Our CEO saw that pricing was theme six, raised by 12 of 200 people, and quietly retired it as the supposed main driver.
- We shipped the onboarding checklist about six weeks later and NPS moved from 34 to 47 in the next survey cycle.
- The outliers-to-watch section flagged two early mentions of a sync bug that became a real fire a month later, which means it basically gave us a one-month warning.
The one rule I never relax
The prompt is told, in capital letters, never to invent a quote. I still spot-check. I pick two quotes per theme and Ctrl-F them in the original spreadsheet to confirm they are real and exact. They always have been, but the day I stop checking is the day I will get burned, because a fabricated-but-plausible quote in a leadership readout is a credibility bomb. Trust the clustering, verify the quotes. That is the deal.
One more practical note for anyone who tries this. The model is good but it is not psychic, so the quality of the clustering scales with how clean your input is. The first time I ran it, I left in the survey metadata, the timestamps, the score numbers, all of it, and the themes were noisier. The second time I stripped it down to just the free-text comments, one per line, and the clusters were noticeably sharper. Five minutes of cleanup buys you a much better readout. It is the closest thing to a free lunch I have found in this whole workflow.
This is now my standard move on any batch of 30 or more pieces of feedback. Grab the Cluster Raw User Feedback into Ranked Themes with Verbatim Quotes prompt on Prompt Dock and run it on your own pile of comments before your next strategy meeting. When a theme turns into a real bet, I drop it straight into my Turn a Rough Idea or Slack Thread into a Team-Ready PRD prompt and the problem statement basically writes itself, in the customer's own words.