Cluster qualitative feedback into themes with frequency counts
Fifty comments in different words and five real underlying themes are separated by clustering someone has to actually do.
- Use it for
- Anyone with a pile of open-ended comments or feedback who needs it organized into real, countable themes.
A large set of open-ended comments — survey responses, support tickets, app reviews — contains real underlying themes, but reading through them linearly makes it hard to actually count how often each theme shows up, since the same idea gets phrased differently every time. This prompt clusters the raw comments into distinct themes, counts how often each one actually appears, and summarizes the sentiment within each cluster.
When not to use this
For a small set of comments easy to read in full, formal clustering adds overhead without much benefit — just read them directly.
Fill in
Here are the comments: {{comments}}
Cluster these into distinct thematic buckets — group comments that
express the same underlying idea even when worded differently. For
each theme, give me a count of how many comments fall into it and a
one-line summary of the sentiment within that group. Flag any comment
that doesn't clearly fit an existing theme instead of forcing it into
one.Why it works
Requiring an explicit count per theme, and flagging comments that don't clearly fit, is what makes the clustering actually quantitative and honest — a loose thematic summary without counts can't tell you which issue is genuinely most common, and forcing every comment into an existing bucket hides real outliers that might matter.
Common mistakes
- ×Producing themes without an actual count, leaving you unable to tell which one is most significant.
- ×Forcing every comment into an existing theme instead of flagging genuine outliers.
Follow-ups
- The largest theme is surprising — help me dig into a sample of those specific comments in more depth.
- New comments came in — help me add them to these existing themes rather than reclustering from scratch.
Did this work?
You have distinct themes with an accurate frequency count each, where every comment maps clearly to exactly one theme, not a loose set of overlapping buckets.
Tested on claude-opus-5. Evidence status is draft; it moves to battle-tested only on recorded runs, never by hand.