Turn ten user interviews into a ranked feature list

The problem. You did the interviews. You’ve got ten transcripts, a hundred scattered quotes, and a strong feeling that there’s a signal in there you’re not seeing. So the interviews sit in a folder and quietly become the most expensive thing you’ll never use.

The play. Hand Claude all of them at once and make it do the synthesis you keep meaning to do. The trick is to make it count and rank, not summarize — and to force it to cite the actual quotes so you can trust the output.

Setup (5 min). Get the transcripts into one folder. If they’re recordings, get transcripts first — most meeting tools export them, and Claude can work from the text. Drop them where Claude can read them (a Cowork project pointed at the folder, or a Claude Code session in that directory). If you keep a USERS.md memory file, this is where it earns its keep.

The prompt:

Read all the interview transcripts in this folder. I'm the founder; I ran
these interviews to decide what to build next.

Do this, in order:
1. Pull out every distinct problem, request, or point of friction a user
   raised. Merge duplicates that are really the same thing worded differently.
2. For each one, count how many separate interviews raised it.
3. Rank them by that count, then by intensity (how strongly the user felt it —
   quote the line that shows it).
4. Flag anything that only one person raised but raised as a dealbreaker —
   those are easy to dismiss and sometimes shouldn't be.
5. Separately, list anything a user asked for that you think I should NOT build,
   and why.

For every item, cite the exact quote and which interview it came from so I can
verify it. Don't invent or infer requests that aren't actually in the text.

What you get. A ranked list of what your users actually said, with receipts. The counting cuts through recency bias (the last interview always feels the most important). The quote-citation requirement is what makes it trustworthy — you can spot-check any line against the source in ten seconds.

What to watch for.

• It will pattern-match if you let it. The “don’t invent requests” line matters. Check the citations on anything surprising before you act on it.
• Count is a signal, not a verdict. Ten people asking for a faster export is real; it’s still your call whether it’s worth building. The tool ranks; you decide.
• The dealbreaker flag (step 4) is the sleeper. The thing one user raised as a reason they’d leave is often more important than the thing five users mildly wanted. Read that section first.

Make it repeatable. Save this prompt. Run it after every batch of interviews. Over a few rounds you’ll see which problems keep climbing the list — that’s your roadmap writing itself.