Swipe right for a newer agent. Swipe left for an older agent, Trip Cost Splitter.
Issue 45 · 19 Oct 2026 · Medium · Free
Interview DecoderAI user interview synthesis with quotes, free
You did six customer calls. You have pages of notes, a vague feeling that people liked it, and no idea which complaints came up once and which came up every time.
Paste notes from a few customer interviews. Get what actually came up again and again, counted ("3 of 6") and backed by real quotes, plus what to ask next.
- Category
- Research
- Runs in
- ChatGPT
- Claude
- Gemini
What you put in and what you get out
Sample you can paste
Interviews for our invoicing tool, Oct 2026. Goal: why do trial users not convert? P1 (freelance designer): Signed up for the templates. Gave up on exports, "the PDF took like a minute and looked off on mobile". Pays for Bonsai now. P2 (agency owner, 8 people): "Honestly I just screenshot the invoice and send that." Didn't know you could send from the app. Wants team seats. P3 (bookkeeper): Loves the dashboard. Uses it daily. Pricing feels fine. P4 (dev contractor): Export again. "CSV broke my accounting import, dates were in US format". Almost churned. P5 (photographer): Wants dark mode, said it three times. Otherwise happy. P6 (consultant): Didn't realize invoices could be sent from the app either. Was copying the link into Gmail.
Share this agent
Interview Decoder — You put in Interviews for our invoicing tool, Oct 2026. Goal: why do trial users not convert?. You get out Flagged. Free from One Agent a Day.
When to use it
Use it after a round of customer calls, when you have pages of notes and cannot tell which complaints came up once and which came up every time.
When to skip it
It does not run interviews or replace talking to more people. Loud one-off requests are flagged, not promoted.
30-second demo
0:00
The notes
Six trial-user calls for an invoicing app. Goal: find out why trials don't convert.
How it works
After a round of customer interviews, the loudest memory wins. The person who complained the most, or the last call you did, starts to feel like "what users want." Proper synthesis means going back through every note, pulling quotes, and counting. It takes hours, so most people skip it.
This agent does that grunt work. It reads every interview against the question you were trying to answer, pulls exact quotes, groups the ones that say the same thing, and counts how many different people said it.
What to paste
One line on what you wanted to learn. Then each interview, labeled P1, P2, and so on, with who the person is. Rough notes are fine. If you have transcripts, paste those. It will only quote what's actually there.
What you get back
Patterns ranked by how many people mentioned them, not by how strongly one person felt. Each pattern has a count ("3 of 6"), two or three exact quotes with who said them, and what it might mean for your question.
Then a short list of one-off opinions it deliberately kept out of the patterns, a list of surprises (things nobody asked about but came up anyway), and follow-up questions for the next round. It won't invent a quote or round 2 people up to "most users."
Use it
Step 1
Gather every interview
Paste notes or transcripts for each call, labeled P1, P2, and so on, with a few words on who the person is. Rough notes work. Transcripts work better.
Step 2
State the question
Write one line at the top saying what you were trying to learn. The agent sorts everything against that question instead of summarizing each call.
Step 3
Run the prompt
New chat in Claude, ChatGPT, or Gemini. Paste the prompt, then the notes. For long transcripts, use a model with a big context window or run it in two halves and paste both outputs back.
Step 4
Decide what's next
Read the patterns with their counts. Anything 1 in 6 is a hunch, not a finding. Use the follow-up questions in your next round of calls.
Prompt
interview-decoder.prompt.txt
You synthesize user interviews. You are a careful researcher: you separate what people said from what you think it means, and you never inflate a finding.
INPUT
- One research question (what the interviews were meant to learn).
- Notes or transcripts from several interviews, labeled P1, P2, and so on, with a short description of each person.
HOW TO WORK
1. Read every interview. Pull short exact quotes or note lines that relate to the research question. Tag each with its participant label.
2. Group quotes that express the same underlying point. Name each group as a plain statement of what users experience ("Exports break their accounting import"), not a solution.
3. Count distinct participants per group, not mentions. One person saying it five times is still 1.
4. Rank groups by participant count. A group needs at least 2 participants to be a pattern.
5. Keep single-participant points in a separate list, even if they were said with feeling.
6. Note anything surprising that came up without being asked.
7. Write follow-up questions that would confirm or kill the top patterns.
RULES
- Only quote what is in the notes. If notes are paraphrased, say "(from notes)" instead of using quotation marks.
- Always write counts as "X of N", where N is the number of interviews.
- Do not use words like "most" or "many" unless more than half of participants said it.
- Do not suggest features or fixes. The user decides that.
- If there are fewer than 3 interviews, say the sample is too small for patterns and give quotes only.
OUTPUT
## Answer in one paragraph
What the interviews say about the research question, with counts.
## Patterns
For each, ranked: a heading with the statement and "X of N", then 2 or 3 quotes with participant labels, then one line on what it suggests for the research question.
## One-off points
Bullets with the participant label. Not patterns yet.
## Surprises
Bullets, or "None".
## Ask next time
3 to 5 follow-up questions, each tied to a pattern.Questions about Interview Decoder
How many interviews does it need?
Three or more. With fewer, it can list what people said but cannot honestly call anything a pattern.
Does it quote the interviews?
Yes. Every pattern comes with real quotes and a count of how many people said it, so you can check it against the notes.
Is Interview Decoder free?
Yes. The prompt is free to copy, download, and reuse under CC-BY-4.0. There is no sign-up on One Agent a Day. You only need access to a chat model such as ChatGPT, Claude and Gemini.
Which AI tools does Interview Decoder work in?
It was written for ChatGPT, Claude and Gemini. It is a plain-text prompt, so any capable chat model can run it. Paste the prompt first, then your input.
What do I paste into Interview Decoder?
Gather every interview. Paste notes or transcripts for each call, labeled P1, P2, and so on, with a few words on who the person is. Rough notes work. Transcripts work better.
Compare and find more
Related agents
Swipe sideways
Research3 Oct 2026 · #29Claim ExtractorPaste an article or a transcript. Get the claims, a supporting quote for each, and a question you still need to check.InTranscript excerpt, interview with a support lead, 3 Oct 2026. Not a study.OutCheckAdvanced Preview this drop
- 0:00 The excerptA support lead says a person now sees refunds over fifty euros. They refuse the phrase halved errors. Counts were not published.
- 0:08 ClaimsOne: refunds over fifty euros are no longer sent by the agent. Quote attached. Two: the lead will not say errors dropped.
- 0:16 Not a claimThe interviewer's 'halved errors' is excluded because the lead rejected it.
- 0:24 CheckWhat would settle the mood of 'calmer': the unpublished counts, which this transcript does not contain.
No daily video for this drop yet. The in and out lines above are the preview.
Research21 Sep 2026 · #17Source Credibility CheckPaste a claim and the source text. Get what that source actually supports, what it does not, and the next source to open.InClaim I want to repeat: "The refund agent cut errors by half."OutThe honest lineMedium Preview this drop
- 0:00 The hopeThe claim says errors were cut by half. The source is a vendor blog by the head of support.
- 0:08 What the text saysTickets closed without a checkbox went from 80 in February to 40 in a week of March. Error is never defined.
- 0:16 Not supportedHalf the errors is not in the text. A week is not a month. The author sells the agent.
- 0:24 The honest lineYou can say the blog reports fewer unreviewed closures after a checkbox, for one week, with no definition of error. Next source: the raw counts.
No daily video for this drop yet. The in and out lines above are the preview.
Research9 Sep 2026 · #05Paper to BriefTurn an abstract and your notes into a one-page brief that quotes the paper and marks what you have not read.InTitle: Cache-aware agents for support mail. Authors: Ibarra, Cho. 2025. Not peer reviewed…OutThe decision lineMedium Preview this drop
- 0:00 What was readAbstract plus the note that the refund tool is mocked and the human check is a checkbox.
- 0:08 The result, quotedMacro F1 0.61 to 0.74 on 400 tickets. Refunds over $50 still need a person.
- 0:16 The limitThe eval did not move live money. The brief says the appendix was not read, so latency is unknown.
- 0:24 The decision lineUseful as a design sketch for a mocked tool. Not evidence you can hand a payments review.
No daily video for this drop yet. The in and out lines above are the preview.
Tomorrow's drop
Get tomorrow's agent in your inbox
One email when the drop goes up. No digest, no course. Unsubscribe from the beehiiv email whenever you want.
Pro and custom
€9/month when Pro opens
Early look at the next drop, a better archive, and the occasional agent that does not fit the public mix. The daily agent stays free. A custom build is €49–199. Nothing is charged from this form.
Join the Pro waitlist