What you'll learn: How to turn raw interview transcripts into themes, jobs-to-be-done, and evidence-backed insights in minutes — with copy-paste prompts and the skills that automate the whole thing.
You finish a round of user interviews. Ten conversations, 45 minutes each. The recordings transcribed clean. And now you're staring at a novel's worth of raw transcript that you have to turn into something your team can act on.
So you start coding. Highlighting. Tagging quotes. Building a themes doc. Somewhere around interview six you lose the thread on what interview two said, scroll back, lose your place again. Two days later you have a synthesis that captures maybe 60% of what was actually said — the parts you remembered to look for. And by the time the roadmap conversation happens, nobody can trace the insight back to who said it or why it mattered.
This isn't a discipline problem. Manual synthesis at that volume is genuinely hard, and it's the first thing that gets cut when the sprint compresses. AI changes the math — not by replacing your judgment, but by reading every transcript with equal depth and handing you structured themes, JTBD, and verbatim evidence to review.
Manual vs. One-Shot Prompt vs. the Skill
There are three ways to analyze interviews with AI, and they're not equal. The difference shows up in coverage and traceability — the two things that make synthesis trustworthy.
| Manual synthesis | One-shot prompt | /user-interview-analyzer skill | |
|---|---|---|---|
| Time for 10 interviews | 1-2 days | 1-2 hours | 15-30 minutes |
| Coverage | ~60% (recency + recall bias) | Variable — depends on the prompt | Every transcript, equal depth |
| Traceability to source | Manual, often lost | Only if you ask for it | Built in — every theme cites the interview |
| Consistency run to run | Low — depends on your energy | Medium — depends on phrasing | High — same framework every time |
Manual is thorough in theory and incomplete in practice. A one-shot prompt is a massive upgrade — it's where this article starts. The skill is what you graduate to when synthesis stops being a project and becomes a button you press after every research round.
The bottleneck was never talking to customers. It was turning what they said into something the team could act on — with the evidence still attached.
The Method: Six Prompts for Interview Synthesis
These work in Claude.ai or Claude Code today. Each one is built around the same rule: be specific, cite the interview, no generic abstractions. Paste your transcripts where indicated and adjust the brackets.
1. Thematic Synthesis Across Transcripts
Start here. This is the prompt that replaces two days of manual coding.
I conducted [N] user interviews about [topic]. Transcripts below,
separated by "=== INTERVIEW [name/ID] ===".
[paste all transcripts]
Synthesize across all of them into:
1. The top 5-7 themes, ranked by how many interviews support each
2. For each theme: which interviews it appeared in, a one-line
description, and 1-2 direct verbatim quotes (with the speaker ID)
3. Patterns that confirm our assumptions about [persona]
4. Signals that contradict each other across segments
5. What's notably absent — topics I expected but nobody raised
Be specific and cite the interview for every claim. I don't want
"customers want better onboarding." I want "5 of 10 users abandoned
setup at the integrations step — Interviews 2, 4, 6, 7, 9."
2. Single-Interview Deep Analysis
Use this when one conversation was unusually rich — a power user, a churned account, a perfect-fit prospect — and deserves its own read.
Here's one user interview transcript:
[paste transcript]
Give me a deep analysis of this single conversation:
- The 3-4 most important things this person told us
- Their workflow today, step by step, in their own words
- Where they expressed friction vs. where they're satisfied
- Any quote that would change how a skeptic on my team thinks
- What I should follow up on if I talk to them again
Quote them directly. Don't smooth their language into corporate
phrasing — if they said "this part makes me want to throw my laptop,"
keep it. The texture is the signal.
3. Jobs-to-Be-Done Extraction
Themes tell you what users said. JTBD tells you what they were trying to accomplish — the level where product strategy actually happens. To go deeper on JTBD, see the full method and prompt set.
From these interview transcripts, extract the jobs-to-be-done using:
"When [situation], I want to [motivation], so I can [expected outcome]."
[paste transcripts]
For each JTBD:
- Classify it as functional, emotional, or social
- Rate confidence: strong (stated directly) vs. inferred
- Cite the interview(s) and a supporting quote
- Note where the current solution falls short of the job
Don't invent jobs to fill out a list. If only three are well-supported,
give me three. I'd rather have three I can defend than eight I can't.
4. Confirm or Challenge Existing Personas
Your personas are hypotheses. Interviews are the test. This prompt forces the data to either back them up or break them.
Here are my current personas:
[paste persona summaries, or the relevant section of personas.md]
Here are [N] interview transcripts:
[paste transcripts]
For each persona, tell me:
- What the interviews CONFIRM (with citing quotes)
- What the interviews CHALLENGE or contradict
- Any behavior, goal, or pain point my personas are missing entirely
- Whether the data suggests a segment I haven't named yet
Be willing to tell me my persona is wrong. Cite the interview for
every confirmation and every challenge. "The data broadly supports
the persona" is not an acceptable answer — show me where it does
and where it doesn't.
5. Pull Verbatim Quotes as Evidence
Themes win arguments when they come with receipts. This prompt builds your quote bank for the next roadmap review or stakeholder deck.
From these transcripts, pull the most useful verbatim quotes:
[paste transcripts]
Organize them under these headers:
- Pain points (the friction quotes)
- Desired outcomes (what "good" looks like to them)
- Competitive / switching signals (tools they use, why they'd leave)
- Surprises (anything that contradicts our roadmap assumptions)
For each quote: exact wording, speaker ID, and one line of context
on what was being discussed. Do not paraphrase or clean up grammar —
I need these exactly as said. If a quote is ambiguous out of context,
flag it rather than including it.
6. The One Pattern to Act On This Sprint
Synthesis fails when it produces a 12-theme document nobody prioritizes. This prompt forces a single decision.
Based on the themes from these interviews:
[paste your synthesis output, or the transcripts]
Tell me the ONE pattern that is both (a) strongly supported by the
data and (b) actionable this sprint. Then give me:
- The evidence: which interviews, how many, the strongest quote
- Why this one over the others (what makes it both real and urgent)
- The smallest next step that would move on it
- What I'd be wrong about if I'm wrong
One recommendation, not a ranked list. Defend it.
Related — The discovery prompts in our 30-prompt collection cover the upstream parts of research — interview synthesis, problem validation, and survey design — that feed into the work on this page.
A Necessary Word on Accuracy
These prompts work, but treat the output as a sharp first draft, not a verdict. Two things to watch.
First, hallucinated quotes are a real risk with raw chat prompting. When you paste a transcript and ask for verbatim quotes, an LLM can occasionally smooth, merge, or fabricate a line that sounds right but wasn't said. Always spot-check quotes against the source transcript before they land in a deck. The instruction "do not paraphrase, flag if ambiguous" in the prompts above reduces this — it doesn't eliminate it.
Second, the model amplifies whatever signal is in the data, including the noise. Sparse notes and poorly structured interviews produce weak synthesis no matter how good the prompt is. Garbage in, confident garbage out.
This is exactly where a purpose-built skill beats a one-shot prompt: the /user-interview-analyzer skill is structured to cite its sources and produce an artifact you can audit line by line against the transcript. You still review it. But you're reviewing structured, attributed output instead of trusting a wall of text.
Related — AI-Powered Discovery: How Claude Code Handles User Research is the broad discovery guide — research planning, multi-source feedback analysis, and turning insights into specs. This page is the deep dive on the synthesis step inside that workflow.
From Prompts to a Discovery Skill
The prompts above are the bridge. They work, and they're a real upgrade over manual coding. But you're still doing setup every time: pasting transcripts, re-specifying the format, re-explaining who your users are, re-checking the output for drift.
A skill removes that overhead. The /user-interview-analyzer skill takes a transcript and produces a structured Interview Snapshot (themes, extracted opportunities, and key quotes with attribution) every time, in the same format. The /research-synthesis-engine skill goes wider: it reads every transcript in your discovery folder at once and synthesizes themes across all of them, mapped to the personas in your context files.
That last part is the unlock. Because these skills read your personas.md and product.md context files, they don't synthesize against generic user archetypes — they map every theme to your actual segments and frame opportunities against your real product. You stop re-explaining your company on every run. The synthesis just knows.
This is what it means to use Claude as a product manager rather than as a smarter search box: the framework, the format, and the context loading happen automatically, every week, without re-prompting. For the full map of which skills handle which parts of the PM job, the AI PM workflows guide lays it out end to end.
Download the User Interview Analyzer free →
Get the full PM Operating System. Every skill, running against your own context — free for 14 days, then $39/mo, cancel anytime. Start your free trial →
Start with the prompts on this page. Graduate to the skill. And when you're ready for the whole system, every discovery skill plus the context files that make them sharp, start a free trial of the full PM Operating System.
FAQ
Can AI analyze user interview transcripts?
Yes. Drop your transcripts into a prompt or into a discovery folder, and AI reads all of them at equal depth — surfacing themes, jobs-to-be-done, and supporting quotes far faster than manual coding. The win isn't only speed. It's coverage: every transcript gets the same attention, instead of the synthesis reflecting only the interviews you remembered best.
Is it accurate, or will it hallucinate quotes?
Raw chat prompting can occasionally fabricate or smooth a quote, so always spot-check verbatim quotes against the source transcript before they go in a deck. Treat the output as a first draft that needs PM review. The /user-interview-analyzer skill reduces this risk by being structured to cite its sources — every theme traces back to a specific interview, so you can audit it line by line.
How many interviews can it handle?
A single prompt handles a batch comfortably — roughly 5-10 transcripts in one pass, depending on length. For larger or recurring research programs, the /research-synthesis-engine skill reads a whole folder of transcripts and synthesizes across all of them at once, which is where the leverage compounds.
Is the user interview analyzer free?
Yes. The User Interview Analyzer is a free skill you can download and run. The full PM Operating System — every discovery skill plus the context files that make synthesis map to your actual product — is part of the paid plan, with a free trial.
About the Author
Ron Yang is the founder of mySecond — he builds and manages PM Operating Systems for product teams. Prior to mySecond, he led product at Aha! and is a product advisor to 25+ companies.