How to Take User Interview Notes You Can Synthesize

While the person is talking, your one job is to keep them talking and to notice what to ask next. That's the whole value of a live interview. The second you start transcribing, you stop hearing the thing worth chasing.

So take yourself off recording duty. You've got three ways to do it:

• Bring a second person whose only job is notes, so you can face the participant.
• Record the session, with permission, and let the words wait for you.
• If you're alone and can't record, write almost nothing. A word and a timestamp when something surprises you, then look back up.

Ask before you record, in one plain sentence: "I'd like to record this so I can pay attention instead of typing. Is that okay?" If they say no, you take the light notes instead; follow local rules.

Markers matter more than full notes. When someone says something that cuts against what you assumed, catch the moment, not a paraphrase. Later you'll want their exact words, not your tired summary of them. And use the room for the one thing only the room can do, the repeat-back: "So what you're saying is the export step is where you give up?" That one question, asked live, beats a page of quiet transcription.

The interview isn't over when the person leaves. It's over about ten minutes later, once you've turned it into evidence while it's still warm. Do this before the next session, every time. A day later you're reconstructing. Ten minutes later you're just remembering.

Keep the card to four things, and no themes yet:

• Quotes. Three to five, in their exact words, in quotation marks. "I stopped trusting it after it lost my draft twice." Not your gloss on it.
• Observations. What they did, separate from what they said. Where they hesitated. Where what they did contradicted what they claimed.
• One takeaway per topic. A single neutral sentence. Not "onboarding is confusing" but "couldn't find where to invite a teammate."
• New questions. Anything you want to ask the next person. This is how the guide gets better between sessions.

Label each card with a participant ID instead of a name. P7, not a person. You'll thank yourself at synthesis, and it keeps the notes clean of detail nobody agreed to have written down.

Notice what's missing: a theme. You don't get to decide what the study means from one conversation. Decide early and you'll spend the rest of it collecting evidence for a conclusion you already reached.

Themes come from the pile, not from any single card. Lay the cards out where you can see them at once, on a wall, a doc, a spreadsheet. Then do the part that's actually analysis.

• Group quotes and observations about the same thing, even when people used different words for it.
• Count distinct participants per group, not notes. Five quotes from one talkative person is one participant, not five.
• Hunt for the ones that don't fit. A contradiction is worth more than another confirmation.
• Keep every candidate theme pinned to the exact quotes under it. If you can't point to who said it, it isn't a finding yet.

This is where people reach for AI, and where it quietly lets them down. Paste in every transcript, type "Read all my interviews and tell me the top themes," and you get back something fluent, confident, and partly made up: themes that read well and were never said. The model is trying to sound like feedback in general, not like your seven people.

Used the other way, it earns its place. Let it take a first pass at grouping and counting, then make it show its work: which quote, which participant, for every claim. Verify each one against the timestamped transcript before it lands in a slide. Keep the participant count next to each theme, so a reader can tell one loud voice from a real pattern. The rule is simple. The transcript is evidence, the theme is your conclusion, and you never let the tool collapse the two.

Everything above assumes a note-taker or a spare hand. If you research solo, especially in person, on a home visit, a shop floor, a hallway catch after a usability test, here's how the whole loop runs on one device, and where you still do the thinking.

Start capture with one tap, from an Apple Watch or your phone. No laptop between you and the person, no bot asking to join. Open by saying who you're talking to and why: "second interview, returning power user, nearly churned last month." That framing lands in the transcript and gives every later step context. Even if you say nothing, each session is its own timestamped recording, and Pastok separates the speakers inside it, so your questions stay distinct from their answers and one participant never bleeds into the next.

After the session, there are three ways to get the evidence card, in rising order of setup:

• Just ask Chat: "Pull three to five verbatim quotes, what they did versus what they said, one neutral takeaway per topic, and new questions. No themes yet." No setup, works the first time.
• Save that as a Skill so it runs itself after every interview. Then add a second Skill that tracks one question across the whole study: "Every time someone talks about pricing, collect their exact words and which participant said it."
• When you're done, ask Chat straight for the synthesis: "Across all my interviews, what are the candidate themes, how many distinct participants back each, and where do they contradict each other?" You can push further and ask for a first-draft readout.

Pastok isn't your repository. The affinity wall stays in whatever tool you already use. What it gives you is trustworthy raw material and a fast first draft. And the rule from the last section still holds: what comes back is a draft, not a finding. That's where the transcript earns its keep. If a quote looks too clean or a theme too tidy, open the timestamped transcript in the web app and inspect the source passage. Checking takes seconds, which is the only reason anyone does it.

Two honest limits. Speaker separation isn't the same as identity, so in a noisy room it mislabels. Check before you attribute a quote. And recording is cloud-processed, so anonymize what's sensitive and get the permission you were going to ask for anyway.

Done this way, Pastok never does the thinking. It just lets you stay in the room, then makes the pile searchable and every claim checkable in seconds.