Why customer interviews feel overwhelming to summarize
There’s this very specific moment during customer research where I hit a wall. It’s after a couple of good Zoom calls, each with 30 minutes of genuinely fascinating back-and-forth, stacked recordings in a folder, the transcripts slowly trickling into Otter or Fireflies or whatever I used that day… and then nothing. I promised myself I’d do summaries the next morning. But the next morning becomes a deadline sprint for something else. By the time I actually open the transcripts, I barely remember who said what.
The pain isn’t the interviews themselves — those usually go well. It’s the synthesis limbo that gets brutal. Context switching between transcription timestamps, trying to find actual insights instead of generic quotes, and that dreaded task of making a slide deck no one fully reads. 😓
So the goal: find a way to get pretty decent interview summaries **without** spending 4 hours rereading every single word. This is exactly where GPT started sneaking into my workflow, and ruined it for a few weeks before it actually worked.
What did not work at first with GPT prompts
I’ll be straight: the very first time I tried using ChatGPT to help with interview summaries, it made everything worse. I pasted a full transcript and prompted something like:
“Summarize this interview into four bullet points based on what the customer is feeling about our product.”
The response was everything I hate about fake business talk —”the user values our platform’s intuitive interface” — like, do they? Where did that come from in the transcript? ¯\_(ツ)_/¯
I tried again. “Give me 3 actionable insights from this conversation.” Nope. Still vague, surface-level observations with no specifics I could tie to the customer’s actual story.
Here’s the real issue: pasting an unedited transcript into GPT leads to garbage output because it doesn’t know what to focus on. You need prompt structure AND input formatting. Which is why the first real fix began when I started pre-editing the input *before* asking for a summary, focusing it around the actual moments that mattered.
How to prep your transcript so GPT actually helps
So this is what finally worked. Instead of dumping the full transcript into GPT and saying “Meet Mr. Transcript,” I picked out only the strong parts of the convo where something *felt* real. Usually these moments have a clear story:
– A moment where the customer says “I didn’t understand [X] at first until…”
– Questions like “Can I do [Y] with it? Because that’s my current blocker.”
– Clear frustration or workaround stories (“So the way I do it now is… it’s kinda backwards”)
I’d manually find 3 to 5 of those in the transcript. Literally just copy-paste the lines where the participant talked for more than a few lines about their setup, expectations, or emotions. Then I’d prompt like this:
> “Summarize these excerpts into key insight bullets. Each bullet should include:
> – what the user wants or expects
> – what they’re frustrated with
> – the workaround or behavior they use now”
That’s it. Simple. And yeah, I had to do the manual lifting of *finding* the good parts of the transcript, but weirdly that wasn’t as painful as reading the whole thing. Once I got used to this, the rest was surprisingly fast.
The consistent prompt format that actually worked
Eventually I started using a template-ish format that worked almost every time. Here’s one that would reliably produce solid summaries:
—
**Prompt:**
You are a UX researcher analyzing real interview excerpts. Your job is to turn 3 to 5 selected excerpts into clear insights that describe each user’s thinking.
For each excerpt, write:
1. Their expectation about the product or process.
2. What actually happened or why it didn’t meet that expectation.
3. How they dealt with that gap — workaround, feelings, behavior.
4. If relevant, a quote from the excerpt itself.
Stick to plain language. Don’t use buzzwords.
**Excerpts:**
[ Paste the 3–5 good excerpts here ]
—
This format saved me. Every single time I used this framing, GPT gave me better language than I could find on my own — and more quickly. The quote part is sneaky useful, too. Sometimes I’d reuse those directly in a presentation or write them on sticky notes when doing card sorting.
Fixing the tone GPT uses in your summaries
Even with good excerpts and a strong prompt, GPT has this tendency to sound like a consultant — too clean, too clinical. I once got a summary that included the line:
> “The user follows a non-standard path through the onboarding sequence due to a presumed lack of clarity in system cues.”
…What? No one talks like that. So to fix this, I started adding casual tone reminders. Here’s how:
Just add to the prompt: *“Write in clear, casual English, as if you’re explaining what the person said to a coworker.”*
Or even better: *“Imagine you’re writing this for your design team whiteboard.”*
With that one extra sentence, the style changed completely:
> “He didn’t realize you had to click the plus sign to add an item, so he gave up halfway and just typed it in a note instead.”
So much better. Humans talk about other humans this way. GPT just needs to be told we want *that* version.
How I batch summarized 10 interviews using this system
Okay this is where it got good. I got through 10 full customer interviews — each about 25 to 30 minutes — in under a day. Here’s the workflow I used:
1. I had Otter transcripts already synced to my laptop. These were rough but good enough to scan.
2. I did one pass through each, pulling 3–5 juicy excerpts and pasting them into a Google Doc. I marked each with the customer’s name and timestamp (just in case).
3. Once done for all 10, I opened ChatGPT in one tab and fed each doc into the same prompt template from above.
4. I saved each summary chunk into a master file.
End result: I had one document with 10 interview summary blocks, each broken down by expectations, frustrations, workarounds, and real quotes. That doc basically became my research report. 😅
Why you still need to fact check your GPT summaries
GPT doesn’t hallucinate the way people fear *if* you constrain it well. But even in good conditions, it loves to slightly over-generalize. For example, one customer said they “couldn’t find the export button on mobile unless they rotated the screen.” GPT turned that into:
> “Users struggle to export data in mobile views due to unresponsive UI.”
Uhhh… okay, slow down. There was only *one* user mentioning that, and the actual issue was the button hiding in portrait mode — not that the entire mobile UI is broken.
So what I do now is highlight anything in the summary that feels like a *leap* in logic. Then I go back to the excerpt or transcript and double-check whether that leap is supported. If not, I revise it manually. Usually I just rewrite slightly:
> “One user mentioned they couldn’t find the export button because it only appeared in landscape view.”
Way more accurate. Still useful.
Pulling themes or trends across all summaries
Once you’ve got clean summaries for each interview, GPT is surprisingly good at identifying common patterns across them — *if* you ask in the right way. What doesn’t work is saying:
“Find the top 3 themes across all this.”
That gets you generic stuff like:
– Users want a better experience
– Confusion on a few features
– Interest in new use cases
Cool. Thanks. Totally useless 😐
Instead, I built this prompt to do actual synthesis:
—
“Based on the following 10 customer interview insights, write down the following:
– 3 recurring frustrations you saw in at least 3 customers
– 2 unexpected behaviors or workarounds users invented
– 2 clear feature opportunities hinted at across multiple conversations
Back each point with 1 quote that roughly matches that idea.”
—
Again, it won’t do your job for you. But man does it accelerate the “What’s showing up the most?” phase. Crossing that threshold from raw quotes to themes in an hour felt like cheating.
Making your product team actually read it
This is the part we all dread, right? You did the work. You made interview summaries that feel human. You caught the patterns. And now half the team won’t open the doc unless you spoon-feed it.
What eventually helped me? I printed 1-sentence bullet insights on sticky notes. Seriously. Color-coded them by theme (yellow for friction, pink for hacked methods, green for surprise positives) and slapped them on a whiteboard in the team room.
Each bullet came straight from the summaries, written GPT-style but fact-checked by me. No fluff. One said:
> “Had to restart onboarding because they accidentally hit back, and that reset everything.”
Nobody missed the point of that one.
I think the best use of GPT here isn’t speed. It’s tone. It let me write *truthful* customer language, minus the cold analyst voice. That’s why people read them.