Claude Prompts for Customer Research That Work

The short version: Claude is strong at customer research because it can hold long transcripts, reviews and survey exports in context and find patterns across all of them at once, but the prompts need structure or you get generic waffle back. Below are the exact prompts I'd use, in order, from raw data to a one-page brief, plus the bit nobody tells you: Claude can't replace talking to customers, it can only help you process what they already told you faster.

Why Claude for this job specifically

I use Claude over ChatGPT for research work because of the context window. You can paste in 20 customer interview transcripts, 300 app store reviews, or a messy survey export with open text answers, and Claude will hold all of it and reference specific lines back to you. ChatGPT does this too now, but Claude's longer context handling has been more reliable in my experience when the document gets past 15,000 words or so.

If you've not set Claude up for business use yet, I've written a separate piece on what to switch on first and the catches that's worth reading before you start feeding it customer data, especially around what you should and shouldn't paste in if any of it is personally identifiable.

The prompt stack: five stages, five prompts

Don't use one giant prompt. Customer research falls apart when you ask Claude to "analyse this and tell me what customers think" because it will hedge, summarise politely, and give you nothing you can act on. Break it into stages instead.

Stage 1: Theme extraction from raw text

Paste in your reviews, survey answers or transcripts and use this:

• "Here are [number] customer reviews/responses for [product name]. Read all of them and list the 8 to 12 recurring themes you see, whether positive or negative. For each theme, give me a one-line description, how many times it appears (roughly), and 2 direct quotes as evidence. Do not generalise, only report what's in the text."

That last line matters. Claude will try to be helpful and infer sentiment that isn't there unless you tell it to stick to the text.

Stage 2: Sentiment and intensity, not just positive/negative

• "Looking at the same data, which of these themes come up with strong emotional language (frustration, anger, delight, relief) versus mild or neutral mentions? Separate them into two lists: 'high intensity' and 'low intensity', and explain why you've placed each one there."

This is the step most people skip. A theme mentioned 40 times in a flat, neutral tone matters less than one mentioned 12 times with words like "furious" or "finally" attached to it. Intensity tells you where to act first.

Stage 3: Jobs-to-be-done framing

• "For each theme from stage 1, rewrite it as a 'job to be done' statement in the format: 'When [situation], I want to [motivation], so I can [outcome].' Base this only on language customers used, don't invent motivations they didn't state."

Stage 4: Gaps and contradictions

• "Are there any contradictions in this data? For example, customers who praise one feature but complain about something closely related. List them and quote both sides."

Contradictions are where the real insight usually sits. If half your customers love your onboarding and the other half call it confusing, that's not noise, that's a segmentation problem.

Stage 5: The one-page brief

• "Turn everything above into a one-page brief for a founder with no time to read raw data. Structure it as: top 3 findings, top 3 risks, and 3 recommended actions ranked by likely impact. Write in plain English, no jargon, no hedging."

A worked example

Say you run a small subscription box business shipping skincare products across the UK and US, and you've just pulled 180 cancellation survey responses from the past quarter. You paste them into Claude in batches of 60 (keeping under the practical limit where Claude starts losing detail on very long single pastes) and run the five-stage stack above.

Stage 1 might surface themes like "too many products I don't use," "shipping delays to the US," and "wanted to pause, not cancel." Stage 2 could show that "wanted to pause, not cancel" carries the highest emotional intensity, with words like "annoyed" and "wish I'd known" repeated. Stage 4 might reveal a contradiction: customers who complain about too many products are also the ones praising the curation quality, which tells you the problem isn't selection, it's quantity and frequency.

The one-page brief that comes out the other end could recommend adding a pause option before a cancel button, which is a cheap product fix compared to the vague alternative of "improve customer satisfaction" that most untreated survey data gives you.

Prompts for before you have data too

Claude is also useful for designing the research, not just analysing it afterwards.

• "I'm about to interview 10 customers about why they churned from [product]. Write me 8 open-ended interview questions that avoid leading language and won't just get me 'it was too expensive' as a lazy answer."
• "Review this survey draft and flag any leading or double-barrelled questions that will bias the answers."
• "Based on this list of customer segments, write 3 different outreach messages inviting each segment to a 20-minute research call, matching tone to how each group talks about us in reviews."

If you're using research findings to write the follow-up marketing, my piece on Claude prompts for marketing copy that convert picks up exactly where this one leaves off, turning customer language straight into ad copy and landing page lines.

The uncomfortable bit

Here's what most posts on this topic won't say plainly: Claude cannot do customer research for you, it can only process research you've already done. It has no access to your actual customers, no memory of past conversations with them, and no way of knowing if your 180 survey responses are representative or just the loudest 15% who bothered to reply. I've seen people treat an AI-generated brief as if it were a finished study. It's not. It's a fast first pass on data you collected, and the collection method still matters more than the analysis tool.

If your sample is skewed (say, all from customers who emailed a complaint, with nobody who quietly left happy), Claude will confidently extract clean themes from a skewed dataset and hand you a brief that sounds authoritative and is quietly wrong. The prompt stack above makes Claude better at reading what's in front of it, not better at telling you whether what's in front of it is the right data in the first place. That part is still your job.

A quick note on tools

If you're deciding between models for this kind of work, I've compared the practical differences in DeepSeek vs Claude, and if you work in a regulated area where confidentiality of customer data matters, it's worth reading how Claude handles sensitive documents in my piece on Claude for legal work, since the same caution applies to customer PII.

If this is more than you want to run yourself, working through research prompts with someone who's set this up before is quicker than trial and error, and it's the kind of thing I cover when I work as an AI consultant for small businesses.

Frequently asked questions

Can Claude analyse customer interview transcripts directly?

Yes, you can paste transcripts straight into Claude or upload them as documents, and it handles long transcripts well due to its large context window, though you should always ask it to quote directly rather than paraphrase so you can check its findings against the source.

How many customer reviews or responses should I paste into Claude at once?

In practice, batches of 50 to 70 reviews per prompt give more reliable theme extraction than dumping in hundreds at once, since very long single pastes can cause Claude to lose detail on earlier sections.

Is Claude better than ChatGPT for customer research?

Claude tends to handle longer documents with more consistency and gives more direct, less hedged summaries, which suits research work, but ChatGPT is a reasonable alternative and the better choice often comes down to which one you already use for other tasks.

Can Claude replace customer interviews altogether?

No, Claude can only process data you give it, it has no access to your customers and can't tell you if your sample is representative, so it should speed up analysis of research you've already collected rather than replace the collection itself.