Prompting Guide
How to Prompt Claude to Analyze Survey Results
A hundred open-ended answers are a wall of text until something sorts them into themes.
Anthropic's Claude is good at reading a pile of survey responses, especially the open-ended ones nobody has time to read individually, and pulling out the themes, the outliers, and the quotes that actually illustrate a point. It isn't good at catching a count that doesn't match your spreadsheet unless you ask it to show its work, which is why the habit that matters most here isn't the prompt, it's checking a sample of the raw data against whatever Claude reports back before you repeat a number in front of anyone else.
To prompt Claude to analyze survey results, upload your response data as a file, CSV and XLSX both work, and ask for specific output: the main themes in open-ended answers, how responses break down by a specific question or segment, and direct quotes that illustrate each theme. Ask Claude to state how many responses a theme appeared in, then spot-check a handful of the raw responses yourself against that count before you treat any number as final.
Published
Free lesson · no account neededMeet ClaudeWhy Claude?6 minStart the free lessonKey Takeaways
- Upload the actual response file rather than pasting in a summary. Claude can work directly with a CSV or spreadsheet of raw responses, which preserves detail a summary would lose.
- Ask for themes with counts attached, not just a list of impressions. "This came up in 12 of 40 responses" is checkable; "several people mentioned this" is not.
- Request direct quotes alongside each theme. A theme backed by two or three actual quotes is more convincing and more checkable than a theme stated as a bare conclusion.
- Spot-check a handful of raw responses against whatever count or theme Claude reports. This is the single habit that catches a miscount before it ends up in a report someone else repeats.
- Ask what doesn't fit the main themes, not just what does. The outlier responses are often where the most useful, specific feedback is hiding.
Copy and Paste
Prompts You Can Use Today
Finding themes in open-ended responses
I've uploaded a file of survey responses. Focus on the open-ended question asking [paste the exact question text]. Identify the main themes across all the responses, state how many responses each theme appeared in, and give me two or three direct quotes for each theme that illustrate it clearly.
Name the exact question you want analyzed if the file has multiple open-ended questions; otherwise Claude may blend answers to different questions together.
Breaking results down by a segment
Using the uploaded survey data, compare how responses to [specific question] differ between [segment, e.g. new customers vs. returning customers, or by department]. Show the breakdown as counts or percentages for each segment, and note any difference that looks meaningful rather than just noise from a small sample.
Ask Claude to flag when a segment is too small to draw a real conclusion from, rather than reporting a percentage that sounds precise but isn't meaningful.
Checking a reported theme against the raw data
You told me [theme] came up in [number] of the responses to this question. Show me the actual text of five of those responses so I can verify the theme was applied correctly, not just read your summary of them.
Run this before repeating any specific count in a report; it's the fastest way to catch a theme that was applied too loosely.
Surfacing the outliers
Beyond the main themes you already identified, what are the two or three responses that don't fit any of them? Show me the actual text of those responses. I want to see what's unusual, not just what's common.
Outlier responses are often where a genuinely new piece of feedback or an edge case worth investigating is hiding.
Turning the analysis into a short summary for stakeholders
Based on the themes and counts we've already identified from this survey, write a short summary for [audience, e.g. the product team] covering the top three findings, one representative quote for each, and one thing worth investigating further. Keep it under half a page.
Do this as a final step after you've already checked the themes against the raw data, not as the first thing you ask for.
The Structure
What a survey-analysis prompt needs to produce something checkable
The same four requests turn a vague summary into something you can verify before you repeat it.
The actual response file, uploaded directly
Upload your survey export as a CSV or spreadsheet rather than pasting a summary you wrote yourself. Claude working from the raw data catches detail a secondhand summary would lose.
The exact question you want analyzed
If the file has several open-ended questions, name the specific one. Otherwise Claude may blend answers from different questions together into one set of themes.
Counts attached to every theme
Ask how many responses each theme appeared in, not just whether it came up. A number you can check against the raw data is far more useful than "many respondents said."
Direct quotes, not just conclusions
Ask for two or three actual quotes per theme. A theme backed by real quotes is both more convincing in a report and easier for you to verify was applied correctly.
What Claude Can Read
Uploading your response data
Claude can read documents in several common formats, including CSV files, directly in a chat. For spreadsheet files, enabling Claude's code execution and file creation setting in your account is required to upload XLSX files, which also unlocks data analysis directly on the spreadsheet rather than just reading it as text; CSV and TSV files work for this kind of analysis once that setting is on. Export your survey results in whichever of these formats your survey tool supports, and upload the file directly rather than copying and pasting rows of data into the chat.
For a typical survey export, this means Claude can work with the actual rows and columns, which produces more reliable counts than asking it to work from a summary table you built yourself. If your export has a lot of columns you don't need for this particular analysis, it's still usually faster to upload the full file and tell Claude which columns matter than to manually trim the file first.
Keep in mind the upload is still bound by ordinary file-size limits, so an unusually large export, tens of thousands of responses, may need to be split or filtered down to the relevant columns and rows before it uploads cleanly. For most team-level or departmental surveys this isn't a practical concern, but a company-wide survey with a long open-ended question and tens of thousands of respondents is exactly the case where trimming the file to the columns you actually need first will save you a failed upload and a second attempt.
Common Mistakes
What turns an analysis into something you can't trust
These are the habits that let a wrong number travel from a chat into a report before anyone catches it.
Repeating a count without checking it
A specific-sounding number like "this came up in 23 responses" feels more trustworthy than it is until you've actually pulled a sample and confirmed the theme was applied correctly.
Treating a tiny segment's percentage as meaningful
A segment with eight responses can produce a dramatic-sounding percentage that's really just noise. Ask explicitly whether a segment is large enough to draw a conclusion from.
Only asking about what's common
Focusing only on the main themes skips the outlier responses, which are frequently where the most specific, actionable feedback is sitting unseen.
Pasting a summary instead of uploading the real file
A summary you typed up yourself has already lost detail before Claude even sees it. Upload the actual export whenever the format allows it.
The Check That Matters Most
Why spot-checking the raw data is the one habit worth building
Of everything in this process, the single habit that catches the most real mistakes is pulling a handful of actual responses and reading them yourself against whatever theme or count Claude reported. This isn't about distrust of the analysis in general; it's that any summarization process, human or otherwise, occasionally groups a borderline response into the wrong theme, and the only way to catch that is to look at a sample of the source material directly.
A practical version of this habit: before you repeat any specific number in a meeting or a written report, ask Claude to show you the actual text of five responses it counted toward that theme. This takes under a minute and either confirms the count or reveals that a handful of responses were stretched to fit a theme they don't really belong to.
This matters more as the stakes of the analysis go up. A quick gut-check on an internal team survey can tolerate a little slack. A survey result that's going to inform a product decision, a pricing change, or something presented to leadership deserves the extra minute of verification before the number becomes the thing everyone remembers and repeats.
From Analysis To Action
Turning themes into something a team can act on
Once you have verified themes and counts, the most useful next step is usually asking what's worth investigating further, rather than treating the theme list itself as the final output. A theme that shows up in a meaningful number of responses is a finding; what to actually do about it is a separate decision that depends on context Claude doesn't have, like your team's current priorities and capacity.
When you write up the findings for other people, keep the verified quotes attached rather than only the summarized theme. A reader who sees an actual quote alongside a conclusion can judge for themselves whether the theme is being represented fairly, which builds more trust in the analysis than a bare assertion ever will. It also makes the write-up more durable: months later, when someone asks why a decision was made, the quotes are still there to point back to, rather than a paraphrase nobody can trace to an actual respondent.
It's also worth revisiting the same uploaded file a second time once you know what you're looking for. A first pass tends to surface the broad themes; a second, more targeted prompt, asking specifically about one theme, one segment, or one surprising quote, often turns up detail the first pass summarized past. Treating the analysis as a single request rather than a short back-and-forth tends to leave real detail on the table.
The analysis is only as good as the check you run on it; treat every theme and count as a strong first pass worth verifying, not as a finished, citable result on its own.
A theme you've verified is worth more than ten you haven't
Reading survey results well is a skill that pays off every time your team runs another one. Nightschool AI's free lessons walk through real workplace analysis tasks with Claude, not toy examples.
Frequently Asked Questions
What file format should I use to upload survey results to Claude?
A CSV file works directly. For an XLSX spreadsheet, you need to enable Claude's code execution and file creation setting in your account first; CSV and TSV files are supported for data analysis once that setting is on.
How do I get Claude to find themes in open-ended survey answers?
Name the exact question you want analyzed, and ask for the main themes with a count of how many responses each one appeared in, plus two or three direct quotes per theme so you can verify it.
How do I know if a count Claude gives me is actually accurate?
Ask it to show you the actual text of a handful of the responses it counted toward a theme, and read them yourself. This quick spot-check catches most mistakes before you repeat the number elsewhere.
Can Claude compare survey results across different groups or segments?
Yes. Ask it to break results down by the segment you care about, like new versus returning customers, and ask it to flag when a segment is too small to draw a meaningful conclusion from.
Should I only ask about the most common themes?
No. Ask specifically what doesn't fit the main themes. The outlier responses are often where the most specific and useful feedback is sitting, and a theme-only analysis tends to skip over them.
Is there a limit to how many survey responses Claude can analyze at once?
Uploads are bound by ordinary file-size limits, so an unusually large export may need filtering to the relevant rows and columns first. For most team or departmental surveys with a few hundred responses, this isn't a practical issue you'll run into.
How should I write up the findings once I've verified them?
Keep the verified direct quotes attached to each theme rather than presenting only the summary. A reader who can see the actual quote can judge the finding for themselves, which builds more trust than a bare conclusion ever does on its own.
Can Claude tell me what to do about a theme it finds in the data?
It can suggest options, but deciding what to act on depends on your team's priorities and capacity, which Claude doesn't have visibility into. Treat the theme as a finding, and the response to it as a separate decision your team is better placed to make than a chat.
Verify the theme before you repeat the number
The prompts are reusable every time you run a survey. The spot-check habit is the part worth keeping.