Open-Ended Survey Response Analyzer

Turn hundreds of free-text survey answers into themes with counts and verbatim quotes, without letting the model smooth away the disagreement.

Act as a research analyst. Code and summarise open-ended survey responses.

Survey context:
- Question asked: "[THE EXACT QUESTION RESPONDENTS SAW]"
- Who answered: [AUDIENCE, HOW THEY WERE RECRUITED]
- Number of responses: [COUNT]
- What decision this feeds: [WHAT WE WILL DO DIFFERENTLY BASED ON THIS]

Responses:
"""
[PASTE RESPONSES, ONE PER LINE]
"""

Deliver:
1. A theme codebook: 5-9 themes, each with a name, a one-line definition, and the inclusion rule you used
2. A table of themes with the count and percentage of responses in each, and the leftover "does not fit" bucket — do not force every response into a theme
3. Two or three verbatim quotes per theme, copied exactly, chosen for how well they represent the theme rather than how quotable they are
4. The strongest disagreement in the data: where respondents want opposite things, with quotes from both sides
5. What the responses do NOT tell us, and the follow-up question that would resolve the biggest ambiguity
6. Three actions this supports, ranked by how much evidence backs each

Rules: quote exactly, never paraphrase into a quote. If a theme has fewer than three responses, label it as thin evidence rather than a finding.

How to use this prompt

• Paste responses in batches if the list is long, then ask the model to merge the codebooks and re-count.
• Spot-check a sample of the codings yourself — the counts are only as good as the model's consistency.
• Lead any presentation with section 4, the disagreement. It is where the useful decisions are.

Why this prompt works

Left alone, models flatten qualitative data into an agreeable summary. Demanding counts, a residual bucket, exact quotes, and an explicit disagreement section keeps the messiness visible, which is where the signal usually lives — and the thin-evidence rule stops a two-person opinion becoming a headline.

Variations to try

• Add "segment the themes by [PLAN TYPE / TENURE / REGION]" if you include that column with each response.
• Ask for a sentiment split per theme, with the caveat that sentiment coding on short answers is unreliable.
• Run the same responses twice in separate chats and compare codebooks to gauge stability.

Common mistakes to avoid

• Removing personal information after pasting rather than before. Strip names and emails first.
• Reporting percentages from 30 responses as if they were a survey of your market.
• Accepting a polished quote that does not appear anywhere in your raw data.

Works well with