Survey Summarizer

Survey data sits in spreadsheets rotting. The default workflow is to skim 20 responses, form a biased impression, and present that as "what customers said." This skill replaces that with the THEME method — Tag, Hierarchy, Evidence, Magnitude, Exceptions — which forces every response through a structured coding pass and produces themes that survive scrutiny rather than themes that survived a coffee break.

Sample size determines what the skill will and won't say. Below 10 responses, it treats them as qualitative interviews with no percentages — each response gets individual attention. 10-29 gets counts, not percentages ("7 of 15 respondents," never "47%"). 30-99 allows percentages but flags segment comparisons as directional only. 100+ unlocks full statistical analysis with reliable segment splits when each segment has 30 or more. These guardrails matter because a "23% prefer X" headline from 13 responses is the kind of claim that survives a meeting and gets a product roadmap wrong.

Bimodal distributions get treated separately rather than averaged. If responses cluster at both 9-10 and 2-3 on satisfaction, reporting the average of 5 describes nobody. The skill detects bimodal patterns, splits the report, and investigates what distinguishes the two groups — which is usually the most important finding in the survey. Same logic for thematic clustering: a theme mentioned by five furious customers outweighs a mild suggestion from thirty, so frequency and intensity are reported as separate columns rather than collapsed into a single rank.

Verbatim quotes are preserved as the strongest evidence. The phrases customers use to describe your value are better marketing copy than anything your team will write — "it's like having a CFO without the £150K salary" beats any sentence that contains "comprehensive financial oversight." The skill flags phrases that appear repeatedly across different respondents as headline candidates, with the count, and leaves them in the customer's words rather than paraphrasing them into corporate-speak.

You get the quantitative snapshot with bimodal flags, the theme map ranked by frequency AND intensity, sentiment overlay per theme, response-rate-driven trust assessment (below 20% gets flagged for non-response bias, above 50% gets called a strong sample), the top three findings with evidence, recommended action, owner, and urgency, and a blind-spots section naming what the survey didn't capture and what the follow-up should ask. Comparative surveys get trend lines per theme and statistical significance flags on between-wave changes. The skill refuses to spin: if data is overwhelmingly negative, it stays negative; if one theme dominates, minor themes don't get equal weight for aesthetics. Observation gets separated from interpretation throughout — "23 respondents mentioned slow response times" is fact; "this suggests understaffing" is inference, and the skill labels which is which.