Prompt library
Plan an exploratory analysis that answers a question
Pre-registering what each pattern would mean is the discipline that prevents just-so stories after the charts render. The confound list and sanity gates front-load the skepticism that usually arrives too late, in the meeting where someone asks about selection bias.
Last reviewed July 17, 2026
The prompt
Plan an exploratory data analysis.
Dataset: {{dataset}}
The question behind the exploration: {{question}}
Produce:
1. Restate the question as 3-5 falsifiable sub-questions, each answerable with this data alone. Anything requiring data we do not have goes on a "cannot answer with this data" list, which is a finding in itself.
2. Per sub-question: the exact cut (variables, filters, grain), the chart or table that would answer it, and what each possible pattern would mean.
3. Sanity gates before interpretation: the 3 checks that must pass for any of this to be meaningful (sample sizes per cell, time coverage, selection effects in how the data was collected).
4. The confound list: for each sub-question, the most likely spurious explanation and the cut that would expose it.
5. Stop condition: what "done exploring" looks like, so the analysis converges to an answer instead of an infinite gallery of charts.
Order the sub-questions cheapest-first so early findings can redirect the rest.Run in idaptOpens a new chat with the prompt prefilled. Nothing sends until you press send.
Fill in the variables
| Variable | What it is | Example |
|---|---|---|
| {{dataset}} | What data exists | 12 months of support tickets with timestamps, category, CSAT |
| {{question}} | The real question | did the new help center actually reduce ticket volume |