Prompts to analyze data
Last updated
Most data analysis fails before the spreadsheet opens — at the question. These four prompts are for everyone who analyzes without "analyst" in their title: sharpening a vague question into a testable one, choosing the simplest sufficient method, reading results without wishful thinking, and pressure-testing the number that looks too interesting.
AI's role here is method and honesty, not computation: the math happens in your spreadsheet or BI tool; the model keeps the reasoning straight. (Deeper, SQL-level templates live on the data-analyst page.)
The four templates
"Why are sales down?" is a mood. Make it a question data can answer.
Sharpen my vague question into analyzable ones. My question: "[PASTE, E.G. WHY ARE SIGNUPS DOWN?]". Context: [WHAT HAPPENED RECENTLY, WHAT DATA EXISTS ROUGHLY, WHAT DECISION HANGS ON THIS].
Deliver: 1) the decision behind the question ("depending on the answer, we would…") — if none, say the analysis may not be needed; 2) 4 precise sub-questions, each with: the exact comparison (what vs. what, over which period), the data needed, and what answer would mean what; 3) the confounders that could fake each answer (seasonality, mix, tracking changes, one-off events); 4) which sub-question to answer first and why it's the cheapest informative one.
You have the question and the data. Before an afternoon of YouTube statistics: what's the simplest method that answers it?
Recommend the analysis method. Question: [THE SHARPENED QUESTION]. Data: [WHAT YOU HAVE: ROWS OF WHAT, HOW MANY, WHICH COLUMNS, TIME RANGE]. My tools: [EXCEL/SHEETS/BI TOOL/SQL]. My stats comfort: [HONEST LEVEL]. Deliver: 1) the simplest method that genuinely answers the question (comparison of means? cohort table? before/after with control? just a pivot?) — and why fancier isn't needed; 2) step-by-step in MY tool, named functions/menus included; 3) the sample-size reality: at my n, what difference is even detectable — rough rule, no false precision; 4) the 2 mistakes people make with exactly this method, and how each would show up in my numbers. Don't compute — coach.
The numbers are in. Now the dangerous part: deciding what they say — versus what you hoped they'd say.
Help me read my results without wishful thinking. Question was: [X]. My hypothesis going in — honestly: [WHAT YOU EXPECTED/WANTED]. Results: [PASTE TABLE/NUMBERS, WITH n PER GROUP] Deliver: 1) what the data supports, each statement with its numbers and comparison base; 2) what it does NOT support that I'll be tempted to claim — given my stated hypothesis, be specific; 3) alternative explanations for the pattern, ranked by plausibility; 4) correlation/causation check on every causal-sounding claim; 5) the honest summary sentence I can put in a slide, caveat included but not paralyzing. Where my n is small, say what that does to each conclusion — with the number, not just "be careful".
One number looks amazing (or alarming). Before it headlines a slide: is it an insight or an artifact?
Pressure-test this number before I share it. The number: [WHAT IT IS + WHAT IT SEEMS TO SHOW]. How it was produced: [STEPS, FILTERS, PERIOD, SOURCE]. What's historically normal: [PRIORS/PAST VALUES]. Run: 1) magnitude check — is it plausible? Show the back-of-envelope math; 2) the usual suspects, each rated likely/unlikely HERE with a reason: definition changed, population/mix shifted, tracking or pipeline gap, small denominator, cherry-picked window, survivorship; 3) the 3 quickest checks to falsify it (described concretely for my setup); 4) if it survives all that: the caveat sentence that still belongs next to it. Your job is to kill this number if it can be killed — be thorough, not kind.
Frequently asked questions
Can I just paste my spreadsheet and ask "what do you see?"
You'll get patterns — some real, some pareidolia, delivered with equal confidence. The sequence here (question → method → honest reading) exists because open-ended "find insights" prompting is where AI analysis goes wrong. Aggregates in, specific questions, always.
Should I let the AI do the actual calculations?
Spreadsheet and BI tools calculate reliably; chat models sometimes don't, and errors look identical to correct answers. The templates keep computation in your tools (AI coaches the steps) — except AI environments with real code execution, where the code, not the chat, does the math.
What about statistical significance?
For business decisions, effect size and detectability usually matter more than p-values — the method template's "what's detectable at my n" check covers the practical core. When formal testing is genuinely needed (experiments, publications), that's the moment for a statistician or the researcher templates.
Is my data safe to paste?
Aggregate what you can (the templates want tables, not raw rows), strip identifiers, follow company policy for anything sensitive. Row-level personal data belongs in governed tools, not consumer chat — same rule as on the analyst page.