AI prompts for data analysts

"Can you look into why signups dropped?" — the ask is a feeling; the plan makes it answerable.

Turn this stakeholder ask into an analysis plan: "[PASTE THE ASK VERBATIM]".
Context: [PRODUCT/BUSINESS, THE METRIC'S DEFINITION, WHAT CHANGED RECENTLY, DECISION THIS INFORMS].
Available data: [TABLES/SOURCES + GRAIN, ROUGHLY].

Deliver: 1) the decision-relevant question, restated precisely (what would we DO differently depending on the answer?); 2) 4–6 testable hypotheses, ranked by prior plausibility, each with: the cut/comparison that tests it, the data needed, what result would confirm vs. kill it; 3) known confounders to control (seasonality, mix shifts, tracking changes); 4) the order to test them (cheapest-informative first); 5) 3 clarifying questions for the stakeholder before I start.
Do not write SQL yet.

The query runs and returns numbers. Whether they're the right numbers is a different question.

Review the SQL below against the schema and intent. Report everything suspicious with confidence labels — I filter, you don't.

Hunt specifically: 1) join fan-out (grain changes that inflate counts/sums); 2) NULL handling in filters and aggregations; 3) date-boundary bugs (timezones, half-open vs. closed intervals, "last 30 days" definitions); 4) filters that silently drop rows (INNER JOIN as accidental filter); 5) window/aggregate mixing; 6) whether the query matches the stated intent at all.
For each finding: the line, why it's wrong or risky, the concrete scenario where it produces a wrong number, and the fix.

Intent: [WHAT THE NUMBER SHOULD MEAN, PRECISELY]
Schema + grain: [TABLES, KEYS, ONE ROW = WHAT]
SQL:
[PASTE]

The finding is solid; the default bar chart isn't. Choose the visual on communication grounds.

Recommend the visualization. Finding: [THE INSIGHT IN ONE SENTENCE, WITH NUMBERS]. Audience: [EXEC / PM / TECHNICAL]. Medium: [SLIDE / DASHBOARD / DOC]. The action we want from the reader: [WHAT THEY SHOULD DECIDE/DO].

Deliver: 1) top recommendation with WHY (what comparison the chart type makes easy — that's the criterion); 2) runner-up and when it'd be better; 3) the trap chart people would default to here and why it misleads; 4) spec for the winner: axes, scale (zero-based? log?), sorting, annotation text (write the actual annotation), what to grey out, title as the takeaway sentence (not "Revenue by Month"); 5) accessibility: color choices safe for color-blind readers.
Data shape: [PASTE AGGREGATED DATA OR DESCRIBE: DIMENSIONS × MEASURES]

Analysis done. Now the paragraphs — where precision usually dies of enthusiasm.

Write the findings section from my results below. Audience: [WHO]. Decision context: [WHAT THIS INFORMS].

Rules: every claim carries its number and comparison base; effect sizes with denominators ("+40 % = from 5 to 7 per week"); uncertainty stated where n is small or the window short — n in parentheses; correlation language unless the design supports causal language — if I wrote "drove/caused" without a causal design, downgrade it and tell me; anything the data does NOT show but readers will assume goes in a "What this doesn't tell us" box.
Format: 1) three-sentence summary (finding, size, recommended action); 2) findings, one paragraph each; 3) the doesn't-tell-us box; 4) next analysis worth doing.

Results (tables/aggregates):
[PASTE]

The number looks surprising. Surprising numbers are either insights or bugs — check before the CEO quotes it.

Sanity-check this result before I publish it. Result: [THE NUMBER + WHAT IT CLAIMS]. How it was computed: [QUERY LOGIC / STEPS IN PROSE]. Base rates and priors: [WHAT'S NORMAL FOR THIS METRIC, LAST PERIODS].

Work through: 1) magnitude check — is this plausible against the priors? Show the quick math; 2) the 5 most likely ways this number is wrong (definition drift, population change, tracking gaps, mix effect, denominator shifts) — rate each likely/unlikely HERE with reason; 3) the 3 fastest falsification checks I should run (queries described in words) before believing it; 4) if it survives: how to phrase the caveat that still applies.
Be the skeptic, not the cheerleader — your job is to kill this number if it can be killed.