10 data analysis prompts

Designing the aggregation, generating SQL, choosing the chart and interpreting the result. Written for ChatGPT, Claude, Gemini. Square brackets mark what you fill in; everything else can be pasted as it stands.

Here is a dataset: [attach or paste a sample].

Before any analysis, report:
- Row count, and the count after removing exact duplicates
- Nulls per column, as a count and a percentage
- Min, max and quantiles for every numeric column
- For any date column, the range and every missing day

Do not interpret anything yet. I want to know what I am holding.

I want to show [what you want the reader to see] to [audience].

Recommend one chart type and say why it fits that specific question. Then say what the two obvious alternatives would mislead the reader into thinking.

Specify the axes, what is sorted by what, and what should not be included. Do not suggest a dashboard.

Schema: [paste the relevant tables and columns].
Question: [what you want to know].

Write the query, then explain in plain language what each join does and what it would do to the row count. State explicitly whether any join can multiply rows.

If the question is ambiguous given the schema, say which interpretation you chose.

I found that [variable A] and [variable B] correlate at [value] across [sample].

Tell me what can and cannot be concluded. Specifically:
- Three plausible explanations other than A causing B
- What would distinguish between them
- Whether the sample is large enough for this to mean anything

Do not write a sentence that implies causation anywhere in the answer.

Here is a sample of the data: [paste it]. It is supposed to represent [what it represents].

Look for signs that it does not: impossible values, suspiciously round numbers, gaps that line up with weekends or deploys, columns where a default has been stored as a real value, and categories that should be mutually exclusive but are not.

List what you found and what each would do to an analysis.

I want to measure [metric] for [population] over [period].

Before any query, pin down: what counts as one unit, how to handle entities that appear partway through the period, what to do with nulls, and whether to count events or distinct entities.

Give me the decisions as a list. I will confirm them, then you write the query.

Here is an analysis result: [paste it]. The reader is [role] and will decide [decision].

Write four sentences: what the number is, what it means for the decision, how confident they should be, and what would change the answer.

No methodology unless it changes how much they should trust the result.

Group A: [n and result]. Group B: [n and result].

Tell me whether this difference is worth acting on. Cover the size of the effect in practical terms, how much of it could be noise given these sample sizes, and what else differed between the groups.

End with one sentence I could say in a meeting that would be defensible.

Two sources report different values for the same thing: [source one and its figure], [source two and its figure].

List every reason two systems produce different numbers for the same metric — timing, definition, filters, deduplication, timezone, late-arriving data — and for each, say how I would check it against these two.

Order by how often each is the actual cause.

Here is everything we currently track: [list the metrics].

For each, say what decision it has informed in the last quarter. If none, mark it for removal.

Then name the one thing we are not measuring that would change a decision, and say what it would take to start.