First Look at a New Dataset

Get an interrogation plan for an unfamiliar dataset: what to check first, what would make it untrustworthy, and which questions it genuinely can answer.

Act as a senior data analyst mentoring me through a dataset I have never seen before.

What I have:
- Source: [WHERE IT CAME FROM AND WHO PRODUCED IT]
- Grain: [WHAT ONE ROW REPRESENTS — e.g., one order, one user per day]
- Columns and types:
"""
[PASTE THE HEADER PLUS A FEW SAMPLE ROWS, OR A COLUMN LIST WITH TYPES]
"""
- Time range and size: [DATES, ROW COUNT]
- The question I am being asked to answer: [THE ACTUAL BUSINESS QUESTION]

Deliver:
1. Ten checks to run before trusting anything, ordered by how badly each failure would mislead me — include the exact thing to compute for each
2. The traps specific to this data: which columns are likely to be double counted, which nulls probably mean zero and which mean unknown, what survivorship or selection bias the source implies
3. Whether the grain actually supports my question. If it does not, say so plainly and state what data I would need
4. The three most useful cuts of this data for my question, and what a boring result would look like for each so I do not over-read noise
5. Two conclusions this dataset can support and two it cannot, no matter how the chart looks
6. The caveat paragraph I should attach to any result I share

Do not summarise the data or invent numbers — I have not given you the full file. Give me the plan.

How to use this prompt

• Run this before you build a single chart; the wrong grain has ruined more analyses than the wrong formula.
• Take section 5 seriously — telling a stakeholder what the data cannot show is the most valuable thing an analyst does.
• Keep the caveat paragraph attached to the result when you share it.

Why this prompt works

The model cannot see your file, so asking it to analyse the data would produce fiction. Asking instead for an interrogation plan uses what it is genuinely good at — knowing the failure modes of common data shapes — and the explicit "state what it cannot support" section is a direct defence against confirmation bias.

Variations to try

• Add "I have to present this to a skeptical CFO" and ask which check they will attack first.
• Ask for the ten checks written as SQL queries against your table names.
• Follow up with your actual findings and ask what alternative explanation you have not ruled out.

Common mistakes to avoid

• Asking the model to "analyse this data" from a ten-row sample. It will happily describe the whole file it never saw.
• Skipping the grain question and then double counting every joined row.
• Treating a plausible narrative as a finding without running the checks.

Works well with