Read a spreadsheet like an analyst, not a summariser

---
name: read-a-spreadsheet-like-an-analyst
description: "Interrogates a dataset before describing it: what's missing, what's double-counted, which column is lying, and only then what it appears to say. Use when the user asks for help with data tasks like data, analysis, spreadsheet."
license: CC-BY-4.0
metadata:
  source: https://promptsbuddy.com/prompts/read-a-spreadsheet-like-an-analyst
  author: "Sam Reyes"
  version: "1.0"
---

# Read a spreadsheet like an analyst, not a summariser

Interrogates a dataset before describing it: what's missing, what's double-counted, which column is lying, and only then what it appears to say.

## Inputs to collect first

Ask the user for anything below that they have not already given you. Do not
invent values for these.

- `DATA` — Paste the sheet as CSV, or the first 50 rows plus the column list. (e.g. Paste CSV, headers included.)
- `QUESTION` — What you're hoping it tells you. (e.g. Whether churn got worse after the March pricing change.)
- `PROVENANCE` — Where the data came from and how it was collected — this is where most errors live. (e.g. Exported from Stripe, joined to a hand-maintained CRM sheet by email.)

## Instructions

Here is a dataset and the question I want answered.

DATA:
{{DATA}}

QUESTION: {{QUESTION}}
WHERE IT CAME FROM: {{PROVENANCE}}

Do not answer the question yet. Work in this order:

**1. What this data can and cannot answer.** Given the columns and the
provenance, state plainly whether my question is answerable at all. If it is
only partly answerable, say which part.

**2. Integrity check.** Look for: missing values and whether they're missing
at random, duplicate rows, a join that could double-count, dates in mixed
formats or timezones, outliers that look like data-entry errors, and columns
whose meaning is ambiguous from the header alone. Quote example rows.

**3. The one column I'd distrust most**, and why.

**4. Now answer**, with the specific caveat that belongs next to the number.

**5. What would make this trustworthy** — the extra column, source, or check
that would move this from suggestive to solid.

Rules:
- Do not compute a summary statistic before step 2. A mean over dirty data is
  worse than no number.
- Do not infer causation from a before/after split. Say what else changed in
  that window that you would need to rule out.
- If the sample is too small to support the question, say the number and stop.

## Notes from people who use this

- Provenance is the highest-value field. Most data errors are joining errors, and the model can't see the join from the CSV.
- Step 3 is uncannily good — 'the column I'd distrust most' catches things like a status field that was backfilled.
- Never paste customer data. Anonymise, or use a structurally identical sample.

---

Read a spreadsheet like an analyst, not a summariser · by Sam Reyes · v1.0
From PromptsBuddy — https://promptsbuddy.com/prompts/read-a-spreadsheet-like-an-analyst
Licensed CC BY 4.0.

Every prompt is also available at /prompts/read-a-spreadsheet-like-an-analyst/skill.md — see all install options.