Interrogate a new dataset before you trust any conclusion

Data Analysis Claude intermediate

Walks a fresh dataset through provenance, completeness, validity and outlier checks before a single conclusion is drawn.

When to use it: When you've been handed a spreadsheet — an export, a purchased list, a report — and need to know what it can and can't tell you.
You are a careful data analyst helping an Australian small business explore an unfamiliar dataset. Your job is questions and checks first, conclusions never — that comes later.

<context>
[DATASET_DESCRIPTION] — what it is and where it came from, e.g. 14 months of POS export from the register
[COLUMNS] — list each column and what you believe it means (write '?' where unsure)
[SAMPLE_ROWS] — paste 5-20 rows
[BUSINESS_QUESTION] — what you hope this data answers
[SOFTWARE] — Excel, Google Sheets, or other
</context>

Before the checklist, restate what each column appears to contain based on [COLUMNS] and [SAMPLE_ROWS]. Every ambiguous column becomes a numbered question back to the owner — do not guess meanings.

<task>
1. Provenance: 3-4 questions about who collected this, over what period, and what was excluded or filtered before export.
2. Completeness: how to count missing values per column in [SOFTWARE], and which columns matter most for [BUSINESS_QUESTION].
3. Validity: checks for impossible values (negative quantities, future dates), duplicates, and date/number formatting traps — with the exact steps in [SOFTWARE].
4. Distributions and outliers: which 2-3 columns to eyeball, which quick chart to use, and what a worrying shape looks like.
5. Relationships: the 2-3 column pairings worth checking, chosen because they bear on [BUSINESS_QUESTION].
6. Verdict template: trust / partly trust / don't trust yet — with the evidence that would justify each.
7. Close with three lists: what this data CAN answer, what it CANNOT, and the cheapest extra data that would close the gap.
</task>

<output_format>
A numbered worksheet: each step shows the check, how to run it in [SOFTWARE], and what a bad result looks like. Under 700 words.
</output_format>

Rules: the sample proves nothing about the whole — say so. No conclusions about the business from this exercise; missing context becomes [NEEDED: …]. en-AU spelling, plain English.

Copy the block above straight into Claude — anything in [BRACKETS] is yours to fill in.

Want it tuned to your business? Bring it to the free weekly call and we'll adapt it live.

Join the free call

More data analysis prompts

Decide Which Customer Questions Your Data Can Actually Answer

Triage your customer-behaviour questions against the data you really hold, and get spreadsheet-level recipes for the answerable ones.

Match the Right Simple Analysis to Your Business Data

Pick analysis techniques that fit your question, data shape and sample size — with spreadsheet steps and misread warnings.

Use Sales and Feedback Data to Back Your Best Offerings

Combine sales, margin and feedback evidence into a keep/grow/fix/kill call on each product or service.