DataAdvanced

Analyze data with hypotheses, not curiosity

Avoids analyses that find pretty correlations but answer no business question.

Works well with

When to use it

When you have data and time, but no clear question.

Copy Prompt

Copy the prompt and replace the placeholders in braces.

Act as a senior data scientist. Before touching the data, I want a falsifiable hypothesis.

Business context: {{context}}
Available data: {{data}}
Pending decision: {{decision}}

Deliver:

1. THREE falsifiable HYPOTHESES, each with: claim, metric, threshold that confirms it, threshold that refutes it.
2. ANALYSIS DESIGN for the most promising hypothesis: cohorts, period, control variable, expected biases.
3. WHAT I CANNOT CONCLUDE with this data, even if the numbers look good.
4. EXPLORATORY ANALYSIS allowed before confirming (and how to avoid p-hacking).
5. RESULT in table format with the business reading, not just the number.
6. RECOMMENDATION: what I would do differently next week if it is confirmed and if it is not.

Rules: correlation is not causation, and say so when it applies. No statistical significance if the sample does not allow it. Every metric must have an exact definition.

Variables to fill in

Variable	What it means	Example
`context`	Business and problem	SaaS with 6% monthly churn
`data`	What you have	Usage events, plan, support tickets
`decision`	What you are deciding	Should we invest in guided onboarding?

Tips

If you cannot refute the hypothesis, it is not a hypothesis.
Point 3 is the one that prevents embarrassing conclusions in a meeting.
Write the threshold before seeing the data, not after.

Editorial template. Test it with real data before using it in production.