Research brief where every claim has a source
researchSeparates what is known, what is inferred, and what is unknown — with citations on the first two.
Question to analysis plan, before writing any SQL
Defines the metric, the population, and the comparison before anyone queries the warehouse.
When to use it
Someone asks 'did it work' and nobody agrees on what that means.
The prompt
Copy into your assistant
Turn this question into an analysis plan. Do not write SQL yet.
Produce:
1. The metric, defined precisely enough that two analysts would compute the same number.
2. The population and the unit of analysis (user, session, order).
3. The comparison: baseline, control, or prior period — and why that one.
4. Known confounders and how the plan handles each.
5. What result would be ambiguous, and what you would do next in that case.
Rules:
- If the question cannot be answered with the available data, say so first.
- Prefer the simplest metric that answers the question over a composite score.
- No p-values; state the decision rule instead.
Question:
<question>
Available data:
<data>What to change
Item 4 is where the plan earns its keep — pasting in the known seasonality or launch dates makes the confounders concrete.
Why it works
Most analytics arguments are definitional, not statistical. Forcing a definition that two people would compute identically resolves the argument before the query is written.
More research prompts
- Customer interview to requirementsTurns a transcript into problems and constraints, and separates them from proposed solutions.
- Research brief where every claim has a sourceSeparates what is known, what is inferred, and what is unknown — with citations on the first two.
- Summarise a paper without overclaimingSeparates what the paper shows, what it suggests, and what everyone will misquote.
Going deeper
- How to know your agent got worseRunning agents in production · Level 3
- When not to use an agentUsing agents · Level 2
- Judging a server before you install itMCP servers · Level 2