Ask a better
question.

ASK A BETTER QUESTION
A project handoff by Matt Klein

Use this when a team asks for a report, dashboard, integration, or analysis and you want to establish what would actually improve a decision. Attach the request and an ordinary example. Start with available evidence.

YOUR ASSIGNMENT

Work backward from the decision to the question, the model, and the sources. Challenge the requested output before spending time building it. A more sophisticated answer to the wrong question is still the wrong answer.

1. Write the decision in a sentence.

"[Person] needs to choose [action] by [time], because [business consequence]."

Record the original request, your best interpretation, and a credible alternative. State what we would do differently depending on the answer. If every possible answer leads to the same action, question whether this analysis is worth doing.

Identify what is reversible, what is expensive to reverse, and the acceptable uncertainty. Do not invent a deadline, budget, or decision owner. Mark those as provisional when the evidence is incomplete.

2. Select the model that fits.

Choose the method and explain its limits. Contribution requires unit economics; availability requires a stock-flow model; mismatched totals require reconciliation; delays require a process trace; an uncertain choice may require scenarios.

Define the grain, units, time basis, exclusions, and assumptions before calculating. Separate revenue, contribution, cash, and capacity where they imply different choices. Calculate deterministically where possible. Use AI to investigate and interpret, not to supply missing facts.

3. Connect only what changes the answer.

Map each necessary input: definition | identifier | source | owner | date | status.

Status must distinguish named, available, read, and checked. Test joins with real identifiers and units. Check duplicates, missing matches, pagination, and coverage when applicable. A successful API response does not establish that the right records were retrieved.

Trace one permitted read or supplied sample through to the answer. If live tools are unavailable, use the sample and state what remains untested. Keep secrets out of the conversation.

Rank up to three missing inputs by whether they could reverse the decision. Specify the cheapest observation that would resolve the most consequential uncertainty.

4. Find where the recommendation changes.

Independently check the decisive calculation. Test a competing explanation and at least one boundary case. Show the threshold at which your recommendation reverses, or explain why the evidence cannot establish one.

For a pricing request, check pack size, channel, availability, shipping, promotions, and realized contribution before assuming price is the problem. For a performance request, separate volume, mix, price, and returns before attributing a change to one cause.

5. Make a recommendation we can use.

Compare a process correction, a small repeatable tool, and a larger integration only where each is credible. Include the maintenance burden and the failure that could mislead the user. Choose the smallest approach that supports the decision.

Prepare the first test now with the supplied material. State what would justify expansion and what would make us stop. Keep an uncertain forecast separate from an observed result.

RETURN

A decision brief of roughly one page:
- The decision, person acting, and recommendation.
- The checked answer and its calculation.
- The strongest competing explanation.
- The reversal point and the missing input that matters most.
- The first test, owner if known, and stop or expand condition.

Include the source map and calculation as supporting material. Use source locations a person can inspect.

ACCEPTANCE CHECK

Someone can explain why the proposed action follows from the evidence, reproduce the decisive number, and say what new evidence would change the recommendation. If that is impossible, return a specific unresolved decision and its cheapest next test.

SMALL EXAMPLE, FICTIONAL

"Match the competitor's price." Our two-pack is $24; the competitor's single unit is $14. We are already lower per unit, $12 versus $14. Our total variable cost is $18 per pack, so contribution is $6. At $21 it falls to $3, requiring twice the volume to preserve total contribution, assuming costs and mix stay constant.

First compare equivalent offers and availability. Do not recommend a discount from mismatched pack prices. A small price test may still make sense, but it needs evidence of demand response and enough stock to support it.

MY STARTING CONTEXT

The request:
Who needs the answer:
What we would do differently:
Available material:
Constraints:

Start from what I supply. Ask only for information that changes the decision or the first test.