Skip to Content

Ask it well

This step is the question itself: the words you type once you know what you want and what you are willing to share. A prompt written with a little care gets an answer you can check; one written on impulse gets an answer that sounds right and hides where it might be wrong.

Why it matters

The tool answers the question in front of it, not the one in your head. Leave out the country, the year or the kind of account, and it fills the gap with the most common case it has seen, which may not be yours. Tell it what you hope is true and it tends to agree. The quality of the answer is set before the reply arrives, by what you put in.

Asking well is also what makes checking possible. An answer to a vague question is a vague answer, with nothing to hold against a source. An answer to a precise question states a figure, a rule or a reason that you can test in the next step. You are not trying to make the tool right; you are trying to make it specific enough to be wrong in a way you can see.

What actually happens

State the facts the answer depends on

Give the tool the handful of facts you kept after the last step: the type of product, the amounts in round numbers, the country or state, the year, and any deadline. Say what you already know. “I have a fixed-rate mortgage in the United States, five years into a thirty-year term” gets a different and better answer than “should I overpay my mortgage”. Every fact you leave out, the tool will assume, and it will not tell you it assumed it.

Ask for the reasoning, not just the verdict

Ask the tool to show its working: the rule it is applying, the arithmetic, the assumptions behind a projection. A verdict on its own cannot be checked; the steps that lead to it can be checked one at a time. If it gives a figure, ask where the figure comes from and which year it applies to. If it names a rule, ask which document the rule is in. The answers to those follow-ups are what you carry into the next step.

Keep the question open

A leading question tells the tool the answer you want, and it usually obliges. “Why is paying off my car loan early a good idea” gets reasons it is a good idea; “what are the arguments for and against paying off my car loan early” gets both sides. Ask for the case against your plan on purpose. The tool is at its most useful when you make it argue with you.

Ask one thing at a time

A prompt that bundles three questions gets a reply that answers one well, one badly and one not at all, and the shape of the reply hides which is which. Split them. The second question can build on the first once you have read the answer. Short exchanges, each with one job, are easier to check and easier to keep track of later.

Ask what it is not sure about

Ask the tool what it would need to know to answer better, and what in its reply you should verify before acting. It does not know its own error rate, so the list is not a measurement, but it often surfaces an assumption you did not realise you had left to it. Treat the list as a starting point for your own checking, not as a guarantee about the rest.

Questions to ask yourself

  • Did I give it the facts the answer turns on, or leave it to guess them?
  • Does my question tell the tool which answer I am hoping for?
  • Have I asked for the reasoning, or only for the conclusion?
  • Is this one question, or several wearing one coat?
  • Could I take this reply to a source and test it, sentence by sentence?

Common mistakes

  • Asking the general question when you have a specific situation, then taking the general answer as yours.
  • Writing the question so the answer is already in it.
  • Accepting a bare recommendation without asking what it rests on.
  • Asking it to confirm a decision you have already made, which produces agreement, not a check.
  • Treating a long, well-organised reply as evidence that the question was answered.

Which knowledge areas apply

  • How language models answer: why the tool fills gaps with the common case and leans toward your framing, which is what asking well works against.
  • Numbers and assumptions: what to ask for when a reply contains a figure or a projection, so the assumptions come out where you can see them.
  • Primary sources: what a usable answer to “where does that come from” looks like, so you recognise one and notice when the tool does not give it.

For practitioners

A client who is doing this step on their own can be taught it in two instructions: give the tool the facts, and ask it for the reasoning. Have them bring you the exact question they typed, not their summary of the answer; the question usually explains the answer. If they have been asking leading questions, show them the same question asked both ways and let the difference make the point. In their file, note the questions they are putting to a tool and any assumption the tool supplied that you had to correct, with the date, so you can see where their information is coming from.

Next step

Check what it tells you: how to test a figure, a rule or a recommendation against a source you can trust before you act on it.