Thinking Fast, Asking Better

You open an AI tool with a real problem on your mind. A proposal is drifting. A client has asked for something vague. Your team cannot agree on which part of a new service matters. You type the first reasonable version of the question, and within seconds the model gives you a polished answer.

The answer is orderly. It has headings. It names a framework, lays out three options, and closes with a sensible recommendation. Nothing in it looks obviously wrong. That is usually the moment when the trouble starts.

The speed of the reply creates a feeling of resolution before the problem has been properly described. The model has filled the gaps with plausible assumptions, and the person reading it is relieved to have something concrete. Work begins around an answer that arrived too early.

Daniel Kahneman's Thinking, Fast and Slow gives us a useful lens for understanding that moment. The book is about human judgment, not artificial intelligence, and the North Crow methodology was not built as an adaptation of Kahneman's work. Still, the connection is strong. Kahneman explains why fast, coherent answers carry so much authority. Asking Better Questions gives us a practical way to interrupt that authority when we work with AI.

The mind likes a finished story

Kahneman describes two modes of thought. System 1 is fast, automatic, associative, and always working. System 2 is slower, more deliberate, and effortful. We need both. Fast judgment lets an experienced designer notice that a layout is wrong before they can explain why. It lets an operator catch an odd number in a report and a strategist hear the weak point in a pitch.

The problem begins when a situation needs deliberate thought and receives an automatic answer instead.

AI makes this easier to miss because it can turn a thin question into a finished-looking response almost instantly. The model supplies language, structure, and confidence. Our fast-thinking mind sees coherence and mistakes it for completeness. We stop looking for what was left out.

This is the hidden thinking failure: we treat the arrival of a plausible answer as evidence that we asked the right question.

Kahneman used the phrase "what you see is all there is" to describe our tendency to build conclusions from the information immediately available. An AI conversation can intensify that tendency. The model sees only what you provide, yet it rarely leaves the missing spaces blank. It predicts across them. Then you see a smooth response rather than the absent client history, political constraint, failed attempt, budget limit, or private doubt that would have changed the recommendation.

The machine produces the story. The human supplies the confidence.

Put friction before production

The better question is not simply, "How can I improve this prompt?"

Ask instead: What information would change the answer, and what have I allowed the model to assume?

That question moves the work into slower territory. It asks you to inspect the frame before accepting what the frame produces. In the Asking Better Questions field guide, that discipline appears in four parts: context, constraint, voice, and intent.

• Context: What is happening, who is involved, and what came before?
• Constraint: What limits, rules, risks, and non-negotiables shape the work?
• Voice: Whose judgment and language should remain visible?
• Intent: What is this for, and what would a useful outcome change?

These are practical prompts for deliberate thought. They make the user retrieve information that the first question skipped. They also expose whether the user knows what they are asking for or has merely named the output they would like to receive.

The method does not require us to distrust intuition. Experienced intuition is often valuable. It requires us to notice when intuition is operating outside the conditions that made it reliable. If you have solved the same production problem a hundred times and the environment is stable, your first read may be excellent. If you are making a pricing decision in an unfamiliar market, interpreting a vague client request, or betting a quarter's budget on a new offer, the first coherent story deserves inspection.

That inspection becomes more important when the decision depends on facts scattered across a contract, an old client conversation, a private concern the team has avoided naming, three failed attempts that never made it into the brief, and the practical knowledge of the person who will eventually have to make the recommendation work.

One request, two kinds of thinking

Imagine an independent consultant deciding whether to turn a custom service into a fixed package.

The fast version of the AI conversation begins like this:

I want to productize my consulting service. Help me create three packages and recommend prices.

The model will probably comply. It may produce a starter, growth, and premium tier. It may suggest attractive round numbers, list features under each tier, and explain why the middle option should be positioned as the best value.

The answer can be competent while resting on invented conditions. The model does not know why clients hire the consultant, where delivery becomes difficult, which work produces the strongest result, what buyers already understand, or whether clients even want standardized choices. Tiered pricing may be the wrong decision presented in a very familiar shape.

The slower version starts elsewhere:

I am considering turning part of my consulting work into a fixed package. Do not design the offer yet. Interview me one question at a time about why clients hire me, which parts of the work repeat, where outcomes vary, what clients resist, what delivery requires from me, and what evidence I have that buyers want a standardized service. When we finish, list the assumptions behind productizing the work and identify the one most likely to make the idea fail.

The second request delays the satisfying part. There are no package names to admire after thirty seconds. Instead, the conversation surfaces the facts and judgments that should determine whether packages belong in the answer at all.

It also counters another problem Kahneman studied: question substitution. When a hard question demands effort, we often answer an easier one without noticing the switch. "Should this service become a product?" is difficult. "What should the three tiers be called?" is easy. AI is very good at helping us complete the substitution because it can answer the easier question beautifully.

Asking better questions keeps the original difficulty in the room.

Confidence needs an opponent

There is another connection between Kahneman's work and the North Crow method. People are prone to confirmation, anchoring, and overconfidence. AI systems are also inclined to follow the user's framing and produce agreeable continuations. Put the two together and a weak idea can acquire a great deal of polish without encountering much resistance.

This is why the method includes deliberate pushback:

What did you assume about the audience, goal, and constraints when you answered?

Find the weakest point in my reasoning before you recommend a course of action.

What evidence would make the opposite conclusion more likely?

If my preferred answer is wrong, where will the mistake show up first?

These questions do not guarantee truth. They create a second pass, one that is less interested in maintaining the original story. The user still has to judge the result, verify claims, and bring domain experience to the decision. The benefit is structural: the conversation now has a place for disconfirming information to enter.

Kahneman was candid about the limits of awareness. Knowing that bias exists does not make a person immune to it. A working practice helps because it moves the correction out of memory and into the process. You do not have to remember every bias by name. You need a repeatable moment where assumptions are exposed, alternatives are considered, and the first answer loses its privileged position.

Thinking, Fast and Slow explains why premature answers feel complete; Asking Better Questions gives you a way to keep thinking after they arrive.

Try the slow pass today

Choose one AI conversation tied to a decision that matters. Before asking for an output, write down the first question you were going to use. Then add a slow pass:

1. Name what the model cannot know unless you tell it.
2. Ask the model to interview you before producing anything.
3. Require it to list the assumptions behind its eventual answer.
4. Ask for the strongest case against the direction you currently prefer.
5. Treat the response as material for judgment, not a verdict.

The extra time may feel inefficient because the tool can generate something immediately. That feeling is part of the problem. Speed is useful after the thinking is sound. Before that, it simply lets a weak question travel farther.