Becoming an AI Power User

You face a decision you have not resolved, so you open AI and ask it straight: should I do this, yes or no? Back comes a confident answer with three tidy reasons, and you notice that none of it is news. The model decided for you, off a single sentence, and you learned nothing you did not already half-believe. That is what asking for a finished answer gets you on a problem that was never finished in your own head. There is a second way of working, one many people never switch into: instead of asking once and leaving with an answer, you stay in the problem and think it through with the model.

Thinking with AI opens ground that asking AI for answers never reaches

There are two ways to sit down with AI, and most people only ever use the first. The first is transactional: you ask, it answers, you leave. That mode is fine, and often best, for a task with a clear right answer. Convert this file. Summarize this page. Draft this reply. You know what done looks like, the model produces it, and you move on.

The second mode is a conversation you stay inside, and it is the one that pays off on hard problems. These are the problems where you do not yet know what you want, where two good options pull against each other, where you cannot even name everything you are weighing. For those, a one-shot answer skips the part that mattered. The model can generate options you would not have reached, name assumptions you did not notice you were making, and hold more moving parts in view at once than you can track in your head. You stay in the problem, reason alongside it, and let the question get sharper turn by turn.

Let each side do what it does better, and aim at the best-aligned call. The model is strong at breadth: it can lay out options, surface assumptions, hold many variables, and argue the other side of your position. You are close to the situation and you own the consequences: you weigh what matters and decide what risk is acceptable. You can also ask the model to check the criteria themselves, since the framing you walk in with is often the part most worth questioning. Both sides get sharper in the exchange. The model gets better context and standards from you, and your reasoning gets tested against options you would not have produced alone. You stay accountable for the choice, and you are after the call that best fits the situation, drawn from whichever judgment is stronger on each part of the problem.

The friction is the feature, not a sign to quit. When the model misreads your question or proposes an option that is off target, that is not a dead end. It is a prompt to say what you really meant and steer the next turn, and the act of restating it usually clarifies the problem for you too. If you abandon the conversation the first time it drifts, you miss the thinking that happens when you stay and repair the direction. The chapter on diagnosing frustrating AI failures treats that repair as a skill of its own.

Why dialogue beats a finished answer comes down to how thinking works

You do not have to know the mechanism to use thinking partnership. The dialogue moves stand on their own, and the rest of this chapter walks them step by step. But two findings from how human thinking works explain why the back-and-forth produces a sharper decision than a one-shot answer does, and they tell you which moves to lean on.

A fluent answer triggers fast acceptance, and overriding it takes deliberate effort you tend to skip. Cognitive scientists describe two kinds of processing. Jonathan Evans and Keith Stanovich formalized them as Type 1 and Type 2 in their 2013 framework, refining Daniel Kahneman's well-known System 1 and System 2 model. Type 1 is fast and intuitive: you see a familiar pattern and reach a conclusion at once. Type 2 is slow and deliberate: you calculate, weigh evidence, and compare options. Most of the time Type 1 runs the show, which is efficient and usually right and also biased in known ways. It overweights vivid examples, anchors on the first number it meets, and prefers explanations that feel coherent over ones that are true. A confident LLM answer lands as Type 1 acceptance, because it looks right, and pushing past it to check takes Type 2 work that drains over a long session. This is why a one-shot answer is risky on a hard call: it invites the lazy yes. A conversation that asks the model to challenge your assumptions and argue the other side can prompt the slower, more deliberate Type 2 checking, dragging into view the considerations a quick yes would have skipped.

Explaining your reasoning out loud sharpens it, even before the model replies. The cognitive psychologist Michelene Chi and colleagues showed in 1994 that prompting people to generate their own explanations improves how well they understand material, beyond what they get from receiving explanations ready-made. When you tell the model why you think an approach should work, what your constraints are, and where you are unsure, you are forced to make your own half-formed thinking explicit, and that act of putting it into words is where part of the clarity comes from. Some of the value of the conversation comes from the model's response. Some comes from what explaining forces you to pin down for yourself. The same effect powers rubber-duck debugging, where a programmer untangles a problem by explaining it to a rubber duck on the desk: the listener does not need to be brilliant, the speaker needs to get articulate.

A five-move dialogue carries you from a vague problem to a clear decision

When power users face an ambiguous problem, they tend to run the same five moves in order. Read them as one pattern with a shape, not five items to memorize: each move sets up the next, opening the problem wide before narrowing it back down to a call you own. The first three moves open the space (describe, surface, generate), and the last two close it (test, decide). The list below names each move and shows the kind of thing you say to make it.

1. Describe the situation without prescribing the answer. Instead of 'write me a plan for X,' say 'here is my situation, here are my constraints, here is what I am trying to achieve. What should I consider first?' You are handing the model the terrain, not the destination.
2. Ask the model to surface your assumptions. 'What am I assuming that might be wrong?' or 'What factors am I not considering?' This one move tends to pay off out of proportion to its length, because the model holds more of the context at once and can spot gaps you have stopped seeing because they feel normal to you.
3. Ask for options, not a recommendation. 'Give me four different approaches with the trade-offs of each' produces more useful thinking than 'what should I do?' The model is strong at breadth; you are close to the situation and accountable for the call. Lean on each where it is stronger.
4. Pressure-test the option you lean toward. 'What could go wrong with option two?' or 'Under what conditions would this fail?' Here the model plays devil's advocate, a role most colleagues are reluctant to take with you, and it does it without flinching.
5. Decide, then evolve the system you used. Make the call, aiming at the result that best fits your situation. Then climb a level: take the dialogue moves that worked and encode them into a reusable thinking-partner prompt, and over time into a standard you keep refining as you learn which questions surface the most. Each ambiguous decision then starts higher up, from a structure that has already absorbed your last few, instead of a blank page. You can even ask the model to read back the transcript and propose how to sharpen the prompt for next time.

Watch the same decision run two ways to feel the difference

Here is one decision worked both ways, so you can see the gap between asking for a finished answer and thinking the problem through. The decision: a consultant is weighing whether to raise an hourly rate. The numbers are a 2026 stand-in, and the point holds whatever the figures or the interface; the same dialogue works typed into a model like Claude Fable today, spoken to a voice assistant or earbuds, glanced through smart glasses, and on through brain-computer interfaces and whatever superintelligent systems come after. The surface keeps changing; staying in the problem and pushing the system to think with you does not.

Bring one real decision and work it in dialogue this week