How do you use AI to think through a problem instead of just getting an answer?

When you type a hard question into ChatGPT, it usually answers at once, and the answer often sounds reasonable. That is useful for facts, but not for decisions that depend on your own situation, values and numbers. A ready-made recommendation skips the part where you discover what you actually think, and it rests on guesses about details you never gave it.

The alternative is to ask the AI to coach your thinking. This is often called the Socratic method, after the Greek philosopher Socrates, who taught by asking questions rather than giving lectures. The AI asks, you answer, and the questions push you to make your reasoning explicit. You end up with a decision you understand and can defend, and the AI ends up with far better context if you later ask for its opinion.

Why the default answer falls short

A Beginner's Guide to ChatGPT stresses that the quality of an answer depends on the context you provide: your role, your situation and what you expect back. With a personal dilemma, most of that context lives in your head. A model that answers immediately fills the gaps with averages. A model that asks questions draws the context out of you first.

Educators reached the same conclusion. In its 2023 guide Teaching with AI, OpenAI published a tutor prompt written by Ethan and Lilach Mollick of Wharton that tells the model to ask one question at a time, find out what the learner already knows, and guide with leading questions and hints. One line in it sums up the idea: "Do not provide immediate answers or solutions to problems." The same approach works for adults working through their own problems.

The rules that make it work

A thinking-partner prompt needs four kinds of instruction:

You can also tell it when to stop questioning, for example after eight to ten questions, so the conversation does not go on forever.

A worked example: a budgeting dilemma

Imagine a household with some savings that is torn between paying off a car loan early or putting the money into an emergency fund. Here is a prompt that sets up the conversation. The numbers are illustrative.

I want to think through a money decision myself. Please act as a patient thinking partner, not an advisor.

My situation: I have $6,000 in savings. I owe $5,500 on a car loan at 7% interest, with 30 months left. I have no other emergency fund. My income is steady but my car is old.

The question: should I pay off the car loan now, keep the money as an emergency fund, or do something in between?

Rules:
- Ask me one question at a time and wait for my answer.
- Do not tell me what to do, and do not rank the options.
- After every two or three answers, sum up in two sentences what you've heard, including anything that seems to contradict something I said earlier.
- If I ask for the answer, remind me of these rules and ask your next question.
- After about eight questions, ask me to write my own decision and my reasons in three sentences. Then point out any assumption I haven't checked.

A good first reply asks something like "What would happen this month if your car needed a $1,500 repair?" Later questions might cover how secure your job is, whether paying off the loan frees up a monthly payment you could then save, and how you would feel with an empty savings account. None of these questions decides anything for you; each makes one consideration visible.

Where to be careful

Two limits apply. First, the AI may still slip into advice after a few turns, especially if you sound unsure. Repeating "Remember: questions only" brings it back. Second, if it states facts along the way, such as typical interest rates or tax rules, treat them as claims to check. The book's warning applies here as everywhere: language models can produce errors in a confident tone, so verify numbers and rules with a reliable source before you act. For serious financial, legal or medical choices, the conversation helps you prepare good questions for a qualified professional; it does not replace one.

How do you ask it to question you back?

Say so explicitly, and make questioning its main job rather than an extra. A line such as "Before you answer anything, ask me the questions you need, one at a time" changes the shape of the conversation. Without it, most chatbots answer first and ask, at best, one polite follow-up at the end.

Give it a reason to ask. "Ask me about anything you would need to know to give good advice" works better than "ask me questions," because it tells the model what the questions are for. The book's advice to ask the model to weigh pros and cons or reason step by step fits here too: you can ask it to question you about each pro and each con in turn.

Steer the type of question when the first ones are too easy. Ask for questions that test your assumptions ("What would have to be true for this to be a bad idea?"), questions about consequences ("What happens in six months if you choose this?") or questions from another person's view ("What would your partner worry about?").

Finally, decide in advance how the session ends. A fixed number of questions, followed by your written conclusion and the AI's list of unchecked assumptions, keeps the exercise focused and leaves you with something concrete to act on.