Use Claude to get you questions, not answers

The standard objection to AI is that it does your thinking for you. People say it about students. They say it about analysts. They say it about anyone who opens Claude before they open a blank page.

The objection is right about how most people use the tool. It's wrong about the tool itself.

If you use Claude the way most people use it, you are giving up your thinking. You ask a question. You get an answer. You read the answer. You either accept it or you don't. The thinking that produced the answer happened inside the model, not inside you, and the only cognitive work you did was deciding whether to keep what came out. That's not thinking. That's grading.

But that's not the only way to use it. Done right, a Claude session makes you think more, not less. It makes you do harder work, not easier work. The reason is that done right means you stop asking the model for answers and start asking it for questions.

The two-hour interview

The other day I sat down to start working on a new project idea. I told Claude to ask me questions, not give me summaries. I kept feeding it information. Every time it looked like it was about to wrap up what I'd said into a tidy reflection, I told it to wait. Just keep taking the input. Don't conclude. Don't tell me what you think I mean. When I finally looked up, almost two hours had passed. Just me sharing information and Claude asking clarifying questions, without summarizing. When we got toward the end and I finally asked for a vision.md articulation of the project, the document was not just good. It was able to anticipate risks and surface conflicts that I would have missed, because we had done two hours of real work to get there instead of fifteen minutes of generic scoping.

Here is what that session looked like as a posture, broken down into its parts.

Open with a directive, not a prompt

I tell Claude I'm not looking for an answer. I'm looking for questions. I want it to challenge me, push in on the places where my logic is soft, and refuse to validate anything that doesn't hold up under pressure. That's the first move and it's not optional. Without it, the model defaults to helpful mode and the session is over before it started.

Then the move that matters most: tell it to wait

Models want to summarize. It's how they demonstrate understanding. You give them three sentences and they reflect back a paragraph, often more elegantly than you said it, and they ask if they've captured your thinking. This feels productive. It is the opposite of productive, because every summary pulls the conversation toward the center of what the model already knows about your topic.

Claude is trained on enormous amounts of data. Enormous amounts of data have a center, and the center is average. The model's first impulse with any input is to map it onto the most common version of that input it has seen before. Premature summaries are the mechanism by which the model drags your specific, peculiar, possibly interesting idea back to the average version of that idea. Once it summarizes, the rest of the conversation is conducted with the average sitting in the middle of the table.

So I tell it: do not summarize. Do not conclude. Do not tell me what you think I mean. I'm going to keep giving you information. When I want a summary, I'll ask. Until then, take it in. Add it to the stack. Don't reduce.

You'll know it's working when the questions stop sounding generic

What you notice when you do this is that the model starts adjusting. After enough input that doesn't fit the average pattern, it stops trying to fit you to the pattern. You can see it happen in the questions. They get more specific. They get less generic. The model recognizes it's somewhere it hasn't been before, and the questions reflect that. This is different than what I was expecting is the signal you're listening for, and you'll know it when you see it, because the model will more or less say so.

That's when the work starts.

Now the questions get good. Some of them, I'm ready for. I've thought about who the user is, so when it asks who specifically would pay, I have an answer. Those questions are fine. They confirm I've done my homework on the pieces I knew were important.

The interesting questions are the ones I'm not ready for. The model asks something that hits a place in my thinking I haven't actually worked through, and I have to stop and think. Sometimes I have to stop the session and go think, because the answer isn't going to come in real time. I come back later with a take, and we keep going. This is the part the standard objection misses. Done right, Claude doesn't replace thinking. It assigns thinking. It hands you specific cognitive tasks that you would not have given yourself, because you didn't know they were the tasks.

The move that earns the whole session

If you give the model enough input over enough time, and you keep it off the average by refusing to let it summarize, eventually it notices something you couldn't notice yourself. It says: earlier, you told me X. Just now, you said something that points toward Y. Those are in tension. Which one do you actually believe.

That moment is what the entire session was for. It's a thing you cannot do for yourself, because you are inside both takes and you can't see them from outside. You can only see your own thinking through your own thinking, which means contradictions hide in plain sight. The model is outside both. It can hold X and Y at the same time and notice they don't fit. When it surfaces that, you have to choose. The choice is real thinking. Hard thinking. The kind that doesn't happen when you're alone with a blank page or alone with a search box.

Why the framing matters

You're giving up your thinking is true if you use Claude as an answer machine. It's the inverse of true if you use it as a question machine that you've explicitly instructed not to take shortcuts. The same tool, used in opposite postures, produces opposite cognitive effects.

The posture has parts. Tell it you want questions. Tell it to push back. Tell it to wait. Don't let it summarize until you ask. Keep feeding it until the model itself recognizes it's off the average. Then let it ask the questions you're not ready for, and go do the work those questions require. Bring back your take. Repeat. Eventually, the model will find the place where two of your takes don't agree, and the session will have done something for you that you couldn't have done by yourself.

That's the opposite of giving up your thinking. It's the most thinking you've done all week.