There’s a version of using AI that’s basically a faster search engine. You have a question, you get an answer, you move on. It works. But there’s a different mode — harder to get into, more rewarding when it clicks — where the AI isn’t giving you the answer so much as helping you figure out what you actually think.
These are not the same use case, and they need different prompts to produce different results.
The difference in practice
Here’s a concrete example. Say you’re deciding whether to take on a freelance client who’s offered more money than you usually charge but comes with some red flags. You could ask AI: “Should I take this client?” You’ll get a balanced list of pros and cons, probably five bullet points on each side, and you’ll feel exactly as uncertain after as before.
Or you could do this instead: dump everything you know about the situation into the chat — the offer, the red flags, the specific thing that felt off in the first call, the fact that you’re three weeks from a slow month — and then ask the AI to tell you what it notices about your description. Not what you should do. What it notices.
That second version is doing something different. You’re not asking it to decide for you. You’re using it as a mirror that’s good at spotting patterns in language. The AI will often point out that you used hedging language about the money but direct language about the red flags, or that you mentioned the slow month almost parenthetically even though it seems load-bearing. Those observations aren’t conclusions — they’re material for you to actually think with.
Why asking for answers is often the wrong move
AI systems are trained to be helpful in a specific way: produce a complete-sounding response to whatever question is asked. That’s great for factual questions. It’s genuinely not great for messy, values-laden, context-dependent decisions where the quality of your thinking matters more than the quality of the information you receive.
When you ask “what should I do,” the model produces an answer that sounds authoritative. Your brain then evaluates the answer rather than the problem. You’re one step removed from your own reasoning, and you often don’t notice that’s happened.
Ask “what am I not saying here” or “what assumption is this argument resting on” and you get a different kind of output — something to push back against rather than accept or reject wholesale.
Prompts that actually shift the mode
A few patterns that consistently work better for thinking than for answering:
- “Here’s my current thinking. What’s the weakest part of it?” — forces the AI to engage with your actual reasoning rather than generate a generic overview of the topic.
- “I’m going to explain this decision out loud. Tell me what assumptions I’m making that I haven’t stated explicitly.” — works especially well for things you’ve been turning over in your head so long that you’ve stopped noticing the buried premises.
- “What would someone who disagrees with this conclusion say — not the obvious pushback, the sophisticated one?” — the “not the obvious pushback” instruction does real work here. Without it, you get the straw man. With it, you often get something genuinely useful.
- “Am I solving the right problem?” followed by a full description of the situation — deceptively simple. Reframing a problem is often harder than solving the one you’ve already framed, and an outside reader (even a non-human one) can spot a misframed question faster than you can.
What actually makes this work
The prerequisite is giving the AI enough raw material to work with. This is the part people skip. Two sentences describing your situation produces generic output, because there’s nothing specific to reflect back at you. The more messily specific you are — the actual numbers, the actual thing that felt off, the actual reason you’re uncertain — the more useful the mirror is.
This is one place where talking out loud (if you’re using a voice interface) tends to beat typing. When you type, you self-edit. You produce a polished, vague summary of your thinking. When you speak, you ramble a bit, catch yourself, contradict yourself — and those moments are exactly what make a thinking conversation productive. The gaps in your own reasoning show up in the ramble.
Having built on top of Claude long enough to have a real opinion: it handles this particular use case better than most of the alternatives I’ve tried. Not because of anything mysterious, but because it pushes back more readily when you ask it to, and because it holds long, messy context — the kind of real situation description that actually enables good thinking — without drifting or losing the thread. For simple Q&A the gap is smaller. For an extended reasoning session, it’s noticeable.
The actual limitation worth knowing
This only works as well as your own willingness to engage with what comes back. It’s entirely possible to use AI as a thinking partner and still just extract the bits that confirm what you already believed, then ignore the rest. The AI can point out a weak assumption; it can’t make you take the observation seriously. That part’s still yours.
There’s also a ceiling on what kind of decisions this is useful for. Values questions — what kind of work do I want, what trade-offs am I actually willing to live with — don’t become clearer just because you’ve had a good thinking session. At some point you have to sit with real uncertainty rather than process it further. More thinking is sometimes the productive move; sometimes it’s the avoidance move dressed up as productivity.
Next time you open a chat with the intention of getting help on a real decision, try spending the first message describing the situation in full, without asking any question at all. See what the model says. It almost always notices something you weren’t planning to address first.
