Most people reach for AI when they need something produced — a draft, a summary, a piece of code. There’s a different use case almost nobody talks about: decisions where you already know the answer is going to be uncomfortable, and your brain keeps finding reasons not to land on it.
Not the trivial stuff. Not “which laptop should I buy.” The harder kind — whether to leave a job, whether to kill a project you’ve sunk months into, whether a business idea is worth pursuing or just feels good to think about. Talking to a friend is awkward because they have opinions about you. Talking to yourself just loops.
AI is emotionally neutral in a way that’s actually useful here. It doesn’t have a stake in your outcome. It won’t be weird about it next week.
The Problem With How Most People Try This
The instinct when you first try using AI for a hard decision is to describe the situation and ask “what should I do?” That gets you a hedge-everything response that technically says nothing. The model lists pros, lists cons, concludes that “ultimately it depends on your personal values and circumstances.” You already knew that.
“What should I do?” is the most underspecified question you can ask. The model has no idea what you actually weigh heavily, what your constraints are, what you’ve already ruled out and why. So it defaults to the safest answer: present the decision as perfectly balanced. Most decisions aren’t perfectly balanced — they just feel that way because the costs of being wrong are asymmetric and you don’t want to face them.
A Better Way to Structure It
The prompt pattern that actually works: do the analysis yourself first, then use the model to stress-test it. Not to generate the analysis for you.
Before you open the chat, write two things — the decision stated as specifically as possible, and your current lean, the direction you’re probably going if nothing changes your mind. One sentence each. In your first message, give the model both and ask it specifically to argue against your lean. Not to present both sides. To make the strongest honest case for the option you’re currently discounting.
This does something different than asking for pros and cons. It forces the model to commit to a direction, and because you’ve explicitly asked for the counterargument, you’re more likely to read it charitably instead of dismissing it as generic hedging.
Then follow up by asking it to identify the assumptions your current thinking relies on, and flag which ones are untestable before you’d have to decide. That second prompt is where the real work happens. Plenty of decisions that feel hard are sitting on one or two assumptions that are either obviously solid or obviously shaky — and you just haven’t named them out loud yet.
What This Actually Looks Like
Say you’re deciding whether to drop a side project you’ve been working on for six months. No revenue. Two active users. Two other things competing for the same hours.
A vague prompt: “Should I keep working on my side project or drop it?” Useless, for the reasons above.
A structured prompt: “I’m deciding whether to shut down a side project I’ve spent six months on. Zero revenue, two active users. My current lean is to keep going for another three months and see if anything changes. Make the strongest honest case for shutting it down now.”
That gets you something you can push back on. The model might point out that “see if anything changes” isn’t a success condition — it’s a way of deferring the decision while opportunity cost keeps accumulating. You might disagree. But now you’re disagreeing with a specific argument, which is more productive than sitting with a vague sense of unease.
The follow-up that usually matters: “What assumptions does my ‘keep going’ position rely on, and which of those could I actually test in the next two weeks?” That’s the one that tends to produce something actionable.
What the Model Is Actually Good At Here
A few things AI does reliably well in this context that humans — even well-meaning ones — often don’t:
- It will say the uncomfortable thing directly. A friend who knows you will probably soften it. The model won’t, if you’ve asked it not to.
- It can hold a lot of context simultaneously and reflect it back without the cognitive overhead of someone trying to remember everything you’ve said across different conversations.
- It’s good at spotting when two things you’ve stated pull against each other. “You said you prioritize flexibility, but you’ve also said you need predictable income in six months — those don’t fit together here.”
What it’s not good at: knowing things about your situation you haven’t told it, reading your actual risk tolerance from stated preferences alone, and noticing when you’ve framed the decision in a way that loads the question. That last one is your job. The model works within whatever frame you give it, so if you describe the situation with a subtle bias baked in, the output reflects that bias back. Same as any analysis tool.
One Genuine Caveat
This works for decisions where the core problem is clarity, not missing information. If you genuinely don’t know a key fact — what something will cost, how a person will respond, whether a market exists — the model can help you figure out how to find out, but it can’t replace finding out. Using structured AI prompting to feel more confident about a decision that still has a real empirical unknown in the middle of it is a trap worth naming.
Also: this isn’t therapy. If the decision is tangled up with something emotionally significant, AI can help with the logical structure, but that’s not the same as processing how you actually feel about it. Confusing the two is easy and not particularly useful.
Where to Start
Pick one decision you’ve been sitting on for more than two weeks. Write your current lean in one sentence. Ask Claude or ChatGPT to make the strongest case against it — not a general overview of trade-offs, specifically against your lean. See whether the counterargument shifts anything, or whether pushing back on it makes you more confident in your original direction.
Worth sitting with: if your reaction to the counterargument is mostly irritation rather than engagement, that’s usually informative too.
