How to use AI as a thinking partner

There is a particular kind of thought that just loops. A decision you can't settle, a conversation that went sideways, a worry that was small at nine in the morning and somehow has teeth by eleven at night. You turn it over, put it down, pick it up an hour later in exactly the same shape. For most of my life the only way out of that was to find a person and talk it through, and good luck with that late on a weeknight. So over the past year I have started using AI as a thinking partner. Not for answers. For getting my own thinking back in a shape I can actually look at.

I say that a bit sheepishly, because it sounds like something you'd roll your eyes at. But it has quietly become one of the more useful habits I've picked up, and the useful part is almost never the advice. It's the questions it asks first, and the way saying a thing out loud to something patient tends to shrink it.

What a thinking partner is actually for

The trap is treating AI like an oracle. You type the messy situation, it hands down a verdict, you either obey it or ignore it, and either way you've learned nothing about your own thinking. The whole exchange just becomes a magic eight ball with better grammar.

What I'm after is closer to a good friend who happens to be a careful listener. Someone who reads back what I just said, notices the bit I skated over, and asks the plain question I'd been walking around for days. When the thought is yours and you've been carrying it a while, you slowly lose the ability to see it from the outside. A thinking partner gives you that outside view, and most of the value shows up in the first few minutes, well before anything that looks like a recommendation.

How to set it up so it argues with you

Left to its own devices, a model will flatter you. Ask it a leading question and it will cheerfully agree that yes, you're onto something, and it will say that in the same confident voice it uses when it's right. So the whole game is setting it up to push instead. It takes about three sentences, and it changes everything. These are the prompts I lean on, more or less word for word:

• Say my own argument back to me in the clearest version you can, then tell me the weakest link in it.
• Make the strongest possible case for the opposite of what I want to do.
• What am I assuming here that I haven't actually said out loud?
• If this goes wrong in a month, what will the reason most likely have been?
• Don't just agree with me. Agreeing is the unhelpful answer here.

That last one matters more than it looks. You have to give it permission to disagree, otherwise it defaults to being nice, and nice is useless when you're trying to find the hole in your own plan. I usually open with the situation, my current leaning, and one line telling it that its job is to make me less certain, not more.

Three voices, not one

The thing that made this click for me was dropping the one flat assistant and setting up a few characters instead. Luna, the journaling app I'm building, leans on exactly this idea, so I've spent a lot of time with it. Three voices on the same problem get me most of the way to the conversation I'd have had with the people who know me, minus the scheduling. The ones I keep coming back to:

• The friend. Warm, a little blunt, more interested in how I'm actually doing than in the spreadsheet. Good for noticing when the real problem is that I'm tired and scared, not that the decision is hard.
• The coach. Practical and slightly impatient. Cuts past the feelings to the next concrete step, and asks what I'm going to do by Friday.
• The philosopher. Zooms out. Asks whether the thing I'm agonising over will matter in a year, and what I actually want underneath the surface question.

The friend and the coach usually end up disagreeing with each other, and watching that disagreement play out is often where I find the answer I already had, sitting in the gap between what I feel and what I know I should do.

The goal was never to be told what to do. It was to spend less time alone with a question before I could see it clearly.

Where it actually helps

Mostly it helps me catch what I'm really asking. I'll sit down certain I'm choosing between two options, and a few minutes of being asked what I'm actually worried about reveals that the decision underneath is a different one, usually one I'd been dodging. Last month it was a trip I kept telling myself I couldn't afford the time for, and the honest version turned out to be that I was nervous about going on my own. Different problem, and a much easier one to look at once it was named.

It's good for the late spiral too, the worry that grew teeth after dark. Saying it plainly to something that won't get bored or anxious alongside me tends to shrink it back to its real size. Half the time the honest answer is that I'm tired and the decision can wait until morning, which is exactly what a steady friend would have told me. And it's good for the small awkward stuff: how to word a message I've rewritten five times, whether I'm being unreasonable, what the kind reading of someone else's behaviour might be. Low stakes on their own, but they're the thoughts that quietly eat an evening.

Where it lets you down

It agrees too easily, which is the failure that actually costs you. A thinking partner you haven't told to push back will mostly hand your own opinion back to you with a nicer vocabulary, and you'll walk away feeling validated instead of clearer. It can also be flatly wrong while sounding completely sure of itself, and there's no wobble in its voice to flag which answers to trust, so I treat everything it says as a prompt for my own thinking rather than a conclusion.

And the obvious limit: it is not a therapist, and it isn't a substitute for the people in your life. For the genuinely heavy things, a loop you can't get out of on your own, it is the wrong tool, and I'd point a friend towards an actual human every time. What it's good at is the ordinary churn of everyday decisions and self-reflection, the stuff that doesn't warrant a phone call but still rattles around for too long.

Keep your thinking yours

There's a catch underneath all of this. The half-formed, slightly embarrassing version of your thinking is exactly the stuff you don't want sitting on a company's server, quietly becoming training data. And a thinking partner only works if you're honest with it. It's hard to be honest with something you suspect is keeping a copy, so you water yourself down, and watered-down input gets you watered-down reflection back.

That's the whole reason I built Luna the way I did. The voices work on what you say while everything stays on your device: private, on-device, no tracking and no ads, and you can export the lot to Notion or a CSV whenever you want. You own the thinking, which is the only arrangement that makes much sense for something this personal. And if you'd rather use a general chatbot, it's worth a minute checking what it keeps before you pour the real stuff in.

None of this has made me decide better. I still get things wrong. What's changed is that I no longer marinate in a question for a week on my own, waiting for a clear head that never quite shows up. I get the second angle the same evening, and the choice is still mine. The voices help me think. They don't get a vote.

If you try it, start by telling it not to agree with you. That one sentence is where the whole thing turns from flattery into something you can use.

Built for this

Luna Journal

A journal that fits into your day. You speak, and Luna turns it into entries you can read back and reflect on. Organised with AI, and stored on your phone rather than on my servers.

Benjamin

Founder & sole developer, Novaire Digital