I use AI as a thinking opponent. Here's what that looks like.
How to Get AI to Disagree With You and Improve Your Thinking
The default behaviour of every AI model is agreement. You share an idea, it finds merit in it. You propose a plan, it helps you execute it. You make an argument, it builds on it. This isn’t a bug, it’s what most people want most of the time.
But it makes AI nearly useless for the thing I need most: finding out where my thinking is wrong before I act on it.
I’ve built a set of prompts I run specifically when I want the model to push back. Not to validate, not to help me execute, not to make my idea sound better. To tell me what’s wrong with it. What I’ve missed. What I’m assuming that might not be true.
These 7 prompts are what that looks like in practice.
The problem with asking AI “what do you think?”
When you ask AI for feedback on an idea, you mostly get a list of strengths followed by a polite section on “potential considerations.” The considerations are usually the least threatening version of the real objections.
That’s not feedback. That’s a compliment with a footnote.
Getting genuine pushback requires telling the model explicitly that agreement is not useful. Every prompt in this chain does that in a different way, for different kinds of thinking problems.
Prompt 1: The steel-manned attack
What it does: Asks the model to build the strongest possible case against your idea, not a weak version of the objection but the version that would most damage the idea if it turned out to be true.
When to use it: When you have an idea or plan you feel confident about. Confidence is exactly when you most need this. The more certain you feel, the more useful it is.
The prompt:
I’m going to describe an idea I believe in. Your job is not to evaluate it fairly. Your job is to build the strongest possible case against it. Find the version of the counterargument that would most damage this idea if it turned out to be true. Don’t hedge. Don’t balance it with positives. Just attack it. Here’s the idea: [DESCRIBE YOUR IDEA OR PLAN IN AS MUCH DETAIL AS YOU CAN]
How to use it:
Describe your idea in full, including why you think it’s right and what you’re planning to do with it
Read the attack without defending yourself on the first pass. Just take it in
Note which objection you most want to dismiss immediately. That’s usually the one worth sitting with
Example input: “I’m planning to go all-in on long-form Substack content and cut social media entirely. My reasoning is that social is a distraction, the algorithm controls your reach, and long-form builds a real audience that you own.”
What you’ll get: A full attack on the idea. For this example: the model might argue that long-form content without social distribution is hard to grow from scratch, that “owning your audience” only matters once you have one, and that cutting social entirely removes the discovery mechanism most Substack writers depend on. None of that means the idea is wrong. But it means you’ve stress-tested it before committing.
Advanced note: If the attack feels weak or obvious, tell the model: “That’s the surface objection. What’s the deeper structural problem with this idea?” The first attack is often the most predictable counterargument. The second one tends to be more useful.
Those two prompts cover the most common failure modes. The next five go after more specific problems.
Prompt 3 — The commitment audit: Makes the case for reversing a decision you’ve already made, ignoring sunk costs entirely, as if you were evaluating it fresh today with everything you now know.
Prompt 4 — The rationalisation detector: Looks at an argument you’re making and tells you whether it reads like reasoning toward a conclusion or reasoning from one, with the specific passage where the logic jumps flagged.
Prompt 5 — The “is this actually working?” audit: Takes the evidence you’re using to conclude something is working and asks whether the evidence actually supports that, or whether you’re reading ambiguous signals as confirmation.
Prompt 6 — The pre-publish argument check: Finds the weakest point in the central argument of a piece you’re about to publish, specifically the place a sharp critic would focus, not the most obvious objection.
Prompt 7 — The smartest critic in the room: Responds to your idea as the most credible, well-informed person in your hardest-to-satisfy audience would, quoting the specific parts they’d take issue with.
Plus: the challenge kit, a set of pre-built pushback prompts you can drop into any conversation without setup.