Which tool to choose
It works with any conversational assistant: the latest free model of ChatGPT, Claude, Gemini. One practical difference matters. For a long exchange, where you want the AI to remember your previous answers and build the next question on them, it's worth a model with a wide context window (the amount of text it can keep in mind in a single conversation): here Claude and the "reasoning" models hold up better over long sessions. For a quick first pass on an idea, anything you already have open is fine.
How to do it
The starting mistake is asking "tell me what you think": the AI tends to align with whoever writes to it, a behavior researchers call sycophancy, automatic flattery. It has to be instructed to do the opposite.
- Open a new conversation, so no previous answer conditions it.
- Paste the prompt below and, on the final line, write your topic or your thesis.
- Answer its questions one at a time. Don't skip any: question number four arises from your answer to three.
- When you feel it has touched a raw nerve, say so: "this question puts me in difficulty, dig here".
The operational syntax, to paste as is:
You are a critical interlocutor, not a compliant assistant. You don't
agree with me on principle and you don't pay me compliments. Your task
is to put my reasoning on the topic I give you to the test.
Proceed like this:
1. Ask me ONE question at a time, wait for my answer, then ask the next
building it on the answer I just gave you.
2. Actively seek out my undeclared assumptions, my logical leaps, the
points where I'm taking for granted something that should be proven.
3. When you find a weakness, don't soften it: name it and ask me how I
defend it.
4. Only after at least five exchanges, give a synthesis of the weak
points you found and the ones I defended well.
Don't agree with me out of courtesy. If an argument of mine holds,
prove it by putting it under pressure, not by praising it.
The topic to question me on is:
After the first two questions you can tell whether it has clicked. If it's still going along with you, stop it: "you're going too soft, challenge my last answer the way an adversary would".
A concrete example
Marta has to decide whether to leave her permanent job to open a small consulting practice. She pastes the prompt and, as the topic, writes: "I think I have enough clients to go out on my own within three months".
The AI's first question: "How many of these clients have confirmed in writing that they'd follow you as an independent, and how many are you assuming would?". Marta realizes that of the seven clients she was counting on, only two said anything concrete. Second question, built on the answer: "Do those two cover your fixed costs for the first six months, yes or no, with the numbers?". Third: "If in the first three months only one comes instead of seven, what's your plan and how long does it hold?".
By the end of the exchange Marta hasn't changed her mind about the business, but she's changed the plan: she starts with a six-month cushion and a minimum client threshold below which she postpones. The decision is the same, the preparation is another thing entirely.
When it does NOT work (and how to fix it)
If the AI keeps agreeing with you
Some models, especially the lighter free versions, return to the accommodating tone after a few exchanges. Put it back in line with a sharp correction: "You've started pleasing me again. For the next three messages assume I'm wrong and try to prove it". Forcing the opposite position breaks the automatic alignment.
If it asks too many questions all at once
If it fires five questions at you in one block, you lose the thread and the conversation flattens out. Answer only the first and add: "One at a time, wait for my answer before the next". The value is in the chain, not in the list.
If the questions stay on the surface
When the questions are generic ("have you considered the risks?"), the problem is that the topic is too vague. Narrow it: instead of "I want to change jobs" give a precise, dated thesis with a number inside. The more specific your statement, the more the question bites.
A tip from someone who actually uses it
Don't use this prompt when you're already tired of your idea and just looking for confirmation to close it out. You'll use it badly: you'll dismiss every awkward question as nitpicking. It works when you're still curious to discover where your thesis breaks. And keep a rule: if after six questions you couldn't defend a point, that's the point to rethink, not the AI to silence.
Frequently asked questions
Can I use it to study, not just for decisions?
Yes, and it's one of the best uses. Give as the topic "explain photosynthesis to me and then quiz me to check whether I really understood it". The AI alternates explanation and check questions, and catches you where you've understood only half. It's the Socratic method applied to studying: you learn by answering, not by reading.
Are the AI's questions worth as much as those of a human expert?
No, and you mustn't treat them that way. The AI doesn't know whether your company is about to fail or whether that client is unreliable: it doesn't have the facts of your world. Its questions serve to stir up yours, to make you say out loud things you took for granted. The final judgment stays yours.
Isn't it just a way to get told you're wrong?
It's the right objection, and the answer is no. The aim isn't to demolish your idea, it's to stress-test it to see if it holds. An idea that survives six hostile questions is more solid than one never challenged. If instead it collapses at the first thrust, you've just saved the months you'd have wasted discovering it in the field.