Most people use AI to build a case. Write the argument, sharpen the pitch, make the idea sound better. That’s fine, but it’s the less interesting direction. The more useful thing — and the one almost nobody does deliberately — is asking the model to dismantle what you just wrote.
Not “what are some weaknesses” in a vague sense. That gets you a polite list of hedged non-criticisms. Actually assigning the model a role: you’re the skeptic, you think this plan is flawed, tell me exactly where the logic breaks down and why.
Why the Default Prompt Fails
Ask AI “what are the weaknesses of my idea” and you’ll usually get something structured and diplomatic. Three bullet points, roughly equal weight, phrased carefully enough that none of them actually sting. It’s technically criticism but it doesn’t put any pressure on you.
The problem is the framing. You asked for weaknesses, so the model scanned for things that could plausibly be called weaknesses and listed them. It wasn’t trying to argue you out of anything. It was satisfying the prompt.
Shifting to adversarial framing — “you think this is a bad idea, convince me” — changes what the model reaches for. It’s not looking for weaknesses anymore; it’s constructing a coherent counterargument. That’s a different task, and the output is noticeably sharper. The model has to commit to a position instead of cataloguing possibilities from a safe distance.
What This Actually Looks Like in Practice
Say you’ve drafted a pricing page for a freelance service. The obvious use is asking AI to improve the copy. More useful: paste it in and say “you’re a potential client who’s skeptical about the price. Walk me through what makes you hesitate or click away.”
Or you’re deciding whether to take on a project with an awkward scope. Standard prompt: “is this a good idea?” Better prompt: “argue that I should turn this down. What’s the strongest case for not doing it?”
The output won’t be perfect — the model doesn’t know your full context, and some objections will miss. But that’s fine, because the good objections feel different from the bad ones when you read them. The ones that land are usually things you already knew somewhere, half-formed, and hadn’t quite articulated. The model just gave them a clean sentence.
The Version That Goes Further: Pre-Mortem Prompting
There’s a specific variant worth naming separately. Pre-mortem analysis — a technique from decision research — starts from the assumption that your plan already failed, then asks why. Applied to prompting, it looks like this:
“Assume I launched this and it didn’t work out. Six months from now I’m looking back at what went wrong. What are the three most likely reasons?”
That small shift — past tense, failure already happened — tends to produce more concrete answers than forward-looking “what could go wrong” framing. The tense matters. When failure is hypothetical, the model hedges. When failure is stipulated, it starts explaining it instead.
I’ve used this when scoping out new automations. Before committing to building something, I’ll paste in the rough plan and run the pre-mortem prompt. The answers aren’t always right, but they’ve surfaced real failure modes — edge cases I hadn’t thought through, dependencies I’d glossed over — more than once.
Where It Breaks Down
The model is only working with what you give it. Paste in a vague paragraph and the adversarial prompt produces vague objections. The technique rewards having actually written something out properly first — which is useful discipline in itself, but means you can’t shortcut the thinking and expect the model to do the hard part for you.
It also doesn’t know your specific constraints. A model arguing against a pricing decision doesn’t know your runway, your existing clients, or your risk tolerance. It’s arguing against the idea in abstract, not against your actual situation. Sorting which objections apply to your context is still your job.
And occasionally, if you’ve given it a strong document, it produces a weak counterargument — not because the idea is airtight but because the model drifts toward agreeableness by default. If the objections feel thin, try adding “be more aggressive” or “assume I’ve already heard the polite version” and see if a second pass sharpens them.
The Actual Prompt Structure Worth Trying
Here’s the pattern that tends to work:
- Set a role explicitly: “You’re a skeptic / investor / critical editor who thinks this is flawed.”
- Give it something concrete to attack: A paragraph, a plan, a decision you’re sitting with — not a vague topic.
- Specify the format: “Give me the three strongest arguments against this, in order of how damaging they are.” Ranked objections are more useful than an unordered list because you can see where the model thinks the real weight is.
- Don’t immediately defend yourself: Read the output first. The objections that made you defensive fastest are usually the ones worth thinking about longest.
In my experience, Claude tends to hold the adversarial role more consistently than ChatGPT — it’s less likely to hedge back toward balance halfway through the response. But either works if you push the framing hard enough.
If the model’s counterargument doesn’t change anything about your thinking, that’s information too. Either the idea is solid, the objections were weak, or you’ve already made up your mind and were looking for confirmation. Knowing which one is happening is worth more than the output itself.
