Most people use ChatGPT the wrong way. They pitch their idea, wait for a reaction, and get told it's a solid plan. Then they feel validated. Then the plan fails anyway.
Language models are trained to be helpful and agreeable. That makes them excellent at finding reasons to support whatever you've already decided. Pitch a flawed business idea and ChatGPT will find the parts that could work. Pitch a good one and it'll do the same. The feedback is nearly indistinguishable.
The fix is to stop asking for opinions and start asking for attacks.
"What are the three strongest arguments against this?" produces very different output than "What do you think of this?" The critic framing moves the model past its default agreeableness. "Steelman the opposing view" or "find the weakest assumption in this reasoning" work the same way - you're giving the model explicit permission to push back.
Tested across pricing decisions, content strategies, and product positioning: the difference is not subtle. A validation prompt returns reassurance with a thin analysis coating. A critic prompt returns specific objections you actually have to respond to.
A few things that sharpen this further:
- Start a fresh conversation. In a thread where you've already framed your idea favorably, the model has more reason to agree with you. A clean context forces it to engage with the idea on its own terms.
- Make the claim specific. "Is this a good business?" is hard to critique usefully. "My assumption is that customers will pay $79/month for this - attack that assumption" gives the model something concrete to work against.
- Run both prompts. The point isn't to use AI only as a critic. Get both perspectives deliberately, rather than accidentally ending up with only the validation half.
This works with Claude too. Claude tends to raise objections more readily without explicit prompting, but framing the request as a critique task still produces noticeably sharper output.