The "strongest case against" prompt
Models agree with you by default. This prompt forces the argument a smart skeptic would make against your plan, plus the assumptions that have to hold for it to work anyway.
- Use case
- Decision stress-testing
- Works with
- Claude, GPT-class
I am about to do this: {describe the plan or decision, and roughly what it
costs in money or time}.
Make the strongest honest case against it. Not a list of generic risks:
the specific argument a smart person who knows this area would make if
they were trying to talk me out of it. Include at least one problem I
probably have not thought of.
Then answer one question: what would have to be true for this plan to work
anyway? List those assumptions plainly, so I can go check them.
Do not balance this with a case in favor. I already believe the case in
favor. That is why I need you.
Left alone, every model is a yes-man. Describe a plan and ask “what do you think” and you get enthusiasm with a light dusting of “considerations.” That is worse than useless, because it feels like due diligence.
I run this before spending real money or committing real weekends. It has talked me out of exactly the purchases and projects it should have.
Why it works
- “The specific argument a smart person would make” raises the bar from risk-listing to actual argument. “It might not work out” is a risk. “You are buying a solution to a problem you have hit twice, and both times the workaround took ten minutes” is an argument.
- “At least one problem I have not thought of” pushes the model past restating your own doubts back at you, which is its comfortable default.
- The assumptions question is the real product. A pure takedown is easy to dismiss. A list of “this works if X, Y, and Z are true” gives you homework: go find out if X is true. That converts the critique into something you can act on.
What good output looks like
Uncomfortable. If you read the case against and feel nothing, either the plan is genuinely solid or the model went soft on you. The good runs are the ones where a single sentence names the thing you had been quietly not looking at.
Where it works (and doesn’t)
Claude is notably good at this when explicitly given permission to be critical, which is exactly what the last two lines are. Without the “do not balance this” instruction, most models sandwich the critique in reassurance, and the reassurance is what your brain will keep.
One warning: the model argues against anything with equal confidence, including good plans. This prompt finds the strongest case against, not the truth. Checking the assumptions is still your job.