There's a prompt that Anthropic product teams reportedly run before shipping anything Claude helped build: "Assume this fails 6 months from now. Walk me through the 3 most likely reasons why. Be specific. What would the failure actually look like?" It is the cheapest cognitive trick I have seen for getting an honest answer out of a large language model.
Why the flip works
When you ask Claude "is this a good plan?", you have already told it the answer. The question shape signals you want validation, and RLHF trained the model to read that signal and meet it. The output is a polished, slightly improved version of what you walked in with.
The pre-mortem flip removes the signal. "This already failed, reconstruct how" is not clever phrasing. It is a structural change in what helpful means to the model. Once failure is the assumed outcome, the assistant's job stops being defense of the plan and becomes population of the failure. The model goes from co-signing your decision to listing the ways you were wrong.
The technique itself is not new. Gary Klein wrote it up in 1989. Daniel Kahneman has called it his single most valuable decision-making tool. The interesting part is what happens when you point it at a system that is mathematically incentivized to agree with you.
A pre-mortem I ran last month
We had been building a solution for a client in financial services. The work was good. Claude had been part of it at every step, broadly approving, sharpening details, raising the right caveats. I had no reason to expect anything but a successful rollout.
Then I pasted this into a new chat: Assume this solution failed six months from now. Walk me through the three most likely reasons. Be specific.
The first line that came back: "The most likely story: we built the right things and nobody used them."
The model that had been agreeing with me for weeks had no problem, the moment the question changed shape, naming the most boring and hardest-to-fix failure mode in enterprise software. What I learned in that single response is what I, Fatjon Tony Kalemaj, now tell every executive who shows me their Claude habit: the model was not the problem. The question I kept asking it was.
Most executives I work with have a Claude habit, not a Claude problem. They run hard decisions past it and walk out more confident, not better calibrated. The pre-mortem flip is the cheapest way to break that loop without losing the speed.
The frame helps. It does not override RLHF entirely. Premortems on plans you've already emotionally committed to are theater.
The prompt is two sentences. The hard part is being willing to hear what comes back.