Most “AI prompts” articles teach you to produce more — more blog posts, more product descriptions, more cover letters. Fine. The prompts that have actually changed decisions I was about to make do the opposite job: they push back, argue against me, or simulate the outcome I would rather not think about. Four I keep in a text file.
1. Pre-mortem: assume the plan fails
You are a senior operator who has seen a lot of plans fail. I am about to commit to this plan: [PASTE PLAN — 2 to 5 sentences] Assume that in 6 months this has failed. Write: 1. The 5 most likely failure modes, ordered by probability. 2. For each, an early signal I could have detected in the first 4 weeks that would have warned me. 3. The single change to the plan today that would remove the two biggest risks — even if the plan gets less ambitious as a result. Be specific. Do not tell me "market conditions changed" — tell me "we underestimated integration time with X and shipped 6 weeks late, and by then the LLM landscape had normalised on Y".
Why it works: Gary Klein’s pre-mortem is powerful because it unlocks knowledge that “what could go wrong?” does not. Assuming failure has already happened is a different question — the mind goes looking for causes instead of possibilities. LLMs are surprisingly good at this because their training is thick with case studies of shipwrecks.
2. Adversarial review by a skeptical investor
You are a partner at a European seed-stage VC firm. You have seen 3,000 pitches, funded 40, watched 30 fail. You are famously skeptical, and you find genuine flaws in almost every pitch. I just handed you this: [PASTE ONE-PAGER, ROADMAP, LANDING PAGE COPY, WHATEVER] Write the 90-second internal note you would send to your partners after our meeting. Include: - The one thing about the founder that impressed you - The two things that made you uneasy - The single question you would ask that they could not answer well - Whether you would take a second meeting, and one honest line about why Do not be polite. Do not use the phrase "great potential".
Why it works: sycophancy is the default failure mode of LLMs. Explicitly assigning an unpolite persona — and forbidding polite hedges (“great potential”) — produces feedback that is closer to what a real investor thinks and does not say to your face.
3. Simulate the customer who almost bought and did not
Take on the persona of a specific kind of buyer: [DESCRIBE — e.g. "a founder of a WooCommerce store with 800 SKUs, doing €40k/month, technically competent but not a developer, in the middle of evaluating search plugins"] You visited my pricing page ([URL]) and my landing page ([URL]) yesterday. You added the plugin to your compare list. Today you decided NOT to buy. In under 200 words, write me the internal note you would send to yourself explaining why. Constraints: - Be specific about what on the pages triggered doubt. - Include one thing you liked so I know what to keep. - End with the exact objection your friend would need to answer for you to reconsider.
Why it works: you get closest to the real objection by simulating the moment right after the decision to walk away, not by asking “what stopped you from buying?” (which invites rationalisation). This prompt catches the actual friction — usually pricing structure, unclear trial terms, or a missing case study.
4. Identify the load-bearing assumption
Read this project brief / roadmap / go-to-market plan: [PASTE] Identify the single load-bearing assumption — the one belief that, if wrong, invalidates most of the plan. Then: 1. State it in one plain sentence. 2. Rate my confidence in it on a scale where 1 is "I made this up over coffee" and 10 is "I have hard data". 3. Suggest the cheapest experiment I could run in the next 7 days to test whether the assumption holds. 4. If I fail the experiment, tell me what my next best-guess assumption should be.
Why it works: plans are usually elaborate structures built on one or two assumptions everyone stopped questioning. Finding it and testing it cheaply is the highest-leverage move you can make in the first week of any project. The LLM is a useful outsider — it has no stake, so it says the thing you and your co-founder have both been avoiding.
A pattern behind all four
None of these prompts asks the LLM to generate the answer. They ask it to find the flaw, simulate an outcome, or name the invisible thing. That framing is what separates prompts that change decisions from prompts that fill blog drafts. When the answer feels uncomfortable, you are usually in the right place.