So, I've stopped asking AI models for advice or brainstorming 6 months ago... Default LLMs are the worst kind of co-founder: they are polite yes-men that validate every single terrible idea you have 😫 You paste in an untested business concept, a vague pricing model, or an overcomplicated architecture. The model responds with 5 enthusiastic paragraphs praising your brilliance, followed by 10 generic bullet points that look like a carousel from 2022 It feels productive in the moment. In reality, it is just statistical flattery that hides the landmines in your plan I only fixed the dynamic when I stopped using AI as an answer engine and started prompting it as an adversarial sparring partner. The model doesn't need to give me answers. It needs to stress-test my thinking until the weak assumptions break. Here is the exact 4-part sparring structure I drop into claude or chatgpt before I make any major architectural or business decision: - Ban all praise and agreement: In the first 2 lines, I explicitly forbid validation. I tell the model: "You are an adversarial board member. Do not summarize my idea, do not tell me it has great potential, and do not be polite. Your only job is to find the operational flaw that will kill this project in 90 days" - Force clarifying interrogation first: Instead of letting the model jump straight into giving suggestions, I add a blocking rule: "Ask me 3 uncomfortable, probing questions about my unit economics, distribution, or technical dependencies before giving any critique" - Identify the hidden operational tax: I make it calculate the unsexy labor: "Assume this idea works. What is the hidden operational friction (maintenance hours, API rate limits, customer support overhead) that will make me hate running this in 6 months?" - The steelman alternative: Finally, I force it to propose the contrarian opposite: "If a competitor wanted to destroy this offer by making it 10x simpler and half the price, what would their MVP look like?"