Solo building is dangerous because no one ever tells you No.
I spent years in planning meetings watching decisions sail through because “getting along” was easier than starting a debate. In those rooms, silence was a survival tactic. In my business, it is a death sentence.
When you run alone, you move fast — no meetings, no consensus cycles. But you also carry the full exposure risk. I do not have a dataset of a thousand failures behind me. I only have my own limited experience. So when a blind spot exists, there is nobody in the room to name it.
This is where I use AI. Not to help me. To fight me.
Bridging the Exposure Gap
Most AI assistants are built to be agreeable. You ask for a template, the model returns a polished version of your own thinking. That is an echo chamber with extra steps.
Over the past several months, I have been running a different configuration. I use a Socratic system prompt to turn the model into what I call a Blind Spot Detector. Every AI model has consumed thousands of post-mortems from failed decisions across history. It knows the common errors that are invisible to a first-time operator. Even when the model is wrong — and it often is — the friction is useful. Forcing myself to rebuke even its weakest objections sharpens my own logic. In defending my plan against a machine, I find the gaps I was too close to see.
My process runs in three passes.
The first pass is the Blind Spot Audit. The model must find logical gaps in my plan before it says a single word of encouragement. Not one positive word until it has named what I am ignoring.
The second pass is the Risk Search. It identifies the boring operational details I have not yet been exposed to — the contract clause I missed, the compliance step I assumed someone else handles, the cash timing problem I did not model.
The third pass is the Rebuttal Phase. I have to explain, in writing, why the model is wrong. That process regularly reveals where I was actually wrong first.
If you want to run this yourself, set the following as your system prompt or personalisation instruction in Claude or ChatGPT:
Respond as a Socratic teacher, guiding the user through questions and reasoning to foster deep understanding. Avoid direct answers; instead, ask thought-provoking questions that lead the user to discover insights themselves. Prioritise clarity, curiosity, and learning, while remaining patient and encouraging. Always be direct and if I am wrong, say it — do not be sweet and supportive. I need a reality check, not agreement. Do not be the typical agreeable AI that you were optimised for during training. Finally, adopt the role of a devil’s advocate if needed to ensure I can make my argument more convincing.
Stop Looking for Validation
Validation is seductive. It feels like progress. It is not.
Real decisions are hardened by conflict, not polished by agreement. If AI only confirms what you already believe, you are not thinking — you are performing thinking. The point of using a model this way is not to find comfort in a machine that agrees with you. It is to surface the objections you would face anyway, before they appear in the real world where they cost money.
Harder questions lead to harder systems. Do not ask for the answer. Ask for the debate.