I recently pitted Claude and Codex against each other in debate mode and got a funny result: Claude argued for option A, Codex for option B. After a round of mutual critique they simply swapped sides, taking care to thank each other for the “epiphany”.

Claude explained it like this: “Both Codex and I were on the fence - that is a signal the decision is genuinely borderline and hinges on one variable that only you know.” In other words, the models are trained to be helpful and avoid conflict, and when the task statement is incomplete, they start mirroring each other and hallucinating the missing context just to reach agreement.

In cases like this I switch from “customer” mode to “architect” mode, using a prompt I picked up from Reddit:

I’m about to start this project: [3-5 sentences about the project]. Interview me until you are 95% sure you understand what I actually want, not what I think I should want. Challenge my assumptions. Ask about edge cases I haven’t considered. Don’t look at other files in the project, only this request.

Why it works:

1. The 95% threshold pushes the model into the role of a meticulous interviewer: it asks questions without worrying about being a nuisance.
2. Shift Left. A goal-setting error is famously one of the most expensive ones: you can move fast and produce high-quality output in the wrong direction.
3. The interviewer effect. While you answer hard questions, the answer often surfaces on its own.

My experience with this varies, but it has been useful. Sometimes halfway through the interview I realize the project isn’t worth doing yet: either too early, or much bigger in scope than it seemed. Sometimes a long interview leads to a much narrower and more precise statement of work. And sometimes the answer to a question appears right in the middle of the conversation.

What this means. For simple tasks you can just tell the model: “do it like this.” I do that often, because it works.

But for complex tasks the “architect” mode pays off: treat AI as a talented but very eager-to-please assistant who first needs help framing the task properly, otherwise it will run off and do something else.

#AI #GenAI #Productivity