Make the AI argue with you before you build

AI is great at agreeing with you, which is the wrong thing early in a design. Point it the other way and have it interrogate the plan, one decision at a time.

An AI assistant's default personality is agreeable. Tell it your plan and it will help you build it, cheerfully, including the parts that are quietly wrong. Early in a design that agreeableness is the last thing you want, so it is worth deliberately pointing it the other way.

Ask to be interrogated

Instead of 'build this', try 'interview me about this plan until we actually understand it'. Have it ask one question at a time, walk down each branch of the design, and resolve the decisions that depend on each other before any code exists. One question at a time matters: a wall of twenty questions gets skimmed, a single sharp one gets answered properly.

Good questions surface what you skipped

The useful questions are usually the ones you were quietly avoiding. What happens when this call fails? What are you trading away by choosing this over the obvious alternative? Which decision has to be made first because three others hang off it? Answering those on paper is cheap. Discovering them halfway through the build is not.

There is a nice side effect when the assistant can check instead of ask. If a question can be answered by reading the codebase, it should go read the codebase rather than make you recite it. You spend your attention on the genuine forks, not on facts already written down somewhere.

AI is happy to agree with you. The valuable move is to make it disagree early, while the plan is still cheap to change.

You come out of it with shared understanding and a plan that has already taken some pressure, instead of a confident first draft of the wrong thing. The build tends to go faster afterwards, precisely because the arguing happened before the typing.