The most expensive mistakes in AI-assisted work are not bugs. They are correct implementations of the wrong thing: a feature built exactly as described, on an assumption nobody said out loud. The fix is to reverse the roles for ten minutes and make the tool interview you.
The tool does not push back because it was not asked to. It builds, and it builds fast, and the wrong assumption is now in the codebase with tests around it.
The prompt
Before building anything that touches a database, a boundary with another system, or money, give the tool this instead of the task:
Interview me about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design one decision at a time. For each question, give your recommended answer. Ask one question at a time. If a question can be answered by exploring the codebase, explore the codebase instead.
That is the whole thing. A colleague turned it into a one-word command and it has travelled to every project since.
Why each sentence is there
One question at a time stops it collapsing into a questionnaire you answer superficially. A session typically runs ten to fifty questions, and you will want to think about most of them.
A recommended answer with each question means you are confirming or correcting a default, not composing from nothing. It keeps the pace up and it kills the fatigue that makes people stop answering carefully.
Codebase first means settled matters get answered from the code, so the questions you see are the ones only you can answer.
What it surfaces
On a dashboard build the interview asked: how is “complete” defined? Is that figure per board or per project? What happens when a task is moved backwards a column? None of those were in the brief. All three would have been resolved by a silent guess, and at least one guess would have been wrong.
Two ways to waste it
Answering “yes, fine” to everything hands the design back to the tool. The point is that you decide; the tool is there to find the decisions.
Not writing down the result. A ten-minute interview produces a design the tool will have forgotten by the next session. End every one by saving the resolved decisions to a file in the repository, or a note in your vault. That file is what the next session reads first.
When not to bother
Single-step tasks, cheap reversible experiments, and problems that are already fully specified. The interview is for the decisions that are hard to undo. For those it is the best ten minutes in the project.