Grilling

Grilling flips the usual direction of a prompt: instead of you describing the task, the agent asks the questions. You tell it what you want to build and instruct it to grill you, one question at a time, until it has no open questions left. The output is not code but a resolved understanding, usually written up as a design concept or a spec.

It works because the model is good at spotting gaps you can't see in your own request. You know what you meant by 'sync the data', so you never said which direction, how often, or what happens on conflict. The agent doesn't know, and in a normal run it would pick answers for you. Grilling surfaces those choices while they're cheap to make, before hours of AFK work are built on them.

Give it rules so it stays useful: one question per turn, no leading questions, stop when the remaining questions are about implementation detail rather than intent, then write the result down. Keep the session short and the context clean, and finish by asking for the spec as a file so it survives as a handoff artifact for the session that actually builds.