Ask, plan, execute

This is the single most useful habit in the whole method, and it is almost absurdly simple. Do not jump straight to agent mode, go build it. Move through three gears instead.

Start by asking. Give the AI the context and tell it what you want or what is broken, but stay in a mode where it explains and discusses rather than acts. You are loading it with the real situation and getting its read before anything changes. This is where you catch a misunderstanding while it is still just words on a screen.

Then ask for a plan. Have it lay out the steps it intends to take, and review them. This is your last cheap checkpoint. If the plan is wrong, the build will be wrong, and a wrong plan caught here costs you a sentence to redirect instead of a long cleanup afterward. Most of the pain people blame on the AI making a mess traces back to skipping this step.

Only then do you execute. By the time the AI acts, it understands the situation, you have agreed on the approach, and you are watching it follow a plan you already vetted rather than improvising in the dark.

This is almost certainly what separates the two studies in the introduction. The developers who got 56% faster were on a fresh, well-scoped task where a clear plan is easy to form. The experienced developers who came out 19% slower were in gnarly mature codebases where an unvetted plan quietly steers the AI into the weeds, and the cost of that shows up later as cleanup. The lesson is not that AI is fast or slow, it is that the plan-review step is where the speed actually lives.

Ask, plan, execute. It feels slower for the first ten minutes and is dramatically faster across the whole build, because you stop paying for misunderstandings in finished code. Make it the default rhythm and you stay in control of work that would otherwise run away from you.