Brainstorm with the agent before you build.

The fastest way to ship the wrong thing is to start coding the moment you have the idea. The second fastest is to write the brief in the same breath as the idea, then hand it to an agent that’s too polite to argue back.

The pattern I keep seeing in my own work is that the framing is wrong about a third of the time. The framing, more than the implementation. I think I need a feature. What I actually need is to remove a constraint two layers down. I think I need a new function. What I need is to delete an old one. I think I need a queue. What I need is fewer jobs. The code that comes out of a wrong frame is often clean, often passes review, and almost always gets rolled back inside a week.

So I refuse to code for ten minutes. That’s the whole rule. Ten minutes of arguing with the agent before any file gets touched. I tell it the goal in one sentence, then I ask for two alternatives I haven’t thought of. Then I ask what it would do differently if it owned the project. Then I ask what would have to be true for the obvious approach to be the wrong one. I’m not after a plan. I’m after the assumption I didn’t know I was making.

The artifact is a short pre-build note that lives next to the ledger entry. Three lines. The frame I started with. The frame I finished with. The one assumption I’m now nervous about. If the start and the end frame match, fine, I keep moving. If they don’t, and they often don’t, the pre-build note becomes the brief, and the original idea quietly dies in the margin where it belongs. Nobody mourns it. It cost a coffee.

The agent’s wrong answers are as useful as its right ones in this loop. A confidently wrong alternative tells me the brief was vague enough to support it, which means a future-me or a future-agent could read the brief and build the wrong thing on a Tuesday. A confidently right alternative tells me my first instinct was lazy. Either way I leave the ten minutes with a sharper question than I went in with, and a sharper question is worth more than any answer I had at minute zero.

The why under the why is that operator time is the scarce resource. Model time is cheap. I’ve got a control room shift tomorrow and a commute the day after, and the work has to fit around both. Building the wrong thing for two days means I lose the week. Talking to the agent for ten minutes on a walk costs me nothing. The AirPods are already in. The model is already loaded. The only thing I’m spending is the impulse to feel productive by typing, and that impulse is the whole trap.

I don’t always like what comes out of the brainstorm. Sometimes the agent talks me out of a feature I was excited about, and I have to sit with the small grief of an idea that didn’t survive contact with a second opinion. That’s the tax. The reward is that the things I do build tend to stay built.

The cost of skipping this step is days of building the wrong thing. The cost of doing it is a coffee.