Prompt like you order coffee

Order a black coffee and you get a black coffee. Nobody hands it back with milk, two syrups, and a dusting of cinnamon you now have to drink around. If you wanted those things, you’d have asked.

Models work the other way by default. Ask for a small feature and you’ll often get an abstraction layer, a config system, and a plugin architecture nobody requested. It all looks plausible, and it’s all in the way. I’ve watched an agent respond to “add a settings toggle” with a plan so elaborate it would have taken a week and never actually shipped. The tell is always the same: lots of scaffolding around the thing, very little of the thing itself.

The problem with all that extra sugar is that you can’t get it back out once it’s dissolved. Complexity is easy to add on the way in and painful to remove later, especially when it’s threaded through files an agent wrote at two in the morning.

So I’ve started prompting for the black-coffee version on purpose. I ask for the smallest change that solves the actual problem, and I add a line telling the model to stop there and wait. Then I add the milk in a second pass, once I’ve tasted what’s underneath. It’s a small habit that has saved me more rewrites than any clever prompt template ever did. I wrote up the longer argument for keeping AI code simple over at Codewalkers.