The Loop Is the Part I Keep Rebuilding

AI can write a lot of code very quickly, but I still need to check whether it understood what I meant. The useful part of the workflow is not the first draft; it is the loop after the draft.

I describe a small piece of work, let the model make a change, and then run the project instead of admiring the diff. The runtime usually tells me which assumption was wrong much faster than another round of prompt wording.

That is why I prefer tasks with a clear boundary. “Make the dashboard better” gives the agent too much room to invent, while “add this state, preserve this API, and prove the empty case works” gives both of us something to check.

The checks are part of the prompt even when they are not written in the prompt. Tests, type errors, browser behaviour, logs, and screenshots turn a vague request into evidence that I can accept or reject.

I also want the agent to show its work in a way I can review. A smaller patch with a useful explanation is easier to trust than a sweeping rewrite that happens to make the test suite green.

I keep rebuilding this loop because every project has a different failure mode. The model makes the first step cheaper; the engineering is still deciding what to observe, what to correct, and what counts as done.