Show the work
Prompts, assumptions, logs, and runbooks should be visible enough to debug a bad run.
About
My bias is toward systems people can inspect: readable code, visible prompts, useful logs, and review points that keep a human responsible for the important calls.
Operating stance
The model call is rarely the hard part. The hard part is deciding what should happen before the model runs, what evidence it should use, where it should stop, how it should explain itself, and what a human should do when it is uncertain.
That is why I start with operational mapping before implementation. Once the task is clear, the build can stay small, observable, and easier to maintain.
Prompts, assumptions, logs, and runbooks should be visible enough to debug a bad run.
The important parts should live somewhere the business can inspect, back up, and change.
Use larger cloud models when they are worth it, but record why and keep the system legible.