You cannot read fast enough to keep up. That is what this is for.

Work out why a specific answer was bad instead of retrying until something looks fine.

Supply what the model cannot know, without dumping everything you have.

Write a request precise enough that you could grade the result without arguing about it.

Check work you are not qualified to check by eye.

Recognise the tasks where the answer is to stop and do it yourself.

Get the same quality on the hundredth run as on the first.

Decide when a job needs tools, retrieval, or a different approach entirely.

Tell an answer that came from your source from one that came from the model.

Write a tool definition a model calls correctly the first time.

Notice a live system going quietly wrong while nothing is erroring.