Fight the cost of inference.
Every time a model runs, you pay for it. The useful tooling is what cuts how often the expensive run happens, and how much of the job you send to it.
Finish it in one shot
A second try is a second bill. Prompt for a finished result the first time: say what done looks like, give only the context that matters, and name the format you want back. Retries and “make it shorter” loops are inference you did not have to buy.
Local and cloud, together
Not every step needs a frontier model. A hybrid setup keeps ordinary work on a local model and sends only the hard step to the cloud. The cloud is the specialist you call, not the machine that sees every draft.
Plan in the cloud, work locally
Ask a cloud model for the plan: the outline, the steps, and what finished looks like. Then give that plan to a local model and let it do the writing and revising. You pay the cloud for the plan, not for every draft.