Parts, Not Whole Tasks

The unit is smaller than a task

Most disappointment starts by delegating at the wrong size.

"Write the quarterly report" is not a task. It is: gather the numbers, work out what changed, decide what matters, structure the argument, draft it, check the figures, adjust the tone for the audience.

Some of those are excellent AI work. One of them, deciding what matters, is the entire job and should not be handed over at all.

Hand over the whole thing and you get something that looks like a quarterly report and contains nobody's judgement.

Splitting one real task

Take something you actually do and break it into steps. For each, ask two questions:

Is this pattern work or judgement work? Transforming, summarising, restructuring, drafting from material you supplied: pattern. Deciding what matters, what to risk, what to say to whom: judgement.

What does a mistake here cost? A clumsy sentence costs nothing. A wrong figure in a board pack costs a lot.

Hand over pattern work with low failure cost. Keep judgement work and anything expensive to get wrong. The middle gets done together, with you watching.

Two ways of working

There are two distinct modes, and they suit different things.

Hand-off. You give a complete subtask with everything needed, it comes back, you review. Clean boundary. Good for well-specified, self-contained work: turn these notes into a draft, extract these fields, translate this.

Interleaved. You and the model alternate closely. Write a bit, ask for a critique, adjust, ask again. No clean boundary. Good for thinking, exploring, and anything where you do not yet know what you want.

Most frustration comes from using the wrong one. Interleaving on a mechanical task is slow. Handing off something you have not thought through produces confident work on the wrong problem.

Which to pick

Could you write down exactly what you want before starting?

If yes, hand it off. Specify properly, let it run, review the result.

If no, interleave. The conversation is how you work out what you want, and that is a legitimate use rather than a failure to prepare.

Work often moves between them. Interleave to figure out the argument, then hand off the drafting once you know what it is.