What to delegate to a model, what to keep, and how to tell
The productivity gain is in choosing tasks well. Delegating the wrong task costs more in review than it ever saved in drafting.
| Ask | If yes | If no |
|---|---|---|
| Can I verify the output quickly? | Good candidate | Review will cost more than doing it |
| Is being 90% right useful? | Good candidate | Keep it — partial correctness is worse than none |
| Is the input already written down? | Good candidate | You will spend the saving on explaining |
| Would a competent stranger manage it? | Good candidate | It depends on context only you hold |
| Is a mistake cheap to undo? | Good candidate | Add a human checkpoint before the action |
| Task | Why it works | Verify by |
|---|---|---|
| First draft of a known document | Blank-page cost removed; you were editing anyway | Reading it as the recipient |
| Summarising a long thread | Compression is checkable against the source | Spot-checking the decisions it lists |
| Reformatting and restructuring | The facts are already yours | A quick diff of the content |
| Explaining unfamiliar material | You will notice when the explanation stops making sense | Following one claim to a primary source |
| Generating options to choose from | Judgement stays with you | Nothing — you are picking, not accepting |
Whether you can verify the result quickly. If checking it costs more than doing it, delegation is a loss no matter how good the draft looks.
They are usually the tasks where you held all the context. If writing the prompt takes longer than writing the document, that is the signal to stop.