Let the agent type: a workflow for people who hate typing

The failure mode people hit with coding agents isn't that the agent writes bad code. It's that they hand it an under-specified task, get something plausible back, and then spend longer reviewing it than writing it would have taken. Laziness done properly is front-loaded.

Write the rules down once

Every trap you've already paid for goes in a file the agent reads on every session. Not style preferences — those are cheap. The expensive ones: the column that has to be added to a grant list, the migration that has to be applied before the code that depends on it merges, the token name that silently collides with a framework utility. Each of those is a bug you only have to have once.

Specify the outcome, not the implementation

• Say what the thing must do and how you'll know it works.
• Say what it must not do — the constraints that aren't obvious from the code.
• Leave the how alone unless you actually care; you usually don't.
• Name the files you already know are involved. It's the cheapest context you can give.

The one review pass

Read the diff, not the explanation. The summary is written by the same model that wrote the code and shares its blind spots. Diffs don't. Look specifically at the edges: what happens on empty input, what happens on the second call, and what got deleted — agents remove things quietly far more often than they add things loudly.

Review the diff. The summary is a story about the diff.

Do that and the loop tightens into something genuinely lazy: describe, skim, ship. The typing was never the valuable part.