For a tiny, obvious edit, a plan may be one sentence. For anything with real choices, ask the agent to inspect first and propose a short plan before it writes.
A useful plan contains decisions
It should tell you:
- Which files or areas appear relevant.
- What the agent intends to change.
- How it will verify the result.
- Which assumptions could change the shape of the solution.
A plan that merely repeats your request adds ceremony. A plan earns its keep when it exposes a fork: two reasonable interpretations, two storage choices, two trigger conditions, or a tradeoff between a local fix and a shared one.
Ask the agent to name those forks, recommend an option, and wait. It should do the investigation; you should retain the decision.
Keep approval small
Approve one bounded task, not a roadmap containing every future improvement. When the approved unit is small, you can compare the eventual diff directly to the plan that authorized it.
Skipping the plan does not skip the decisions. It hides them inside the code.
This really happened. An agent building a retry queue had to decide exactly when an item should go back in the queue. Two readings were possible. It quietly picked the one suggested by nearby code, and it picked wrong; the human had meant the other. The habit that prevents it: say the condition out loud, name the other reading, and wait for the pick. That pattern is entry AF-14 on the watch-out list.
When you are ready for a deeper version of this habit, continue to Make the agent interview you.