Plan at the level of risk
A one-line copy change does not need ceremony. A change spanning data, UI, and deployment does. Planning is most valuable when it exposes assumptions, ownership boundaries, and verification steps before edits make those choices expensive.
Execution needs a finish line
Agents are more reliable when they know what evidence closes the task. A focused test command, a required route, a screenshot size, or a performance budget is more useful than an abstract request to make something production-ready.
Review the behavior first
A tidy diff can still solve the wrong problem. Start review with the user flow, edge cases, and visible result. Then inspect the implementation for maintainability and safety. That order prevents code style from becoming a substitute for product judgment.