The weakest coding-agent prompts describe a change and leave completion as an exercise for the agent. Add a button. Fix the API. Refactor the service. That might be enough to generate a diff, but a diff is not the same thing as finished work. The best improvement is usually a clear definition of done.
For me, that means observable checks: the route returns the expected status, the old behaviour still passes, the UI renders at the relevant widths, the migration works on existing data, and the production-safe dry run is clean. OpenAI’s original Codex guidance makes the same practical point from another angle: agents do better with configured environments, reliable tests, and clear documentation.
A definition of done also helps the agent make smaller decisions without repeatedly asking what matters. It can choose an implementation while staying anchored to the acceptance criteria. When something cannot be verified locally, the gap becomes explicit instead of disappearing behind a confident summary. That makes review faster because the evidence arrives with the change.
This is not special prompt magic. It is the same discipline good engineering work has always needed, made more visible because an agent will follow vague instructions literally. Tell it what to change, where the boundary is, and how both of you will know the job is complete. That will improve the result more reliably than adding another paragraph of motivational prose.