WORKFLOW · 5 MIN READ

Give every agent a job you can verify

A practical way to turn vague AI coding requests into assignments with evidence and a clear finish line.

Start with an outcome

“Improve this app” is an understandable goal, but it gives an agent too many possible directions. Begin with a concrete outcome: “The voice shortcut should open a recorder in Code mode, and the transcript should appear in the selected terminal without running it.” A specific outcome makes it easier to choose the right files, test the change, and know when the work is done.

A useful brief says who is affected, what happens today, what should happen instead, and which boundaries matter. If you want the agent to draft a command but wait for you to press Enter, say that explicitly. The boundary is part of the feature.

Ask for evidence, not a confident summary

Good agent work leaves traces you can inspect. For a bug fix, ask which failure was reproduced, which path changed, and what test or manual check passed. For a design change, ask to see the page at desktop and phone widths. For a workflow, ask for the state transitions and the failure path.

Keep the evidence close to the work. A terminal transcript, file diff, and approval request in the same workspace make review faster than a polished paragraph that says everything looks fine.

Use a small finish line

Split large ambitions into reviewable steps. A useful sequence is: reproduce the problem, fix the smallest cause, test the affected path, then decide what remains. That sequence does not stop you from building a powerful product. It keeps each addition trustworthy.

When a change touches permissions, accounts, payments, or deployment, make the final action visible. The agent can prepare the implementation and checks; you should be able to inspect what will happen before anything goes live.