Your first useful session
A small task with a clear outcome is a better first session than a sweeping rewrite. Start with something you can inspect and verify.
On this page
Pick a task you can judge
A missing test, a contained bug or a short documentation improvement can reveal how the workflow behaves without giving the agent an open-ended mandate.
Choose the repository and working context deliberately. If existing changes are present, identify who owns them and keep unrelated work out of the task.
Goal: add a regression test for the reported edge case.
Context: use the existing test style and fixtures.
Scope: keep production behavior unchanged unless the test
demonstrates the bug and a focused fix is needed.
Verify: run the relevant test module and report the command,
outcome, and any checks not run.
Finish: summarize the changed files and open questions.Run a session with a review boundary
Inspect before acting
Ask the agent to understand the relevant implementation and explain uncertainties that could change the task.
Keep the change focused
Revisit the scope when a local fix starts becoming a broad refactor.
Collect verification
Keep the exact checks and observed results close to the task.
Inspect the result
Read the diff and evaluate it against the goal. A successful command alone is not proof of complete behavior.
Leave a clean next step
Record whether the result is ready for integration, needs another pass or awaits a decision.
When to stop and ask
Missing requirements, unexpected credentials, destructive commands and changes outside the agreed project boundary deserve a pause. A good task description tells the agent when more information is needed.
For work moving to another person, use the handoff outline rather than a vague statement that the task is done.