How to use a coding agent on one real issue
A useful coding session ends with a reviewable change and evidence that it works. Give the agent a clear boundary and inspect the result.
What you will make: A focused change with evidence for its behavior and a human review.
Reproduce and scope
Read the repository’s instructions and reproduce the issue before editing. Record the actual and expected behavior. If you cannot reproduce it, ask for the next diagnostic step instead of treating the first explanation as the cause.
AI can help: Help inspect the affected code and identify diagnostic steps.
Your review: Confirm the problem and expected behavior.
Find current context
Inspect the nearby implementation and installed dependency versions. Use version-matched documentation. Confirm whether the project already has a component or helper for the job before introducing a new dependency.
AI can help: Find version-matched documentation and related code.
Your review: Check that the examples match the actual environment.
Implement a small change
Work on a branch and request a focused change. Ask the agent to stop when requirements are ambiguous or access needs to expand. Review the diff while it is still small enough to understand.
AI can help: Draft a focused change on a branch.
Your review: Review dependencies and prevent unrelated changes.
Test and inspect
Run checks that exercise the changed behavior. Include a relevant failure case. Inspect generated tests to make sure they would have caught the original problem. Record what ran and what remains untested; a confident explanation is not a test result.
AI can help: Run relevant checks and summarize failures.
Your review: Inspect the diff and verify that tests cover the actual risk.
Approve and release
Review the final diff, dependencies and data handling. Have the responsible person approve merge and deployment through the normal process. After an authorized release, monitor the original symptom and keep a way to recover if the change fails.
AI can help: Draft release notes and a monitoring checklist.
Your review: Approve merge and release through the project’s normal process.
Keep a record of what worked.
Use the pilot scorecard to record quality, total time and failures. These walkthroughs are suggested ways to work; their tool combinations have not been tested end to end.
Get the scorecard ↗Publisher references
Reviewed October 6, 2026. Follow the publisher’s current setup instructions for your account and version.