GitHub Copilot

Choose the execution location before you write the prompt. Local agent mode modifies the development environment you are using, while the cloud agent works in a separate GitHub environment and prepares a branch. For a small function cleanup, reviewing a local suggestion may be easier than configuring a delegated task; for a repository issue, a branch-based handoff may be useful.

State the behavior that must survive: public names, return types, exception messages and ordering rules. Ask for the smallest structural change that improves one named problem. GitHub documents naming, conditional simplification and method extraction, but the useful output for a maintainer is a reviewable change with preserved behavior, not a shorter file at any cost.

• Keep the public function signature unchanged.
• List the observable behavior before suggesting a replacement.
• Reject new dependencies or unrelated formatting unless requested.
• Run the existing checks against the original and changed code.

Editorial assessment

Free is a reasonable starting point for a small local suggestion when an allowance is available. Pro is the first paid individual tier with cloud-agent and code-review access; higher tiers primarily need justification from your actual workload and feature access. Do not buy a large allowance merely because the repository is large: a narrow change can use little context, while repeated whole-repository conversations can consume much more.

Credit cost depends on the selected model and input, output and cached tokens. Repeating a prompt after adding logs or more files changes the work being billed. Keep the first experiment small, record the dashboard balance before and after it, and budget for review iterations; that observation is useful for your repository but should not become a universal claim about everyone else’s allowance.

Cloud-agent budgeting also needs GitHub Actions minutes, alongside AI Credits. Code review consumes Actions minutes on private repositories too. Check the two allowances separately before delegating a long-running task; a credit balance alone can hide the Actions cost of the workflow.

Do not infer that from an organization content-exclusion setting. GitHub identifies unsupported modes and indirect context limitations, including information supplied through IDE features. For a confidential project, confirm the exact mode and policy with its owner before providing context; use a sanitized fixture if the policy does not clearly permit the proposed workflow.

1. Write assertions for standard totals 99 and 100, express total 100, and an unknown shipping method.
2. Ask Copilot to replace nested branching while preserving all four results and the function signature.
3. Inspect the diff before accepting it; reject changed error text, broad file edits or a new dependency.
4. Run the assertions, inspect usage and save the final diff with the original fixture.

Pass only when the results remain 5, 0, 12 and unsupported, and you can explain the new branch order. Fail if the express case becomes free or the unknown method silently receives a standard price. If access or quota is unavailable, stop the task rather than treating an incomplete response as a successful refactor. For a separate code explanation, the ChatGPT and Claude profiles describe general assistants; the editorial policy explains the evidence standard.