GitHub Copilot Evals
InstallSearch /github/evals

Improve an eval

Refine a success rule that misses a mistake, while keeping the original evaluation.

Choose your app

Find your setup instructions
Create an eval form bound to the sample release-note skill while workspace trust is off.
Describe the task and expected behavior for this source. Evals Canvas · sample data

Choose your app to see how to start. Start with the first-eval guide for the general steps.

Save a separate version

Start from the original eval and a specific missed mistake. Review and save separate files; keep the original. Saving a new version does not prove it is better. Asking for a draft does not give permission to save.

Stay in the Improve conversation linked to the original eval, not Canvas’s inline Create panel. Follow the walkthrough and save/recovery rules.

In supported Canvas hosts, select the original eval, describe your change, and choose Send to Copilot. Improve keeps the original and asks only questions that change cases or checks. Existing answers and declined suggestions stay in effect.

Finish review and saving in conversation. Canvas keeps a read-only snapshot; closing it doesn’t cancel work, and reopening it doesn’t resend. After saving, optionally View saved eval in Canvas where supported. This opens the saved candidate, not an unsaved draft, and never runs it.

If the asset needs fixing instead

If the check caught a wrong answer, ask to improve the skill or agent or run /eval improve <skill-or-agent> in CLI or the App to choose focus areas and guide the loop.

Keep comparisons fair

An easier check can raise a score without improving AI work. Changed checks need the same answers; changed tasks need controlled new runs.

Improvement guide · Compare