eval-harness

Eval-driven development for Claude Code requires creating standardized eval definitions in .claude/evals, then running commands to execute checks and generate reports with pass@k and pass^k metrics.

Capability and regression evals validate new features and prevent regressions by continuously checking changes against structured evaluation criteria defined before implementation.

Software testing evals support multiple grader types including code-based, model-based, and human graders to score changes within the integrated workflow from definition to reporting.

Yes, you can define custom evaluation criteria before implementation by creating a standardized eval definition file under .claude/evals and executing it through the formal workflow commands.

Measuring pass@k and pass^k reliability metrics is handled natively by the eval-harness framework, which generates these metrics automatically when you run the report command on your eval definitions.