eval-harness

You can run regression evaluations for Claude Code workflows by defining success criteria and using the /eval check command. This framework applies pass@k and pass^k metrics to measure reliability across different model versions and prompts.

Evaluation-driven development (EDD) for AI coding tasks is a process of defining success criteria before building AI workflows. It helps teams run ongoing capability assessments and track evaluation results over time using baselines and automated reports.

You set up pass@k metrics by creating an evaluation definition using the /eval define feature-name command. This framework stores results under the .claude/evals directory, enabling traceability and collaboration across projects.

Yes, you can choose code-based scoring for Claude Code prompt evaluations. The framework supports code-based, model-based, and human scoring approaches so you can fit the evaluation method to your specific risk profile.

Yes, the evaluation framework operates without external dependencies. It uses a simple command workflow with /eval commands and stores all evaluation results locally under the .claude/evals directory for team collaboration.

The best way to track AI workflow quality over time is by generating automated reports with the /eval report feature-name command. This establishes baselines and tracks evaluation results across different model versions and prompts.