eval-harness

To set up AI model evaluation for Claude Code sessions, define your first eval with /eval define feature-name, then implement, run checks, and generate a report using /eval report feature-name.

Eval-driven development for prompt engineering validates AI behavior through capability and regression evals, enforcing pass@k and pass^k metrics to improve development reliability across model updates.

Yes, the AI testing framework supports code-based, model-based, and human evaluators as graders, providing flexible scoring options for validating Claude Code session behavior.

The best way to prevent regressions in AI workflows is implementing regression evals that enforce pass@k and pass^k metrics, storing definitions and logs under the .claude/evals directory.

No external dependencies are required to run capability evals; the framework operates self-contained within Claude Code, managing eval definitions and logs under the .claude/evals storage pattern.

Defining evals before implementing features establishes clear capability and regression benchmarks, ensuring the end-to-end workflow from definition to reporting maintains AI development reliability.