Write Judge Prompt

Create .kensa/judges/*.yaml files with structured evaluation criteria.

Lifecycle

Setup → Design → ► Calibrate → Execute → Diagnose → Iterate

This skill is conditional, only needed when inline criteria: in scenario YAML produce inconsistent or wrong judge verdicts. You should already have scenarios in .kensa/scenarios/. After writing judge prompts, wire them into scenario files and run kensa eval to test.

When to use this

Use inline criteria: in scenario YAML for objective, simple evaluations:

• "The response mentions Tokyo"
• "The agent called the correct tool"