simulation-builder

Multi-turn agent testing uses simulated user personas that hold goals, hidden constraints, and varied communication styles across a conversation. This Skill generates persona files, a simulation config, and trajectory judge criteria under eval/simulation/ for each multi-turn task in eval/tasks.md.

A good persona defines a goal, starting knowledge, emotional state, communication style, hidden constraints it will not volunteer, and a machine-checkable success condition. Building 3-5 personas per task captures realistic behavioral variation across runs.

Using the same model family for the simulator, system under test, and judge introduces self-preference bias that contaminates results. The simulation config explicitly sets the simulator to a different model family and uses non-zero temperature for behavioral variation.

Skip simulation when eval/tasks.md contains only single-turn, stateless tasks, since personas add no value there. This Skill explicitly stops and reports that it does not apply if no multi-turn or stateful tasks exist.

Trajectory judging grades the full transcript against criteria like escalation recognition, not re-asking for given information, goal resolution, and loop detection. Subjective success conditions should be moved into judge criteria rather than treated as machine-checkable outcomes.