What problem does it solve? Building offline evaluation datasets for AI agents often produces inconsistent, fabricated, or non-representative test cases. This Skill explores an agent's real prompts, tools, business rules, and code first, builds a capability map and mutually exclusive scenario plan, then generates traceable, self-consistent JSONL benchmark data only after user confirmation. ## Core Features & Use Cases - Capability and Scenario Exploration: Reads agent prompts, tool definitions, business rules, and code to build a capability map with evidence-backed, mutually exclusive scenarios before any data generation. - Two-Stage Confirmed Generation: Produces a small sample JSONL set for user review first, then generates the formal dataset only after explicit approval, with rejected cases written to a separate rejected.jsonl. - Validation and Coverage Tooling: Ships scripts to inspect agent bundles, validate capability maps, check JSONL structure and cross-field consistency, and report scenario coverage against the capability map. - Use Case: Point it at a tool-calling or RAG agent directory, confirm the discovered capabilities and scenarios in Chinese, and receive a validated offline evaluation dataset for regression testing or synthetic multi-turn dialogue. ## Quick Start Explore the agent in ./agent_bundle, summarize its capabilities and concrete scenarios in Chinese, and wait for my confirmation before generating any sample JSONL data.
