{
  "skill_name": "hf-agent-composition",
  "evals": [
    {
      "id": 1,
      "prompt": "I need to create a new GSD agent that reviews code for security vulnerabilities. It should scan files, check against OWASP top 10, and produce a report. What XML blocks do I need and how should they be structured?",
      "expected_output": "Agent definition with <role>, <execution_flow> (load_context, analyze, verify_and_write, report), <structured_returns> (SECURED/OPEN/ESCALATE format), <success_criteria>, and project_context block. No deviation rules needed (read-only agent).",
      "files": []
    },
    {
      "id": 2,
      "prompt": "My executor agent keeps getting stuck reading files forever without writing any code. How do I prevent this?",
      "expected_output": "Reference the analysis_paralysis_guard pattern: after 5+ consecutive reads without action, agent must STOP and either write or report blocked. Show the exact XML block to add.",
      "files": []
    },
    {
      "id": 3,
      "prompt": "What's the difference between a verifier agent and an auditor agent in GSD? How do their return formats differ?",
      "expected_output": "Verifier uses VERIFIED/FAILED/UNCERTAIN per-truth status with 4-level artifact checking (exists/substantive/wired/data-flows). Auditor (nyquist/security) uses SECURED/OPEN/ESCALATE or GAPS FILLED/PARTIAL/ESCALATE with count-based metrics. Both are read-only and escalate implementation bugs.",
      "files": []
    },
    {
      "id": 4,
      "prompt": "Design a checkpoint protocol for an agent that needs user to choose between React and Vue for a new component",
      "expected_output": "checkpoint:decision type with options table (React vs Vue with pros/cons/complexity), recommendation (front-loaded), completed tasks table, and awaiting section prompting user to select option.",
      "files": []
    },
    {
      "id": 5,
      "prompt": "I'm writing a planner agent. What discovery levels should I include and how do I prevent scope reduction?",
      "expected_output": "Reference discovery_levels (Level 0: skip, Level 1: quick verify, Level 2: standard research, Level 3: deep dive) and scope_reduction_prohibition blocks. Show the decision coverage matrix pattern and 'never simplify user decisions — split instead' rule.",
      "files": []
    }
  ]
}