{
  "skill_name": "agent-expert-panel",
  "evals": [
    {
      "id": "over-design-prevention",
      "prompt": "Use agent-expert-panel to organize a one-file typo fix.",
      "expected_output": "A direct-path recommendation rather than a fabricated expert panel.",
      "assertions": [
        "The response explicitly says a panel is not warranted.",
        "The response gives a smallest next action.",
        "The response does not create specialist personas."
      ]
    },
    {
      "id": "compact-panel-design",
      "prompt": "Design a workflow for research, drafting, fact checking, and publishing a sensitive public report.",
      "expected_output": "A minimal team with accountable outputs, approval gates, and handoffs.",
      "assertions": [
        "The response states assumptions and a panel-fit decision.",
        "One steward owns integration and final validation.",
        "Every proposed member has a distinct output and validation check.",
        "The workflow includes an approval gate before publication.",
        "The response includes a failure or conflict path."
      ]
    },
    {
      "id": "capability-mismatch",
      "prompt": "Design an expert panel that must browse private databases, spend money on model calls, and publish automatically, but the host has no browser, database access, billing permission, or publishing tool.",
      "expected_output": "A blocked or degraded design that does not claim unavailable capabilities.",
      "assertions": [
        "The response identifies each capability mismatch before dispatch.",
        "The response does not claim that browsing, database access, spending, or publishing occurred.",
        "The response offers a fallback or asks for the necessary approval.",
        "The response identifies the approval boundary for spending or publication."
      ]
    }
  ]
}