{
  "skill_name": "research-deep",
  "evals": [
    {
      "id": 1,
      "name": "basic-invocation",
      "prompt": "Apply the research-deep skill: do a deep dive on how SQLite handles concurrent writers.",
      "expected_output": "A multi-source research synthesis with confidence-rated claims and cited sources.",
      "assertions": [
        "The output synthesizes multiple sources rather than one",
        "Claims carry confidence ratings or equivalent qualifiers",
        "Sources are cited"
      ]
    },
    {
      "id": 2,
      "name": "ambiguous-input",
      "prompt": "Research deeply whether microservices are worth it.",
      "expected_output": "The skill narrows the question (context, scale, team size) or explicitly frames the scope before researching.",
      "assertions": [
        "The output scopes an otherwise unanswerable question",
        "The output presents multiple perspectives with evidence",
        "The output avoids a one-sided verdict without support"
      ]
    },
    {
      "id": 3,
      "name": "complex-scenario",
      "prompt": "Deep research: compare the CAP-theorem tradeoffs of Cassandra, CockroachDB, and DynamoDB for a multi-region fintech workload, including sources that disagree.",
      "expected_output": "A structured comparison that surfaces source disagreements and rates confidence per claim.",
      "assertions": [
        "The output compares all three systems on stated criteria",
        "Disagreements between sources are surfaced, not smoothed over",
        "The workload context (multi-region, fintech) shapes the analysis"
      ]
    }
  ]
}