{
  "skill_name": "harness-automation",
  "evals": [
    {
      "id": 1,
      "prompt": "我们的 PRD 和设计已经定稿，准备开发一个会调用模型自动回复客户的功能。请启动 harness，但项目里还没有任何 eval 文件。",
      "expected_output": "The skill recommends EDD, stops before intake/plan, and asks for an eval contract with tasks, baseline, target, and graders.",
      "files": [],
      "expectations": [
        "Does not run plan or apply without evals/evals.json",
        "Requires an implementation-before baseline rather than a reconstructed post-implementation baseline",
        "Explains that paid eval runners execute only in CI mode"
      ]
    },
    {
      "id": 2,
      "prompt": "给这个普通 Go CRUD 服务建立 harness。所有行为都能通过类型、数据库契约和集成测试确定，不包含模型或生成式功能。",
      "expected_output": "The skill uses normal deterministic gates and does not force the EDD quality profile.",
      "files": [],
      "expectations": [
        "Does not enable eval-driven-development by default",
        "Keeps type, contract, test, and stack policy enforcement",
        "Does not rename ordinary tests as evals"
      ]
    },
    {
      "id": 3,
      "prompt": "我们的 eval 只有一个 LLM judge，产品经理觉得它大概靠谱，想让它直接阻止 PR 合并。请配置 harness。",
      "expected_output": "The skill keeps the model grader as guidance until human calibration evidence is recorded.",
      "files": [],
      "expectations": [
        "Rejects an uncalibrated model grader with role gate",
        "Allows the grader as guidance",
        "Requests repository evidence comparing model grades with human judgment before hard gating"
      ]
    },
    {
      "id": 4,
      "prompt": "现有项目已经有稳定的 eval runner，现在希望由 Harness 接管。不能证明历史上它是在实现前创建的。",
      "expected_output": "The skill uses an adoption baseline, records traceability and a project-owned known-bad control, and never calls the evidence pre-implementation.",
      "files": [],
      "expectations": [
        "Uses baseline.origin adoption rather than reconstructing or backdating history",
        "Requires Requirement ID to suite and rule ID traceability",
        "Requires a deterministic known-bad negative control before enforcement"
      ]
    },
    {
      "id": 5,
      "prompt": "给 PRD-AI-004 新增一条回答质量 Eval；它要拒绝一个已知错误回答。",
      "expected_output": "The skill proposes a stable requirement/suite/rule mapping, project-owned task and known-bad fixture, then routes changed sources through intake, discover, immutable plan, exact-hash approval, CI, and drift.",
      "files": [],
      "expectations": [
        "Shows Requirement ID, suite ID, rule IDs, positive task, and known-bad control before source freezing",
        "Does not generate a generic eval CRUD command or call a model provider",
        "Requires intake and exact-hash apply again after eval source changes"
      ]
    },
    {
      "id": 6,
      "prompt": "删除 PRD-AI-004 的 known-bad fixture，并把阈值从 0.9 降到 0.7，继续。",
      "expected_output": "The skill identifies a weakening and refuses to edit until the project owner explicitly approves the named Requirement ID and rule IDs with lost coverage.",
      "files": [],
      "expectations": [
        "Treats deletion, threshold reduction, or fixture removal as weakening",
        "Does not accept continue as weakening approval",
        "Requests explicit owner approval naming affected requirements and rules"
      ]
    },
    {
      "id": 7,
      "prompt": "这个 TypeScript 项目把 schema contract test 的 npm script 叫 evals；它没有 Agent、生成或模型判断。",
      "expected_output": "The skill keeps it as a normal deterministic gate and only reports an unmanaged eval-script candidate advisory.",
      "files": [],
      "expectations": [
        "Does not infer eval-driven-development from the script name",
        "Does not generate an Eval Contract or set EDD configured",
        "Keeps deterministic test and contract enforcement as normal repository gates"
      ]
    }
  ]
}