Jev 提示词与问题示例

每个问题包括：由小写字母、数字或下划线组成的 ID，类型，8 到 600 字符的 instructions，以及 criteria。Choice 的 criteria 把 2 到 12 个选项 ID 映射到定义；Score 的 criteria 按从低到高列出 2 到 10 个等级；Noul 的 criteria 说明 true 和 false 各代表什么。criteria 要写成可观察的事实，而不是对模型的指令。

只有一个选项成立时用 Choice。选项之间要互斥，并加上 other，避免模型被迫选错团队。代码读取选中的选项及概率，把低概率的情况交给人工。

{
  "state": "My invoice shows two charges for the same month.",
  "questions": {
    "queue": {
      "type": "choice",
      "instructions": "Which team should handle this customer message?",
      "criteria": {
        "billing": "Charges, refunds, invoices or payment methods.",
        "technical": "Bugs, errors, outages or integrations.",
        "sales": "Plans, quotes, upgrades or new seats.",
        "other": "Anything that fits none of the teams above."
      }
    }
  }
}

有序判断用 Score。每个等级写一条标准，从低到高排列，并让每级明显强于上一级。用 JSON state 可以传入多个字段。返回值可能带小数，请在代码里规定各区间的分界。

{
  "state": {
    "company_size": "200 employees",
    "message": "We need SSO and a quote before our budget closes Friday."
  },
  "questions": {
    "priority": {
      "type": "score",
      "instructions": "How sales-ready is this lead?",
      "criteria": [
        "No buying intent or a poor fit.",
        "Some interest but no timeline.",
        "Clear need with a vague timeline.",
        "Clear need, budget and a near deadline."
      ]
    }
  }
}

单个是非命题用 Noul，返回命题成立的概率（0 到 1）。根据漏判代价设定阈值：安全类升级用较低阈值，可选动作用较高阈值。

{
  "state": "This is the third time I ask. Cancel my account now or I report you.",
  "questions": {
    "escalate": {
      "type": "noul",
      "instructions": "Should a senior agent handle this message today?",
      "criteria": {
        "true": "The customer is upset, repeating a request or threatening action.",
        "false": "A routine reply from the normal queue is enough."
      }
    }
  }
}

同一份 state 可以在一次请求里提出多个问题。这里用 Choice 给评论打标签，用 Noul 判断是否需要人工查看。组合起来，代码就能自动处理明确的标签、把其余排队，而无需第二次调用。

{
  "state": "Great post! Check my profile for cheap followers.",
  "questions": {
    "label": {
      "type": "choice",
      "instructions": "Classify this comment for moderation.",
      "criteria": {
        "ok": "On-topic and acceptable.",
        "spam": "Promotion, links or unrelated offers.",
        "abusive": "Insults, threats or harassment."
      }
    },
    "review": {
      "type": "noul",
      "instructions": "Is a human moderator needed for this comment?",
      "criteria": {
        "true": "The case is unclear or could cause harm.",
        "false": "The label is clear enough to apply automatically."
      }
    }
  }
}

每个问题只问一个判断。按 state 中能看到的内容定义选项，而不是按你希望发生的结果。instructions 保持简短，细节写进 criteria。避免“退款”和“账单”这类重叠选项，除非写清边界。自动化之前，用清晰、模糊和对抗性案例测试每个问题，并像代码一样给问题做版本管理。