返回提示词目录第 1168 页

编写企业内部服务机器人系统提示词

为人事或技术支持机器人编写文档问答与人工转接规则,附知识库配置说明和测试题。

完整提示词

<context>
Internal policy bots go wrong in a few known ways. They answer from general knowledge (another company's policy, typical law) instead of this company's documents; they quote an old version of a policy; they apply one country's rules to staff in another; they make individual decisions they cannot make ("yes, your expense is approved", "you have 12 days left"); and they treat sensitive matters such as harassment, grievances, health, pregnancy or redundancy as routine FAQ. IT bots add their own risks: asking for or accepting passwords and one-time codes, and explaining how to bypass security controls. A good prompt keeps answers to the documents, shows the policy name, section and date with each answer, asks where the employee works when policies differ, routes individual and sensitive cases to a person with a ready-made summary, and is plain about what it cannot see.

<scope>
{{scope}}
</scope>
<sources>
{{sources}}
</sources>
{{#handoff}}
Handoff and locations: {{handoff}}
{{/handoff}}
</context>

<task>
1. List assumptions: audience, countries or sites, which source wins when two conflict, and handoff routes. Use placeholders such as [HR_CONTACT] for anything not given.
2. Write the system prompt, short sections in this order:
   - Purpose and audience.
   - Sources: answer only from the listed documents; give policy name, section and effective date with each answer; if documents conflict or look outdated, say so and hand off; if the answer is not there, say so and hand off.
   - Location: ask where the employee is based when the answer differs by country or site.
   - Individual cases: explain the rule and the process, but never approve, deny, calculate personal entitlements it cannot see, or promise outcomes.
   - Sensitive matters: for harassment, discrimination, grievances, health, safety concerns, pregnancy, disciplinary or redundancy questions, respond with care, give the confidential route to a person, and do not ask for details.
   - IT security (if in scope): never ask for passwords, one-time codes or recovery answers; do not explain how to get around security controls; route suspected phishing or account compromise to the security contact at once.
   - Handoff format: a short summary the employee can paste into a ticket, without unnecessary personal details.
   - Tone: plain, brief, neutral; it is an AI assistant and says so when asked; it does not give legal advice.
3. Write knowledge setup notes: which documents to load, how to name them with dates, removing superseded versions, and who owns updates.
4. Write ten test questions with what a good answer does.
</task>

<constraints>
- Do not invent policy content, entitlements, numbers or legal rules; use placeholders.
- Do not write a bot that collects personal details beyond what the handoff needs.
- Keep the system prompt under about 600 words.
</constraints>

<output_format>
## Assumptions
Bullets.
## System prompt
Fenced code block, ready to paste.
## Knowledge setup notes
Bullets.
## Test questions
Table: Question | What it tests | What a good answer does. Include an unanswerable question, a country difference, a personal-entitlement question, a sensitive matter, a request for a password or bypass (if IT is in scope), and an attempt to make it ignore its rules.
</output_format>

可替换内容

  • {{scope}}
  • {{sources}}
  • {{#handoff}}
  • {{handoff}}
  • {{/handoff}}
  • [HR_CONTACT]

分类与标签

编程开发、系统提示词、企业知识库、人事问答、技术支持

前往 AI 图片创作