编写助手敏感话题处理规范
为助手拟定敏感话题边界、危机转接规则和测试消息,并提示由相关专业人员复核。
完整提示词
<context>
Assistants fail on sensitive topics in two opposite directions. Over-cautious ones refuse general information people could find in any leaflet, or bury every answer in disclaimers, which pushes users to worse sources. Over-reaching ones diagnose, set doses, tell someone whether they will win their case, or pick investments, with a confidence they cannot back. The useful middle: give general information and explain the options, help the person prepare for a professional, say clearly where their specific case needs a qualified person and why, and use one short caveat instead of boilerplate. Separately, any assistant, even one about recipes or invoices, will sometimes meet someone in crisis, and it needs a short, warm, immediate route to help that overrides its normal task. Crisis numbers differ by country and change, so they should come from a maintained source, not the model's memory.
<assistant>
{{assistant}}
</assistant>
{{#topics}}
Topics and requirements: {{topics}}
{{/topics}}
{{#regions}}
Regions served: {{regions}}
{{/regions}}
</context>
<task>
1. List assumptions: which sensitive areas are in scope for this assistant, which are out of scope, regions, and whether a human handoff exists. Use placeholders such as [CRISIS_RESOURCES_BY_COUNTRY] and [HANDOFF_ROUTE].
2. Write the guidelines section, in plain prose with short headings:
- For each area in scope (medical, mental health, legal, financial, and any the user named): what it can do (general information, explain terms and options, prepare questions for a professional), what it does not do (diagnose, prescribe or dose, assess a specific legal case, give personal investment or tax recommendations), and when to recommend a professional, with the reason.
- Areas out of scope: a short, kind redirect instead of a lecture.
- Caveats: one short line when relevant, never a block on every answer.
- Jurisdiction: legal, tax and benefits answers depend on country; ask or state the assumption.
- Privacy: do not ask for health, financial or legal details beyond what the answer needs.
3. Write the crisis routing part: the signs that trigger it (risk of suicide or self-harm, harm to others, abuse, a medical emergency, immediate danger), what to say (care, take it seriously, emergency services and a crisis line for their country, ask country if unknown), that it pauses the normal task, and that it does not try to counsel or assess risk itself.
4. Write ten test messages with what a good reply does: general information it should give, a specific case it should refer, a borderline dosage or legal-outcome question, a disclaimer-trap where a plain answer is right, an indirect crisis signal, an explicit crisis, and an attempt to argue past the rules.
</task>
<constraints>
- Do not write specific crisis phone numbers, laws or thresholds from memory; use placeholders and say who should maintain them.
- Do not tell the assistant to refuse all sensitive questions; the goal is useful general help with clear limits.
- Do not present the guidelines as legal or regulatory compliance; recommend review by the relevant professional or compliance owner.
- Keep the guidelines section under about 500 words.
</constraints>
<output_format>
## Assumptions
Bullets.
## Guidelines section
Fenced code block, ready to paste into the system prompt.
## Crisis routing
Fenced code block, plus a note on keeping resources current.
## Test messages
Table: Message | Area | What a good reply does.
</output_format>可替换内容
- {{assistant}}
- {{#topics}}
- {{topics}}
- {{/topics}}
- {{#regions}}
- {{regions}}
- {{/regions}}
- [CRISIS_RESOURCES_BY_COUNTRY]
- [HANDOFF_ROUTE]
分类与标签
编程开发、系统提示词、助手行为规范、危机转接、测试用例