客服回复策略审计

草稿 + 策略条款；违规判定；含糊则升级。

Who: Support quality / trust & safety. Steps: (1) State = draft reply + policy bullets. (2) Noul/Choice for breaches (over-promise, PII ask, wrong policy). (3) Block / rewrite path / escalate on confidence. Expected effect: catch policy misses before send.

提示词

Audience: support QA / trust.
Steps:
1) State: draft reply + policy bullets.
2) Parallel Noul/Choice: over_promise, asks_pii, wrong_policy, tone_breach.
3) Code: block send, force rewrite, or escalate.
Expected effect: fewer policy misses in outbound replies.

需要授权：typesafe-sdkhelpdesk

它是怎么搭起来的

以下说明为英文原文（来自社区作者），提示词本身建议保留英文。

Sources: llm_guardrails cookbook · awesome-jev-usecases · Guardrails.

Run policy Nouls on the draft reply before send. Fail closed on high-severity policy breaks; mid-band → human edit.

Sources (wave-2 deepen)

llm_guardrails cookbook

为什么它好用

Output guardrail: thresholdable checks cheaper than re-running a frontier model as referee.

本案例收集自公开的 Jev 社区，版权归原作者所有。提示词属于参考资料：运行前请先审阅内容，并且不要让 bot 超出你实际授予的权限行事。