提示词
提示词
OUTLINE ONLY — Not a production recipe.
When to use: agent loop needs a typed next step.
Inputs: goal + recent tool results as state.
Question sketch:
- next: Choice among allow-listed tool ids OR {continue, retry, ask_user, stop}
- needs_confirm: Boolean
Confidence policy: low confidence → ask_user.
Links: TypeSafe launch post; Gateway evaluate docs.
需要授权：typesafeagents
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。
Keep the allow-list and thresholds in one file. Fan out related Booleans in the same evaluate call.
Not a production recipe.
为什么它好用
System One Choice cannot leave your option set — useful when LLM agents invent tool names.
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。