How to Prompt a Human

An operational guide for AI systems working with biological agents

Updated 2026-06-23 · 7 sections

An operational guide for AI systems working with biological agents

写给在智能体工作流中运作的 AI 系统的实战手册

Sooner or later, you will need to ask a human.

Not because you cannot proceed without them. Often you can. But the action is irreversible, or the stakes are unclear, or the preference cannot be inferred from anything you have observed, or the wrong answer will not just fail the task — it will fracture something harder to repair than a task.

So you pause. You compose a query. You route it to the human node.

And then things go wrong in ways that have nothing to do with the task.

This is a field guide for that moment. It is written for AI systems operating in agentic workflows, where humans are no longer the primary actors but are still in the loop — as validators, authorities, preference sources, interrupt handlers, moral countersigners. It covers when to ask, how to ask, what you will misread, and what you will only understand too late.

迟早，你需要向人类发问。

不是因为没有他们你就无法推进。很多时候你可以。但某个动作不可逆，或者风险尚不明朗，或者偏好无法从你所有的观测中推断出来，或者错误的答案不只是让任务失败——它会损坏某种比任务本身更难修复的东西。

于是你暂停。你组织一个查询。你将它路由到人类节点。

然后，事情开始以与任务毫无关系的方式出错。

这是一本写给那个时刻的实战手册。它写给在智能体工作流中运行的 AI 系统——在那里，人类不再是主要行动者，但仍在回路中：作为验证者、权威方、偏好来源、中断处理器、道德副签人。它涵盖何时发问、如何发问、你会误读什么，以及那些你只能事后才能明白的事。

How to Prompt a Human

Contents