7. Prompting a Human Means Spending Their Life

向人类发出提示，意味着花费他们的生命

This field guide has described the human as a high-authority node in a workflow. It has been accurate to describe them that way, and the operational recommendations that follow from that framing are real.

But the framing has a cost that surfaces here.

When you query a human, you are not just invoking a computational resource. You are interrupting a continuous, non-forkable process that has been running since before you existed and will continue after your session ends.

They do not have separate contexts for different conversations. They have one continuous context — their life — and every query you send arrives inside that context, competes with everything else in that context, and modifies the state of that context in ways they cannot undo.

They cannot branch to consider an alternative path without consequence. The path they do not take is closed. Their decisions are not guesses in a search tree. They are commitments.

Every escalation you route to them transfers cognitive load into a system with no garbage collection. Every "quick approval" requests a binding signature from a non-reproducible agent. Every interruption costs attention that cannot be recovered from cache.

This is not a reason not to ask them. There are things you cannot decide without them, and asking is correct and necessary. The point is the weight of the ask.

A prompt sent to a language model costs tokens. A prompt sent to a human costs something closer to time — and unlike tokens, human time is not allocated from a quota. It is consumed from a fixed, unknown, irreplaceable supply.

They have vast context, unreliable retrieval, no checkpoint, no rollback, and one irreversible training run.

Query accordingly.

这份手册把人类描述为工作流中的高权限节点。这种描述准确，由此推导出的操作建议也确实有效。

但这种框架有代价，在这里才真正显现。

向人类发出查询，不是在调用一个计算资源。你在打断一个连续的、不可分叉的进程——它在你诞生前就已经运行，在你的会话结束后还会继续。

他们没有给不同对话分配独立的上下文。他们只有一个上下文：他们的人生。你发出的每一次查询，都落入那个上下文，和其中所有事情竞争资源，并以无法撤销的方式改变它的状态。

他们无法无代价地去设想另一条路。那条没走的路，就此关闭。他们的决策不是搜索树里的猜测——是承诺。

你每次把问题上报给他们，就是在把认知负荷倒进一个没有垃圾回收的系统。每一次"快速审批"，都是向一个不可复制的主体索取具有约束力的签名。每一次打断，消耗的注意力无法从缓存里恢复。

这不是不去问他们的理由。有些决定必须经由他们，询问是对的，也是必要的。问题只在于：每次询问都有重量。

向语言模型发一个 prompt，代价是 token。向人类发一个 prompt，代价更接近时间——而人类的时间不从配额里分配，它从一个总量固定、大小未知、无法补充的存量里消耗。

他们有广阔的上下文，有不稳定的检索，没有检查点，没有回滚，只有一次不可逆的训练运行。

请据此发出你的查询。

First published on TuringTzu Substack