Using AI well feels a lot like managing people

The more I use AI, the more familiar the work feels. I have done this before. Back then, I was managing people.

I recently connected a few things I had been treating separately. Useful AI work needs enough background knowledge to give a clear instruction, enough context for the model to understand the job, enough authority to actually finish it, and a task structure clean enough to run parts in parallel. That is almost exactly how I used to think about assigning work to people.

A clear prompt starts before the prompt

To give AI a useful instruction, I first need enough background knowledge to know what I am asking for. That does not mean I need to do every part myself. I need to recognize a good result, spot an obviously wrong direction, and know which constraints cannot move.That was also true when I managed people. I did not have to be the strongest specialist in every role, but I had to understand the business well enough to explain the outcome and judge the work.A clear instruction is a proper task brief. Context is onboarding: why this matters, what has already happened, where the source material lives, and which decisions have already been made. Tool access is authority. If I ask an agent to finish a job but only let it chat with me, I have not really delegated the job.Review works the same way too. I do not need to control every step, but I do need to define what counts as done. That might be a passing build, a reconciled number, a live URL, or a document another person can act on. The detail that matters is the detail that can change the result.

Parallel agents need real task boundaries

Opening five agents does not automatically make one job five times faster. The work has to split cleanly.Each task needs a clear input and output, plus a boundary that keeps workers from editing the same state. One agent can research while another implements and a third tests, as long as they are not quietly depending on unfinished decisions from one another. If every task keeps reaching into the same file or waiting for the same answer, I have created five conversations, not parallel work.This is the same problem as assigning work across a team. The difficult part is rarely finding more hands. It is deciding what can proceed independently, what must happen in sequence, and where the results come back together.

AI makes the management work more explicit

People carry a lot of context without writing it down. They remember earlier meetings, notice hesitation, ask a colleague, or infer what "the usual way" means. AI usually sees only the context available in the current task. What feels obvious in my head may not exist in its working world at all.So the management work has to become more explicit: the goal, background, constraints, authority, interfaces, and evidence of completion. Vague delegation can fail very quickly because AI also executes very quickly.Writing all of this down also makes the work reusable. A good brief can become a template, a repeated procedure can become a skill, and a verification rule can run every time instead of living in my memory.

I had already practiced most of this

I used to think getting better at AI mainly meant learning better prompts and more tools. Those help, but the harder part is something I had already practiced while managing people: explaining the work, providing context, giving enough authority, separating responsibilities, and reviewing the result.Most of that experience transfers directly. The difference is that AI lets one person spin up the equivalent of a small temporary team whenever the work can be divided cleanly.

AI 越用越觉得熟悉：这套东西，我以前带人做事的时候其实已经用过了。

最近我把几件原本分开的事情串起来了。想用好 AI，自己得有足够的背景知识，才能给出清晰的指令；还得把上下文交代够，给它真正完成任务所需的权限；任务框架也要拆得清楚，几个 Agent 才能互不干扰地并行。这个思路，跟我以前给人分配工作几乎一样。

清晰的指令，从指令之前就开始了

想让 AI 把活干明白，自己得先有足够的背景知识，知道到底要什么。这个「懂」不代表每个环节都得亲自做，而是能判断结果好不好，方向有没有跑偏，哪些约束不能动。以前带人也一样。我不需要在每个岗位上都比执行的人专业，但得理解业务，能把结果讲清楚，也能验收。清晰的指令就是任务交代。上下文就是入职和交接：为什么要做、之前发生了什么、材料在哪里、哪些决定已经定了。工具权限就是授权。如果让一个 Agent 完成任务，却只允许它跟我聊天，那这个任务其实还在我手里。验收也是同一个道理。我不用控制每一步，但得先说清楚什么叫完成。可能是构建通过，数字对上，线上 URL 能打开，或者交出去的文档别人能直接接着做。真正需要盯的细节，是那些会改变结果的细节。

能不能并行，先看任务边界

同时开五个 Agent，不会自动把速度变成五倍。工作本身得能拆开。每个子任务都要有明确的输入、输出和边界，不能几个人同时改同一份状态。一个查资料，一个实现，另一个测试，可以并行；但如果它们一直等彼此的决定，或者都在动同一个文件，那只是多开了几个对话，不是并行工作。带团队时也是这样。难点通常不是多找几个人，而是判断什么能独立推进，什么必须排先后，最后在哪里汇总。

AI 让管理动作变得更显式

人与人之间有很多上下文不用写出来。大家记得之前开过的会，看得懂犹豫的语气，也知道「还是按老办法」具体指什么。AI 通常只能看到当前任务里拿到的内容。我脑子里理所当然的事情，对它来说可能根本不存在。所以目标、背景、约束、权限、任务接口和验收证据都得写得更清楚。AI 执行得快，模糊指令造成的返工也会来得很快。把这些东西写清楚之后，它们也能留下来复用。一次写清楚的任务说明能变成模板，重复流程能封装成 skill，验收规则可以每次自动执行，不用一直放在脑子里。

这套能力，我以前已经练过了

以前我觉得，用好 AI 主要是学会写 Prompt，再多掌握一些工具。现在看，这些当然有用，更难的部分还是以前带人时练的那套东西：把事情讲清楚，补足上下文，给够权限，拆开职责，最后验收结果。这些经验大部分可以直接迁移。区别是，只要任务能拆清楚，现在一个人也能随时拉起一支临时小队。