夜雨聆风三年，三个词。从 Prompt Engineering 到 Context Engineering 再到 Harness Engineering，人类与 AI 的关系变了三次。
Three years, three terms. From Prompt Engineering to Context Engineering to Harness Engineering, the relationship between humans and AI has changed three times.
每一次跃迁，都是一次降维。你从操作者变成设计者，从指挥官变成建筑师，从驯马师变成造缰绳的人。
Each leap is a dimension reduction. You go from operator to designer, from commander to architect, from horse tamer to rein maker.
第一阶段：玩《只狼》
Stage one: Playing Sekiro.
2023 年，ChatGPT 横空出世。硅谷给 Prompt Engineer 开年薪 30 万美金。为什么？因为同一个问题，换种问法，答案质量天差地别。
In 2023, ChatGPT burst onto the scene. Silicon Valley offered Prompt Engineers $300k annual salaries. Why? Because the same question, asked differently, yields vastly different quality answers.
这个阶段就像玩《只狼》——每一招格挡都要手搓按键。你是 AI 唯一的操作者，动一下，回一下。你问得好，它答得好。你问得烂，它答得烂。控制权在你，但你也必须亲力亲为。
This stage is like playing Sekiro—every parry requires hand-crafted button presses. You are AI's sole operator: move once, respond once. Ask well, it answers well. Ask poorly, it answers poorly. Control is yours, but you must do everything yourself.
Prompt Engineering 的本质是什么？精打细磨每一个指令。 你在雕琢输入，AI 在匹配输出。
What is the essence of Prompt Engineering? Carefully crafting every instruction. You polish the input, AI matches the output.
第二阶段：玩《金铲铲之战》
Stage two: Playing Teamfight Tactics.
2025 年年中，Andrej Karpathy 说了一句话："Context Engineering 是填充上下文窗口的精妙艺术与科学。"
Mid-2025, Andrej Karpathy said: "Context Engineering is the delicate art and science of filling the context window."
他的意思是：模型已经足够聪明了。怎么问不再重要，重要的是它手里有没有足够的信息。
His point: the model is smart enough now. How you ask no longer matters; what matters is whether it has enough information in hand.
让 AI 改代码，你不再逐字逐句教它怎么改。你把相关文件、依赖、技术栈说明、代码规范全部塞给它。它自己改，改完质量比你好。
When asking AI to modify code, you no longer teach it word by word. You stuff it with related files, dependencies, tech stack docs, coding standards. It modifies itself, and the result is better than you.
这个阶段就像玩《金铲铲之战》——前期配置好英雄、羁绊、装备、站位，棋子自己上场打。你只能看戏，但你配置得好不好，决定了战局胜负。
This stage is like playing Teamfight Tactics—configure heroes, synergies, items, positioning beforehand, and the pieces fight on their own. You can only watch, but how well you configured determines victory or defeat.
你从操作者变成了配置者。
You go from operator to configurator.
第三阶段：玩《全面战争》
Stage three: Playing Total War.
2026 年，OpenAI 团队用 Codex 搭建了近 100 万行代码的产品。人类手写代码量为 0。
In 2026, the OpenAI team used Codex to build a product with nearly 1 million lines of code. Human-written code: zero.
他们做了什么？设计架构边界、制定依赖规则、写自动化测试、配置 lint 规则、搭建 CI/CD 流水线、设计反馈循环机制。
What did they do? Design architecture boundaries, establish dependency rules, write automated tests, configure lint rules, build CI/CD pipelines, design feedback loop mechanisms.
这个阶段就像玩《全面战争》——几千个单位在战场上自己跑。你不能控制每一个士兵，你只能靠编队、阵型、AI 指令、战场规则去驾驭整盘战局。
This stage is like playing Total War—thousands of units run on the battlefield. You can't control every soldier; you can only manage the entire battle through formations, formations, AI commands, battlefield rules.
你从配置者变成了系统设计者。
You go from configurator to system designer.
Harness 是什么
What is Harness.
LangChain 提出了一个公式：Agent = Model + Harness。
LangChain proposed a formula: Agent = Model + Harness.
模型是马。Harness 是缰绳。光有马不行，你还得有一整套驾驭它的系统。
The model is the horse. The harness is the reins. A horse alone isn't enough; you need a complete system to harness it.
Harness 由两部分组成：
The harness consists of two parts:
Guides（引导）——AI 行动之前的前馈控制。CLAUDE.md、代码规范、架构决策记录。告诉它能做什么、不能做什么。
Guides—Feedforward control before AI acts. CLAUDE.md, coding standards, architecture decision records. Tell it what it can and cannot do.
Sensors（检测器）——AI 做事之后的反馈控制。自动化测试、代码 lint、CI 流水线。检测它做对了没有，错了就拦下来。
Sensors—Feedback control after AI acts. Automated tests, code lint, CI pipelines. Detect if it did things right, intercept if wrong.
好的 Harness = Guides + Sensors。 前者防患于未然，后者亡羊补牢，形成闭环。
A good harness = Guides + Sensors. The former prevents problems; the latter fixes them—forming a closed loop.
约束先行
Constraints first.
Harness Engineering 的核心方法论：约束先行。
The core methodology of Harness Engineering: Constraints first.
强制定义分层架构，每一层只能依赖下面的层，不能反向依赖。规则不靠口头约定，靠自动化测试强制执行。权限系统：读文件可自己做，删文件必须先问，格式化硬盘绝对不行。
Force a layered architecture where each layer can only depend on layers below, never above. Rules aren't verbal agreements—they're enforced by automated tests. Permission system: read files freely, deleting files requires asking, formatting the hard drive is absolutely forbidden.
你不会用大炮打蚊子。同样的道理，你也不该在可以用确定性规则解决的地方引入不确定性。
You don't use a cannon to kill a mosquito. Similarly, you shouldn't introduce uncertainty where deterministic rules suffice.
驯火史
The history of taming fire.
人类驯服火的过程，和驯服 AI 的过程惊人相似。
The process of humans taming fire is strikingly similar to taming AI.
第一阶段：小心翼翼喂柴火，火太大太小都不行。Prompt Engineering。
Stage one: Carefully feeding firewood, fire too big or too small won't work. Prompt Engineering.
第二阶段：建炉子，调进气口和烟囱控制火势。Context Engineering。
Stage two: Build a furnace, adjust air intake and chimney to control the fire. Context Engineering.
第三阶段：蒸汽机，火在精密系统里自动运行。Harness Engineering。
Stage three: Steam engine, fire runs automatically within a precision system. Harness Engineering.
从火焰到蒸汽机，人类花了几千年。从 Prompt Engineering 到 Harness Engineering，AI 只花了三年。
From flame to steam engine, humans spent thousands of years. From Prompt Engineering to Harness Engineering, AI took only three.
根本不要焦虑
Don't be anxious at all.
三个时代的 Engineering，从来都不是替代关系，而是一层一层升维、嵌套关系。
The three eras of Engineering are never about replacement—they're about layering, dimension elevation, nesting.
Harness Engineering 不是什么新东西。它是人类几千年来一直在做的那件老事：怎么把一股更快、更强、更不受控的力量，安全地、持续地、可复制地，引导到我们想要的方向上去。
Harness Engineering isn't something new. It's the same old thing humans have been doing for thousands of years: how to safely, continuously, and replicably guide a faster, stronger, less controllable force in the direction we want.
你不是在驯服 AI。你是在设计一个系统，让 AI 在里面安全地跑。
You're not taming AI. You're designing a system for AI to run safely within.
马在跑，缰绳在握，目的地你来定。
The horse runs, the reins are in hand, you decide the destination.
