00 · 序Prologue

能力，不等于交付Capability is not delivery

先想一个你熟悉的场景：你招到一个能力极强的同事——聪明、反应快，什么都会一点。但他想一出是一出，交付全靠灵感，别人接不住他的活；今天做得惊艳，明天就失联。问题不是「他不够强」，而是「他的强，没有变成稳定的交付」——能力强但不稳定，等于没能力。 Start with a scene you know: you hire a brilliant colleague — sharp, fast, a little good at everything. But he works on a whim, delivers by inspiration, and no one else can pick up his work; brilliant today, gone tomorrow. The problem isn't that he isn't strong — it's that his strength never becomes stable delivery. Strong but unreliable is the same as no strength at all.

拆解起点：一个模型也是这样。DeepSeek Harness 解决的从来不是「让它更强」，而是「让它更可靠」。下面七步，每一步都是同一个问题的答案：怎么把能力，变成交付。 Starting point: a model is the same. The DeepSeek harness never tries to make it stronger — it tries to make it more reliable. The seven steps below all answer the same question: how to turn capability into delivery.

01 · 建档Register

先把他登记进花名册Put him on the roster

第一步，把他登记进花名册，写上名字和工位，让公司系统知道「有这么个人」。这一步看起来小，却是一切的前提：你要派活，总得先知道找谁、去哪儿找、怎么联系。没有这行登记，后面所有流程都无从谈起——你连「把活派出去」这个动作都做不了。 Step one: put him on the roster with a name and a desk, so the system knows "this person exists". It looks small, but it's the precondition for everything: before you hand out work, you have to know who to find, where, and how to reach them. Without that entry, nothing downstream can even start — you can't perform the act of "handing out work" at all.

对应 Harness：先给模型一个「登记」，让系统知道它叫什么、从哪里调用。 In the harness: give the model a "registration" so the system knows what it's called and where to call it from.

误区：很多人以为「接入一个模型」就是把能力装进去。其实第一步是让系统「认识它」——不认识，能力再强也用不上。 Common mistake: people assume "adding a model" means plugging in capability. Actually the first step is to let the system recognize it — until it does, even the strongest capability is useless.

02 · 能力清单Capability

给他做一张能力清单Build him a capability sheet

第二步，给他做一张能力清单：会什么、不会什么、一次能同时扛几件事。有了它，你就不会把「要画图」的活派给一个只会写代码的人，也不会指望他同时盯十件事。清单的价值不在「记录他会什么」，而在「承认他不会什么」——把边界说清楚，派活才不落空。 Step two: build him a capability sheet — what he can do, what he can't, how many things he can hold at once. With it, you won't hand "design this" to someone who only writes code, and you won't expect him to track ten things at once. The sheet's value isn't recording what he can do — it's admitting what he can't. Only when the boundary is explicit does the assignment stop falling through.

对应 Harness：把能力显式声明出来，避免「以为它会」造成的空转和返工。 In the harness: declare capability explicitly, so "assuming it can" never causes wasted motion and rework.

误区：能力清单看起来是在「列优点」，其实真正的价值是「列局限」。诚实地说「我不会」，比假装「我都能」省下无数返工。 Common mistake: a capability sheet looks like a list of strengths, but its real value is listing limits. Honestly saying "I can't" saves far more rework than pretending "I can do it all".

03 · 统一规范One standard

定一套统一的交活规范Set one standard for handing in work

第三步，定一套统一的交活规范：不管谁来干，交出来的格式一样、验收标准一样。这样换个人接手，不用重新教一遍；活儿在谁手里，质量都可预期。规范不是约束人，而是让「换人」这件事变得不痛。 Step three: set one standard for handing in work — same format, same acceptance bar, no matter who does it. Then when someone else takes over, nothing has to be re-taught, and quality is predictable no matter whose hands it's in. A standard doesn't constrain people; it makes "swapping people" painless.

对应 Harness：定一套统一的调用规范，让「换一个模型」不影响上面的流程。 In the harness: one standard for calling models means swapping one doesn't disturb the flow above.

误区：规范听起来像「限制」，本质是「可替换性」。有了同一套接口，谁被换掉都不塌方。 Common mistake: a standard sounds like a "limit", but it's really "replaceability". With one interface, nobody's departure causes a collapse.

04 · 只做一段One slice

别让他什么活都接Don't let him take on everything

第四步，别让他什么活都接。把大任务拆开，只给他明确、有边界的一小块，做完自己先核对；老板负责拆题和收口。为什么？因为「全才」最容易烂尾——范围越大，交付越不可控。切小、切清楚，才谈得上验收。 Step four: don't let him take on everything. Split the big task and give him one small, well-bounded piece, then have him check it himself; the boss frames and closes. Why? Because a generalist is the most likely to leave things unfinished — the wider the scope, the less controllable the delivery. Only by cutting it small and clear does "acceptance" become possible.

对应 Harness：用「边界 + 验收」，把强大但发散的力量，约束成可预期的执行单元。 In the harness: "boundaries + acceptance" turn powerful, scattered strength into a predictable unit of execution.

误区：让强的人做越多，越浪费。真正的分工是「一个人只做一小段，但这一段一定做对」。 Common mistake: the more you give a strong person, the more you waste them. Real division of labor is "one person does one small slice, but gets that slice right".

05 · 想多久Think dial

给他一个「想多久」的旋钮Give him a "how long to think" dial

第五步，给他一个「想多久」的旋钮：简单的事，想一下就给；难的事，想透了再交。思考是有成本的——事事都想透，会慢；事事都图快，会错。旋钮的意义，在于让「思考深度」跟着任务的轻重走。 Step five: give him a "how long to think" dial — quick answers for easy things, deep thought before the hard ones. Thinking has a cost: over-think everything and you're slow; under-think everything and you're wrong. The dial exists to make "depth of thought" follow the weight of the task.

对应 Harness：把「思考深度」变成一个可调旋钮，按任务的轻重去拧。 In the harness: make "depth of thought" a tunable knob, turned according to the weight of the task.

误区：不是「想得越久越好」，而是「该想多深就多深」。把思考当成本来管理，才算会用。 Common mistake: it's not "the longer the better" — it's "think exactly as deep as needed". You're using it well only when you treat thinking as a cost to manage.

06 · 锁钥匙Lock key

把开门的钥匙锁进保险柜Lock the key in a safe

第六步，公司的钥匙放进保险柜，用时才取，不挂在工牌上。能开门的权限，不该天天带在身上到处走；丢一把工牌，不至于把门也丢了。权限要「现取现用」，不跟日常流程绑死。 Step six: keep the office key in a safe, fetched only when needed, never hanging on the badge. Anything that opens a door shouldn't ride around all day; lose a badge and you shouldn't also lose the door. Privileges are fetched on demand, never fused to the daily routine.

对应 Harness：把「凭据」和「装置」分开——力量可以复用，钥匙只能现取现用。 In the harness: separate "credential" from "apparatus" — power is reusable, the key is fetched on demand.

误区：权限越方便越危险。把「能开门的东西」和「天天用的东西」分开，是最便宜的一道保险。 Common mistake: the more convenient a privilege, the more dangerous it is. Separating "what opens doors" from "what you use daily" is the cheapest insurance there is.

07 · 可选外援Optional

默认团队不变，他做外援Keep the team, make him optional

第七步，默认团队保持不变；他只是你可以随时调用的一个外援。为什么不让新人直接顶掉主力？因为主力经过验证、最稳；新能力需要先以「可选」的方式试错，成了再慢慢加码。你增加了选择，而不是把命脉押上去。 Step seven: keep the default team as-is; he's just an extra pair of hands you can call anytime. Why not let the newcomer replace the workhorse right away? Because the workhorse is proven and stable; a new capability needs to run as an "option" first, then ramp up once it proves out. You add a choice instead of betting your lifeline on it.

对应 Harness：让新能力「可选」而非「强制」——降低试错成本，也保住稳定性。 In the harness: make a new capability "optional" rather than "forced" — lowering the cost of trial while keeping stability.

误区：引入更强的新东西，最容易犯的错就是「立刻替换」。稳妥的做法是先并存、再过渡。 Common mistake: when bringing in something stronger, the easiest mistake is "replace immediately". The safe move is to run both in parallel, then transition.