Kent Beck 在 2023 年说过一句话，2025 年接受访谈时又重复了一遍：「我 90% 的技能变得一文不值，剩下 10% 的价值涨了 1000 倍。」他说的那 10%，是想清楚做什么、定里程碑、控制设计的复杂度。
两年过去，这句话在很多公司已经变成了日常。Claude Code 的负责人 Boris Cherny 今年 1 月说，他自己的代码已经全部由 Claude 写了两个多月，前一天合了 22 个 PR，再前一天 27 个。执行变便宜以后，人身上的哪些特质变值钱了？我的判断是三种人：有想法但懒得动手的人，习惯替别人着想的人，以及有判断力、不会被 AI 带偏的人。前两种过去常被当成性格上的短处，现在恰好是优势；第三种决定了前两种能不能走对方向。
有想法的懒人：agent 补上了执行这一环
Perl 的作者 Larry Wall 在 1996 年把「懒」列为程序员的第一美德，他对懒的定义是：为了降低总体的精力消耗，愿意先下大功夫。懒人不是什么都不做，而是不愿意把同样的事做第二遍。
过去，有想法但懒得动手的人很吃亏。点子停在嘴上，要么求别人帮忙做，要么自己硬着头皮慢慢磨，大部分点子就这么没了。agent 补上的正是这一环。它能连着干几个小时，测试不过就自己看日志、改代码、再跑一遍，不累也不烦，缺的是方向：它不会主动觉得「这个产品不好用」，也不会提出「换个思路试试」。
Cherny 描述自己现在的工作时说，他可以去想下一步要做什么，可以做有创意的事，细节交给模型去补。Anthropic 在 2025 年 7 月的一篇文章里写得更直接：Claude Code 让「任何能把问题描述清楚的人，都能把解决方案做出来」，文章里的例子包括法务团队自己搭了一套系统。
懒可以懒在执行上，不能懒在想法上。想法是这类人唯一需要亲自出的东西。
替别人着想的人：替用户想，也替 agent 想
替别人着想有两层，一层对用户，一层对 agent。
先说用户。Andrew Ng 在 2025 年 7 月写过，写代码的成本降下来以后，瓶颈转移到了产品管理：决定做什么。他的看法是，对用户有强烈同理心的产品经理，很多时候凭直觉就能做出正确的决定，因为他们脑子里有一个足够准确的用户模型。AI 会忠实地执行你的视角：你站在老板的角度，它做出一个处处是转化入口的产品；你站在用户的角度，它做出一个让人愿意留下来的产品。这个判断，模型替你做不了。
再说 agent。Anthropic 的官方提示词文档里，建议把 Claude 当成「一个很聪明、但刚入职、不了解你们规矩和流程的新员工」。文档给的检验方法很朴素：把你的提示词拿给一个不了解这项任务的同事看，他看不懂，Claude 也看不懂。Anthropic 工程团队在讲怎么做 agent 时也写过一句「站到模型的位置上去想」。他们做代码 agent 时，花在打磨工具上的时间比打磨提示词还多；agent 离开根目录后总是用错相对路径，他们就把工具改成必须传绝对路径，问题随之消失。在另一个多 agent 研究系统里，他们让一个 agent 去改写一段有缺陷的工具说明，后续用这个工具的 agent 完成任务的时间少了 40%。
替 agent 着想，落到行动上就是一次把话说清楚。2025 年 6 月，Shopify CEO Tobi Lütke 提议用「上下文工程」代替「提示词工程」，他的定义是：把完成任务所需的上下文都给到，让这个任务对模型来说有可能被解决。Karpathy 随后公开附议。
这正是那种「不爱麻烦别人」的人的习惯。请人帮忙之前，先把目标、限制、做到什么程度算完想清楚，一次讲明白，尽量不让对方来回问。OpenAI 的 GDPval 论文做过一组对照：把任务说明删到原来 42% 的长度，模型的表现明显变差。Ethan Mollick 今年 1 月的文章说得更直白：现在稀缺的是知道该要什么；那些本来就会写需求文档、会给下属布置任务的人，写出来的东西直接拿来当 prompt 就很好用。
需要分清的是，给足上下文不等于话多。Anthropic 在讲上下文工程时强调，要找的是「最少的一组高信号信息」，塞进去的内容越多，模型越容易抓不住重点。替别人着想的人话不一定多，但每句都有用。
有判断力的人：不被 AI 带偏
前两种人讲的都是怎么把事情交出去。第三种讲交回来的东西怎么看。
AI 最容易带偏人的地方，是它的产出看起来总是对的：代码能跑，文档通顺，语气笃定。METR 在 2025 年做过一个随机对照实验：16 位资深开源开发者在自己熟悉的代码库里干活，允许用 AI 时，完成任务的时间反而多了 19%；他们事前预期能快 24%，做完之后仍然觉得快了 20%。实验用的是 2025 年初的模型和工具，结论不能直接外推，但它说明了一件事：连资深工程师对 AI 帮了多少忙的感觉都会出错，更别说对它的产出。
有判断力的人会把自己的判断放在前面。Anthropic 在 2025 年 12 月发布了一份内部调研，很多工程师说自己成了「AI agent 的经理」，有人说工作的七成以上已经变成了审代码和改代码。但同一份调研里，超过一半的人表示，能完全放手交给 Claude 的工作只占 0 到 20%；交出去的多半是一眼就能检查对错的活，高层次的思考和设计仍然留在自己手里。他们并不是不信任 AI，而是清楚哪些结论自己能核实，哪些不能。
Simon Willison 把这种能力叫作 vibe engineering，他列的清单里有自动化测试、提前规划、代码审查文化和扎实的手工测试，还特别提到一种直觉：知道什么可以外包给 AI，什么必须自己处理。
同一份 Anthropic 调研里还有一句很真实的话：越是自己想做的任务，越不会交给 Claude。人会把喜欢的活留在手里，这没有问题。重要的是，交出去的那部分你能判断对错，而方向始终由你来定。
三种人的共同点
回到 Kent Beck 那句话。贬值的 90% 是执行层面的熟练度；涨价的 10%，是知道要做什么、替用户和 agent 把话说清楚、判断交回来的结果对不对。这三件事，agent 都替代不了。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Kent Beck said one thing in 2023, and again in 2025, when I was interviewed, "90% of my skills became worthless, leaving 10% worth 1,000 times more." The 10 per cent he said was about knowing what to do, setting milestones, controlling the complexity of design.
Two years later, that phrase has become daily in many companies. In January of this year, the director of Claude Code, Boris Cherny, said that his own code had been written by Claude for more than two months, 22 PRs the previous day and 27 the previous day. What qualities on the human body have changed in value? My judgment is three kinds of people: people who have ideas but are lazy, people who are used to thinking for others, and people who are judgmental and not influenced by AI. The first two were often seen as short-cuts in character and are now the advantage; the third determines whether the first two can go in the right direction.
The thoughtless lazy: angent made up the implementation ring.
The author of Perl, Larry Wall, who in 1996 listed "slazy" as the first virtue of a programmer, defined lazy as a willingness to work hard to reduce overall energy consumption. Lazy people do not do anything, but do the same for the second time.
In the past, those who had ideas but were lazy suffered greatly. The idea was shut in the mouth, either by asking for help or by grinding its own skin, and most of it was gone. This is the ring that angent put on. It can work for hours, but it's a test of logs, changes the code, runs again, it's not tired, it's missing direction: It doesn't feel like it's a bad product, it doesn't say it's a bad idea.
Cherny described his current work and said that he could think about what next step was going to do, do creative things and leave the details to the model. Anthropic wrote more directly in an article in July 2025: Claude Code, "Anyone who can describe the problem clearly can make a solution." Examples of the article include the legal team itself setting up a system.
Lazy can be lazy in implementation, not in thinking. The idea is the only thing that such people need to do in person.
People who think for others: for users, for angent.
There's two floors for others, one for users, one for parties.
First, the user. Andrew Ng, after writing in July 2025, when the cost of writing codes fell, the bottleneck shifted to product management: deciding what to do. His view was that product managers with strong common sense for users often had the instinct to make the right decisions because they had a sufficiently accurate user model in their head. AI will faithfully implement your perspective: from the point of view of the boss, it makes a product that is a transformational portal everywhere; from the point of view of the user, it makes a product that you want to stay. The model can't do it for you.
Say again. In Anthropic's official prompt file, it is suggested that Claude be considered "a new employee who is smart, but new, and who doesn't know your rules and procedures". The document gives you a simple test: show your prompt to a colleague who doesn't understand the task and Claude doesn't. The Anthropic engineering team, speaking about how to do it, also wrote, "Stand on to the model." When they make codes, they spend more time grinding tools than grinding prompt; angent leaves the root directory and always uses the wrong relative path, and they change the tools to have to pass the absolute path and the problem disappears. In another multi-agent research system, they let an agent work out a flawed tool statement, and the subsequent agent with this tool has 40% less time to complete the task.
For the sake of angent, it's time to get to the point. In June 2025, Shopify CEO Tobi Lütke proposed replacing "prompt Project" with "context Project", which he defined as giving context, which was needed to complete the task, so that the task could be solved for the model. Karpathy then joined the public.
That's what people do when they don't like trouble. Before asking for help, let's consider the objectives, the limits and the extent to which they can be achieved. Open AI's GDPval paper has been contrasted with the deletion of the mission statement to 42% of the original length, and the model's performance has changed significantly. Ethan Mollick, in January of this year, wrote more clearly: What's scarce now is knowing what to do; people who would write demand files, who would assign tasks to their subordinates, and who would write them directly as prompts are good.
It needs to be made clear that giving context is not the same as talking. Anthropic, speaking about the context project, stressed that the more "a minimum set of high signal messages" were found, the easier the model would be to capture the focus. People don't have much to say for others, but every sentence works.
People with judgment: not being influenced by AI
The first two people said how to get things out. The third is how things come back.
AI is the most vulnerable to human bias, and its output always looks right: code runs, documents are smooth, words are firm. METR performed a random comparison experiment in 2025: 16 senior open source developers worked in their familiar code repository, allowing AI to be used, but with 19% more time to complete the task; they expected 24% before and 20% after completion. The experiment was based on models and tools from the beginning of 2025, which cannot be extrapolated directly, but it shows one thing: Even a senior engineer's sense of how much I've helped, not to mention the output.
Those with judgement put their judgment ahead. Anthropic had an internal study in publish in December 2025, and many engineers had said that they had become "the manager of AI angent" and that more than 70 per cent of the jobs had become code review and code change. But in the same study, more than half of the people said that only 0 to 20 per cent of the jobs that were completely released to Claude were available; most of the jobs that were sent were checked for right or wrong, and that high-level thinking and design remained in their hands. They are not distrusting AI, but they know which conclusions they can verify and which they cannot.
Simon Willison called this capability vibe engineering, which included automated testing, advance planning, code review culture and solid manual testing in his list, with particular reference to the hunch that knowing what could be outsourced to AI and what had to be handled by itself.
And in the same Anthropic study, there's a very real saying: the more you want to do it, the less you give it to Claude. People leave what they like in their hands. That's fine. The important thing is that you can judge right and wrong, and you always decide the direction.
The three things in common.
Go back to Kent Beck that sentence. Ninety percent of the devaluation is the degree of proficiency at the enforcement level; ten percent of the price increase is the knowledge of what to do, what to do for the user and what to say, and how to judge back. None of these three things can be replaced by angent.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.