新智元报道
Claude Code的「二号人物」，也走了！
4日凌晨，Claude Code与Cowork产品总负责人、创始产品经理Cat Wu（Catherine Wu）发出告别信，宣布上周五是她在Anthropic的最后一天。
这已经是短短10天里，Claude Code团队离开的第4位核心成员。
算上10天前同日离职的CLI核心Thariq Shihipar、桌面版主创Anthony Morris，以及工程经理Sunil Shah，在Anthropic递交IPO招股书前后的关键时期，这支创始团队正经历一场密集的离职潮。
如果说Thariq是Claude Code对外的头号布道者，Anthony是把它从终端搬上桌面的人，那Cat Wu就是创始人Boris Cherny最核心的搭档。
硅谷媒体曾把Boris和她并称为Anthropic的「蝙蝠侠与罗宾」，一个写下了最初的原型代码，一个负责整个产品的方向与落地。
而在告别信里，她聊起了自己两年前刚接触Claude Code时的场景。
在加入Anthropic之前，Cat Wu做过华尔街交易台实习生、硅谷工程师，也当过三年风险投资人。
无论换到哪个岗位，她都闲不住，总想自己动手做点东西出来。
Cat Wu全名Catherine Wu，2017年本科毕业于普林斯顿大学计算机科学系。大四那年，她和6个朋友一起开发了一款支持手写公式与图表的协作白板应用Scratchwork。
毕业后，她先去摩根大通股票衍生品交易台实习，随后进入YouTube和早期的Scale AI担任产品工程师，又在开源数据编排公司Dagster带队做出了核心商业化产品Dagster Cloud。
2021年10月，她加入知名风投机构Index Ventures，当了近三年投资人。
哪怕坐在风投办公室里看项目，她也没放下写代码的习惯，经常自己写程序爬X上的动态来寻找早期创业公司，看看哪些开源项目正在火起来。
工作之外，她还是个喜欢自己动手做皮具的人。
疫情期间，因为喜欢朋友送的一只意大利Il Bisonte皮包，又苦于公寓太小放不下电动工具，她干脆在亚马逊上花40美元买了一套新手工具包，决定在家手工缝皮包。
为了买到用久了会慢慢变成深棕色的正宗意大利植鞣牛皮，她逛遍了皮具论坛。最初因为买错成很硬的美国雕刻皮，她拿着一把不太稳的小折刀，费了半天劲也割不透那层厚牛皮。
踩了几次坑后她终于摸清门道，用3毫米斩刀一锤一锤打孔、拿木棒打磨封边，在公寓书桌上亲手缝出了整整一个系列的真皮包。
Cat Wu在公寓书桌上用斩刀手工打孔（左）与亲手缝制出的成品皮包（右）
在风投看了三年项目后，她还是觉得亲手做产品更有意思，于是决定重回科技公司。
2024年8月，Cat Wu加入Anthropic，最初在研究团队担任产品经理，负责把前沿模型能力做成面向客户的产品。
那年秋天，Boris Cherny做出的Claude Code刚刚在公司内部开放试用，最早发在内部Slack上的演示甚至只收到两个表情包回复。
Cat先拿它写Streamlit应用来分析用户反馈，随后试着干了一件大胆的事，让Claude Code帮她从零搭建一个强化学习环境（RL env）。
她在告别信里回忆，那次她让Claude连续跑了十几个小时，自己一行代码都没手写，最后居然真把整个环境搭出来了。
这次体验彻底震撼了她，让她当即决定转岗加入刚刚起步的Claude Code团队，成为创始产品经理。
随着Claude Code在2025年5月大火，她和Boris Cherny成了硅谷最抢手的搭档，连头号竞争对手都坐不住了。
2025年7月1日，Cursor母公司Anysphere突然出手，把Boris Cherny和Cat Wu双双挖走，分别委以首席架构师兼工程主管、产品总负责人的重任。消息一出，整个硅谷都炸了锅。
没想到事情很快迎来了大反转。
仅仅过了大约两周，在印着新头衔的名片大概还没印出来之前，两人又双双回到Anthropic，创下了硅谷AI圈有史以来最短的跳槽纪录之一。
回归之后，两人一路把Claude Code做成了年化收入数十亿美元的核心业务。
在看日常数据时她还发现，公司内部连法务、财务和运营同事都在学着用终端里的Claude Code，非技术人员的使用比例超过了20%。
她立刻意识到，不写代码的普通人同样需要这类工具，于是带队做出了面向普通办公人群的Claude Cowork和Tag。
几十个人的小团队，能在一年多里同时做出命令行、IDE插件、桌面端、网页版和Cowork，很大程度上得益于Cat Wu独特的带队方式。
在今年3月发表的文章《在AI指数级增长下做产品》和多次公开演讲中，她分享过团队日常工作的三条经验。
Cat Wu在旧金山Code with Claude 2026大会上发表演讲
在传统软件公司里，产品经理往往要先聊几个月客户、写好长篇需求文档，才能等来工程师动手写代码。
但在Cat的团队里，几乎所有产品经理都是全职工程师出身，产品、设计与工程之间的分工界限非常模糊。
Cat Wu画出的角色演变：产品、设计与工程从单向流水线变成深度交融
她在今年夏天的AI工程师世界博览会上透露过团队的一条规矩，「如果我们有个点子，却没能说服任何一个工程师去写，那产品经理就自己动手把它写出来。」
团队取消了传统的汇报站会，改成花半天时间让Claude Code直接做出能跑的原型来讨论。
她甚至给自己写了一套测试工具，让Claude自动打开本地应用、模拟真实用户点击界面来验收新功能；碰到报错也不用问人，直接让Claude去读团队的Slack聊天记录，看看是不是已经有同事遇到过同样的问题。
同时，她鼓励全员在主线工作之外去跑「支线任务（Side Quests）」，拿一个下午的时间自由尝试新点子。
如今广受欢迎的Claude Code桌面版（正是10天前离职的Anthony与搭档Alice的作品）、AskUserQuestion反问工具以及Todo待办清单，全都来自这些一下午的支线尝试。
团队之所以这么讲究速度，是因为底层的模型能力进步太快了。
从2024年10月的Sonnet 3.5开始，每当内部训出新模型，Cat都会测同一道题目，让Claude Code给开源白板工具Excalidraw加上一个「表格工具」。
早期的模型每次都做不出来；到了2025年6月的Opus 4，它开始偶尔能跑通，团队赶紧录下视频当成发布会演示。
而不到一年后的Opus 4.6，已经稳到可以在几千名开发者面前现场直播演示。
第三方机构METR的评测数据也印证了这种变化。
Claude Code刚立项时，模型只能完成人类21分钟的软件任务；16个月后，Opus 4.6已经能独立完成人类近12小时的工程任务，任务时长足足提升了约41倍。
在她看来，做AI产品就像在一块不断隆起的地面上盖楼，任何半年以上的路线图都会被新模型作废。
因此她立下了一条「三个月报废原则」，哪怕明知一个功能三个月后就会被新模型淘汰，只要今天能让产品更好用，就立刻动手上线。
一旦下一代模型自带了这项能力，团队就会干脆地删掉当初的补丁代码。仅在升级Opus 4.6时，她们就删掉了20%的系统提示词，把更干净的上下文留给模型。
她也从不让团队盯着竞争对手抄作业，并直言「如果你老想着竞争对手，你的执行速度永远会比自己的极限慢上两周到一个月」。
既然不盯着竞争对手，那下一步做什么就全看真实用户的日常使用习惯。
Claude Code早期为了安全，每执行一步操作都要弹窗让用户确认，但Cat很快发现了一个问题。
她观察到，当用户出于习惯对99%的弹窗看都不看就直接点「同意」时，权限弹窗就失去了原本的安全作用，真正危险的操作反而容易被顺手放过去。
为此她和团队推出了「自动模式」，在底层加入一个轻量级安全分类模型，自动放行绝大多数常规操作，只在真正有风险的时候才提醒用户。
到了今天，开启自动审批（Auto-approve）已经成为大多数开发者的默认用法。
平时她也非常留意用户的吐槽和折腾。发现开发者写完网页老是手动切到Chrome插件里复制粘贴测试，她就拍板把两边打通，做成了内置的Claude Code with Chrome。
看到社区有人抱怨Token消耗太快，她没有拿套话敷衍，而是主动联系了几位发帖的用户打视频电话，对着他们本地的对话记录一起排查到底是哪里用掉了太多Token。
回看Cat Wu在Anthropic的这两年，她自己的心态也经历了一场不小的变化。
以前在公寓里手工缝皮包时，她是一个连几毫米孔距误差都要反复调整的完美主义者，习惯把每个细节都攥在自己手里。
但在那篇长文的末尾，她却坦言，做AI产品教会她的第一堂课，就是「学会放手」。
她在文中写道，做AI产品就像在冲浪，最重要的事情是留在浪头上，作为一个完美主义者，学会放弃对每一个细节的把控，是她最难适应的转变。
两年前，大家连一行终端命令都不敢让AI自动运行；两年后，成百上千万用户已经习惯放手让AI自己干活，而她也迎来了自己在Anthropic的最后一天。
在她的个人Substack简介栏里，五年来一直只写着三个英文单词，「Not sure yet.（还没想好）」。
站在Anthropic冲刺IPO的前夜，这群亲手做出Claude Code的人正一个个挥手告别。
至于下一步去哪，就像她常说的那样，在飞速向前的AI浪潮里，唯一要做的就是始终留在浪尖上。
参考资料：
https://x.com/_catwu/status/2117731934340337676
https://claude.com/resources/articles/product-management-on-the-ai-exponential
https://arstechnica.com/ai/2026/05/claude-codes-product-lead-talks-usage-limits-transparency-and-the-lean-harness/
https://aidb.news/researchers/catherine-wu
编辑：摩西
不喜欢
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。
New Ji-won.
Claude Code's character number two, too!
In the early hours of 4 April, Claude Code and Cathrine Wu, the founding product manager, sent a farewell letter announcing last Friday as her last day in Anthropic.
This is the fourth core member of Claude Code left in just 10 days.
The CLI core, Thariq Shihipar, whose departure took place 10 days ago, the desktop owner Anthony Morris, and the engineering manager Sunil Shah, are experiencing an intense wave of separations in the critical period before and after Anthropic presented the IPO book.
If Thariq was Claude Code’s leading preacher, Anthony was the one who moved it from the terminal to the desktop, Cat Wu was the founder Boris Cherny’s core partner.
Silicon Valley media used to write down the original prototype code for Boris and her, also known as "Battleman and Robin" as Anthropic, one responsible for the direction and landing of the entire product.
And in her farewell letter, she talked about her first contact with Claude Code two years ago.
Before joining Anthropic, Cat Wu was an intern at Wall Street Trading Desk, a Silicon Valley engineer and a venture investor for three years.
In any position she can't help but do something herself.
Cat Wu, full name Catherine Wu, graduated undergraduate from the Department of Computer Sciences at Princeton University in 2017. In that year, she and six friends developed a collaborative whiteboard application to support handwritten formulas and charts.
Upon graduation, she went to the Morgan Chase Stock Derivatives Exchange, then to YouTube and the early Scale AI as product engineer, and made the core commercial product Dagster Cloud in the Dagster team of open source data organization.
In October 2021, she joined the prestigious Index Ventures as an investor for almost three years.
Even sitting in a wind office to watch the project, she did not put down her habit of writing codes and often wrote her own program to climb X to look for early start-up companies to see which open-source projects were on fire.
Outside of work, she's a man who likes to make leather.
During the epidemic, she bought a beginner kit on Amazon for $40 and decided to sew a handbag at home because she liked an Italian Il Bisonte bag sent by her friend, and because her apartment was too small for electric tools.
In order to buy it for a long period of time and slowly turn it into a dark brown real Italian cow-skin, she wandered through the skin forum. At first, she had a small, unsettled crutches, which took half a day to cut through the thick cow skin, because she had bought it wrongly.
After stepping on several pits, she finally touched the doorway, hammered the hole with a three-millimeter knife, sharpened the edge with a wooden stick, and sewd an entire series of real purses with her hands on the desk of the apartment.
Cat Wu's hand-stabbing (left) and hand-sewn finished purses on the apartment desk (right)
After three years on the project, she found it more interesting to make the product herself and decided to return to the company.
In August 2024, Cat Wu joined Anthropic, initially as a product manager in the research team, to make the front-line modelling capability a client-oriented product.
In the autumn of that year, Claude Code, by Boris Cherny, had just opened a trial inside the company, and the first demonstration on the inside of Slack had received even two emoticon responses.
Cat used it to analyze user feedback and then tried to do a bold thing to get Claude Code to help her build an enhanced learning environment from scratch.
In her farewell letter, she recalled that she had left Claude with no hand-written code in her line of work for over a dozen hours, and that she had finally put out the whole environment.
The experience shocked her by her immediate decision to join the newly established Claude Code team as a founding product manager.
With Claude Code in the May 2025 fire, she and Boris Cherny became the best partners in Silicon Valley, and even the number one rival could not sit.
On 1 July 2025, the Cursor parent company, Anyshere, took out both Boris Cherny and Cat Wu with the responsibility of the Chief Architect and Chief Engineer and Chief Product Officer respectively. The whole Silicon Valley blew up the pot.
I didn't think it would be a big turn.
Just about two weeks later, before the card with the new title was probably printed, both returned to Anthropic, creating one of the shortest jumps ever recorded in the Silicon Valley Ai Circle.
Following their return, Claude Code was a core business with billions of dollars of annualized income.
Looking at the daily data, she also found that Claude Code, a member of the company's legal, financial and operating colleagues, was studying the terminals, using over 20 per cent of the unskilled.
She immediately realized that ordinary people who did not write needed the same tools, and led a team to create Claude Cowork and Tag for ordinary people.
Dozens of small teams have been able to produce command line, IDE plugins, desktops, web pages and Cowork simultaneously for more than a year, largely thanks to Cat Wu ' s unique way of bringing together.
She shared three lessons from the daily work of the team in her article " Making a product under the growth of the AI Index " and in several public speeches in March of this year.
Cat Wu addressed the Code with Claude 2026 in San Francisco.
In traditional software companies, product managers often have to talk about clients for months and write long demand files before engineers can start writing codes.
In Cat ' s team, however, almost all product managers are full-time engineers and the division of labour between products, designs and works is very blurred.
Cat Wu's role evolution: product, design and engineering from one-way waterline to depth
She revealed the team's rule at the AI Engineers' World Expo this summer: "If we had an idea, but we didn't convince any engineer to write it, the product manager would write it himself."
The team cancelled the traditional reporting station meeting and changed it to a half-day session to allow Claude Code to make a running prototype.
She even wrote herself a test tool that enabled Claude to automatically open a local application, simulate a real user click on the acceptance check new feature; she did not have to ask anyone about errors, and asked Claude directly to read the team's Slack chat record to see if colleagues had experienced the same problem.
At the same time, she encouraged the whole team to run "Side Quests" outside the main line of work and take an afternoon to try new ideas freely.
The popular Claude Code desktop version (the work of Anthony and his partner Alice, who left 10 days ago), the AskUserQuestion counter-examination tool and the Todo to-do list) all came from these afternoon tributaries.
The team's speed is due to the fact that the bottom modeling capacity is moving too fast.
Starting with Sonet 3.5 in October 2024, every time a new model is trained in-house, Cat will measure the same subject, allowing Claude Code to add a "table tool" to the open-source whiteboard tool Excelidraw.
The early model was never done; by June 2025, Opus 4, it started running occasionally, and the team quickly recorded the video as a publish demonstration.
Less than a year later, Opus 4.6 has stabilized to be able to deliver live presentations in front of thousands of developers.
This change is also confirmed by the assessment data of the third-party agency METR.
When Claude Code first set up the project, the model could only complete a 21-minute human software mission; 16 months later, Opus 4.6 had been able to do almost 12 hours of human engineering on its own, a task that had increased by about 41 times.
In her view, making an AI product was like building a building on a rising ground, and any road map for more than six months would be invalidated by a new model.
So she put in place a three-month "end-of-life principle" that even if she knew that a function would be eliminated from the new model after three months, she would go online as soon as she could make the product work better today.
Once the next generation model takes this capability, the team simply deletes the original patch code. Only when the Opus 4.6 was upgraded, they deleted 20 per cent of the system prompt, leaving the cleaner context to the model.
She never lets the team stare at the competition's copying, and she says, "If you keep thinking about the competition, your execution will always be two weeks to one month behind your limits."
Since next step does not stare at competitors, it depends on the daily usage habits of real users.
Claude Code had been able to identify users with a bullet window for each step of the operation in the early days of the operation, but Cat soon discovered a problem.
She observed that the permission bullet window had lost its original security function when the user had been used to "consent" without looking at 99 per cent of the bullet windows, and that the real dangerous operation was easily left unattended.
To that end, she and her team introduced an "automated mode" by adding a lightweight safety classification model to the bottom, automatically releasing most of the routine operations and reminding users only when there is a real risk.
To date, AutoApprove has become a default for most developers.
She was also very careful about the user's vomiting and twitching. When it was discovered that the developers were always manually cutting into the Chrome plugin to copy the paste test, she boarded both sides and made Claude Code with Chrome.
It was observed that some members of the community complained that Token was being consumed too quickly and that instead of using a set of words, she had volunteered to contact several of the users who had sent the posts to make video calls and to run together their local conversation records to find out where too many Token had been used.
Looking back at Cat Wu's two years in Anthropic, her own mindset has also undergone a considerable change.
She was a perfectist who had to adjust several millimeters of pore error when she was in a hand-suture bag in the apartment, and used to keep every detail in her hands.
But at the end of that long article, she confessed to doing the AI. product to teach her the first lesson of "learning to let go."
She wrote that doing an AI product is like surfing, and the most important thing is to stay on the wave and learn, as a perfectist, to give up on every detail, is the most difficult change for her.
Two years ago, a line of terminal commands was not allowed to run automatically; two years later, millions of users had become used to letting AI work for herself, and she had her last day in Anthropic.
In her personal Substack profile column, there have been only three words in English for five years, "not sure yet."
On the night before Anthropic sprinted on IPO, the people who made Claude Code were waving goodbye.
As for where next step went, as she often said, the only thing to do in the fast-forward wave of AI was to stay on top of it.
References:
https://x.com/_catwu/status/2117731934340337676
https://claude.com/resources/articles/product-management-on-the-ai-exponential
https://arstechnica.com/ai/2026/05/claude-codes-product-lead-talks-usage-limits-transparency-and-the-lean-harness/
https://aidb.news/researchers/catherine-wu
Editor: Moses
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.