← Back to Global Robotics News/Editorial detail
Global Robotics News / EDITORIAL DETAIL

Princeton CS Star Leaves Anthropic, Having Helped Make Claude Code a Hit

In the early morning hours of the 4th, Claude Code and Cowork's General Manager of Products, Cat Wu, Catherine Wu, issued a farewell letter announcing that Friday was her last day at Anthropic. This was the fourth core member of the Claude Code team to le

图片

  It's the first time in the history of the world.

Claude Code's "person number two" is gone!

On the morning of the 4th,Claude Code sends a farewell letter to the CEO of Cowork, Cat Wu, the founding product manager, Catherine Wu.She announced last Friday as her last day at Anthropic.

图片

This is the fourth core member to leave the Claude Code team in just 10 days.

According to CLI's Core Thariq Shihipar, who left 10 days ago, desktop founder Anthony Morris, and engineering manager Sunil Shah, the company's head of software development, said:During the critical period before and after Anthropic filed its IPO, the founding team was experiencing a steep turnaround.。

图片

If Tariq was the number one preacher outside Claude Code, and Anthony was the one who moved it from the terminal to the desktop, then Cat Wu was the number one preacher outside.The founder, Boris Cherny, is a key partner.。

She is also known as "Boris" and "Boris" by Silicon Valley media."Batman and Robin" by AnthropicOne wrote the original prototype code, one was responsible for the direction and landing of the entire product.

In her farewell letter, she talks about her first encounter with Claude Code two years ago.


From Princeton to the Vampire CircleShe brings in the tools to billions of dollars in products.

Before joining Anthropic, Cat Wu was a Wall Street intern, a Silicon Valley engineer, and a three-year venture capitalist.

No matter where she works, she can't stay and always wants to do something with her own hands.

When investors write their own codeHandmade paper bags with small knives

Cat Wu full name Catherine Wu graduated in 2017 with a Bachelor's degree in Psychology.The Department of Computer Science at Princeton UniversityWhen she was four years old, she and six friends developed a whiteboard application called Scratchwork that supports handwriting formulas and graphs.

After graduation, she went first.General Morgan stock derivatives trading platformInternship, then entryYouTubeand the earlyScale AIAs a product engineer, I work in open source data compilation companiesDagsterThe team developed the core commercialized product Dagster Cloud.

In October 2021, she joined a well-known venture capital firmIndex VenturesI've been an investor for almost three years.

Even sitting in a venture capital office and looking at projects, she hasn't given up the habit of writing code.Frequently writing their own programs to climb the dynamics of X in search of early startupsWe'll see what open-source projects are on fire.

In addition to her work, she is also a person who likes to make leather with her own hands.

During the epidemic, she was just on Amazon because she liked an Italian Il Bisonte leather bag sent by her friends, and because her apartment was too small to put up electrical tools.$40 to buy a new kitI decided to make a handmade leather bag at home.

In order to buy the original Italian leather that would slowly turn dark brown, she walked through the leather forums. She initially bought the hard American carved leather with an unstable knife.It took half a day for a chicken to cut through that thick skin.。

After stepping a few times on the crater, she finally touched the doorway, drilled a hole in it with a 3-millimeter razor blade, and rubbed the edge with a wooden stick.I had a whole series of real leather bags on my apartment's desk.。

图片

Cat Wu drilled a hand-held knife through an apartment's desk (left) with a hand-held leather bag (right)

After watching the three-year project, she still found it more interesting to do the product herself, so she decided to return to the tech company.

I didn't hit a line of code for ten hours.She then joined Claude Code.

In August 2024, Cat Wu joined Anthropic, initially as a product manager in the research team, responsible for transforming cutting-edge model capabilities into customer-oriented products.

In the autumn of that year, the Claude Code by Boris Cherny had just been put to the test inside the company, and the first demonstration on the internal Slack even received only two expressions.

Cat used it to write the Streamlit app to analyze user feedback, and then tried to do something bold.Let Claude Code help her build a reinforced learning environment from scratch.。

She recalled in her farewell letter,This time, she had Claude run for ten hours without writing a single line of code.。

This experience shocked her so much that she immediately decided to join the newly formed Claude Code team as a founding product manager.

The Cursor is a double jump with BorisIt's only two weeks before lightning strikes back.

With Claude Code's fire in May 2025, she and Boris Cherny became Silicon Valley's most sought-after partners.

On July 1, 2025, the parent company of CursorAnysphere suddenly gets out of hand and digs out both Boris Cherny and Cat Wu.The committee is headed by the chief architect and the engineering director, the general manager of the product. The news came out that the whole Silicon Valley was blown up.

I didn't expect things to turn around so quickly.

Only about two weeks later, before the new title was printed, the two returned to Anthropic in pairs.It also set one of the shortest jump records ever in the Silicon Valley AI circle.

图片

After returning, the pair made Claude Code their core business with billions of dollars in annual revenue.

Looking at the daily data, she also found that legal, financial and operational colleagues within the company are learning to use the Claude Code in the terminal, and non-technical personnel use more than 20 percent.

She immediately realized that ordinary people who don't write code also need these tools, so she took her team to do something for ordinary people in the office.Claude CoworkandTag。


Don't write long documents, don't plan for six monthsHow she brought out Claude Code's team

Small teams of dozens of people, able to create command lines, IDE plugins, desktop edges, web versions, and Cowork at the same time for more than a year, have benefited greatly from Cat Wu's unique bandwidth.

In an article published in March this year as a product designer and public speaker on the rise of the AI Index, she shared three experiences of the team's daily work.

图片

Cat Wu speaks at the Code with Claude 2026 convention in San Francisco

The reporting station is not open, and a desktop version is made using the "afternoon tea branch".

In traditional software companies, product managers often have to talk to customers for months, write long requirements documents, and wait for engineers to write the code by hand.

But in Cat's team, almost all product managers are full-time engineers, and the line between product, design and engineering is very blurred.

图片

Cat Wu's characters evolve: products, design and engineering go from a one-way flow to a deep fusion

She revealed a rule to the team at the World Exhibition of AI Engineers this summer:"If we have a point and we can't convince any engineer to write it, the product manager writes it himself".

The team canceled the traditional reporting station meeting and changed it to a half-day session to allow Claude Code to directly build prototypes that could run for discussion.

She even wrote herself a test kit that allowed Claude to automatically open a local app, simulate a real user clicking interface to accept new features, and not ask anyone if there was a problem.Just have Claude read the team's Slack chat record and see if anyone else has had the same problem.。

At the same time, she encouraged everyone to run away from mainline work."Side Quests" is the title of this episode.Take the afternoon off and try something new.

Today, it is widely popular.Claude Code for the desktop(This is the work of Anthony and his partner Alice, who left office 10 days ago.)AskUserQuestion Questions and AnswersandTodo waiting for the listIt's all coming from these afternoon line attempts.

The reason the team is so fast is because the underlying modeling capabilities are progressing too fast.

A test subject witnessed a 41-fold increaseWhen the model is upgraded, 20% of the prompts are removed.

Starting with Sonnet 3.5 in October 2024, Cat will test the same theme each time a new model is trained internally.Claude Code gives the open-source whiteboard tool Excalidraw with a "tablet tool".。

Early models failed each time; by June 2025, Opus 4, which began to run occasionally, the team was quick to record the video as a conference presentation.

And less than a year later, Opus 4.6 is stable.Live demonstrations in front of thousands of developers。

This change is also confirmed by the evaluation data of the third-party institution METR.

When Claude Code first conceived the model, it was able to complete only 21 minutes of human software tasks; 16 months later, Opus 4.6 was able to independently complete nearly 12 hours of human engineering tasks.The number of missions increased by about 41 times.。

图片

In her view, making AI products is like building a building on a rising terrain, and any roadmap that is more than six months old will be scrapped by the new model.

So she set up a"The three-month principle of repeal"Even if it is clear that a feature will be phased out in three months, get online immediately if you can make the product more useful today.

Once the next generation of models have this capability, the team will simply remove the original patch code.Just when they upgraded to Opus 4.6, they removed 20% of the system prompts.It's a good idea to leave a cleaner context for the model.

图片

She also never let the team focus on competing copywriting and was straightforward."If you keep thinking about your competitors, your execution speed will always be two weeks to a month slower than your limit".。

Since you're not staring at your competitors, the next step is to look at real users' daily usage habits.

I found that "99 percent of the bulletproof windows are unseen".Check Bugs by making video calls with Spit users

In the early days of Claude Code, for security reasons, every step of execution required a pop-up window to confirm the user, but Cat soon discovered a problem.

She observed,When users don't look directly at "agree" because they are accustomed to seeing 99% of the pop-ups, permission pop-ups lose their original security function.It's easy to let go of what's really dangerous.

To this end, she and her team introduced the "Automatic Mode", which incorporates a lightweight security classification model at the bottom, automatically performing the vast majority of conventional operations, reminding users only when there is a real risk.

To date, Auto-approval has become the default method for most developers.

She was also very attentive to the user's vomit and folding. When she discovered that the developer had written the page manually and had copied the paste test into the Chrome plugin, she tapped the board and opened it on both sides.Claude Code with Chrome。

When she saw people in the community complaining about Token being consumed too quickly, she didn't take the word for it.Users who actively contacted several posts made video calls to check out where they used too many tokens together with their local conversation records.。


From the perfectionistsTo learn how to "let go".

Looking back on her two years at Anthropic, Cat Wu's own mindset has also undergone a minor change.

She used to be a perfectionist who adjusted even a few millimeters of hole gaps repeatedly when she was in her apartment, and used to take every detail into her own hands.

But at the end of that passage, she confesses,Her first lesson in AI was "learning to let go".。

She writes that doing AI products is like surfing, the most important thing is staying on top of the wave.As a perfectionist, learning to give up control of every detail was her most difficult transition to adapt to.。

Two years ago, we didn't even have a single terminal command to let AI run automatically; two years later, hundreds of millions of users were getting used to letting AI work on their own, and she was on her last day at Anthropic.

In her personal Substack profile, she has written only three English words for five years."I'm not sure yet.。

Standing on the eve of Anthropic's IPO, the group of people who worked on Claude Code waved their goodbyes.

As for where to go next, as she always says, in the fast-moving wave of AI, the only thing to do is to stay on top.

RoboInd editorial channelResearch, robotics companies and market intelligence for cross-border sourcing.Explore the channel ↗
Share this
← Previous postBan Lifted: OpenAI Apologizes 11 Days After a User Was Suspended for Dismantling T...2026-11-04 Next post →A 5,000-Word Specification Yields Just 14% from Claude: Testing 28 Agent Combinati...2026-11-04
End of editorialExplore more from Global Robotics News ↗