Claude Code founding product manager Cat Wu announces departure from Anthropic

The "number two" of Claude Code has also left!

In the early hours of the 4th, Cat Wu (Catherine Wu), head of product for Claude Code and Cowork and founding product manager, sent out a farewell letter announcing that last Friday was her final day at Anthropic.

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

Counting Thariq Shihipar, a core member of the CLI who left the same day 10 days ago, Anthony Morris, the creator of the desktop version, and engineering manager Sunil Shah, this founding team is going through a dense wave of departures at a critical time around Anthropic's filing of its IPO prospectus.

If Thariq was Claude Code's top external evangelist and Anthony was the one who moved it from the terminal to the desktop, then Cat Wu was founder Boris Cherny's most central partner.

Silicon Valley media once dubbed Boris and her Anthropic's "Batman and Robin": one wrote the original prototype code, the other was responsible for the entire product's direction and execution.

And in her farewell letter, she talked about the scene when she first encountered Claude Code two years ago.

From Princeton to the venture capital world, she turned an internal small tool into a multibillion-dollar product

Before joining Anthropic, Cat Wu had been an intern on a Wall Street trading desk, a Silicon Valley engineer, and a venture capitalist for three years.

No matter which position she moved to, she could never sit still and always wanted to build something with her own hands.

When an investor still writes code herself and hand-sews leather bags with a knife at home

Cat Wu's full name is Catherine Wu, and she graduated with a bachelor's degree in computer science from Princeton University in 2017. In her senior year, she and six friends developed Scratchwork, a collaborative whiteboard app that supports handwritten formulas and diagrams.

After graduating, she first interned on the equity derivatives trading desk at JPMorgan, then went on to work as a product engineer at YouTube and early-stage Scale AI, and later led a team at the open-source data orchestration company Dagster to build its core commercial product, Dagster Cloud.

In October 2021, she joined the well-known venture capital firm Index Ventures and spent nearly three years as an investor.

Even while sitting in a venture capital office reviewing projects, she never gave up her habit of writing code, and often wrote programs herself to scrape X feeds in search of early-stage startups and see which open-source projects were catching on.

Outside of work, she is also someone who enjoys making leather goods by hand.

During the pandemic, because she liked an Italian Il Bisonte leather bag a friend had given her and was frustrated that her apartment was too small to fit power tools, she simply bought a $40 beginner's toolkit on Amazon and decided to hand-sew leather bags at home.

In order to buy authentic Italian vegetable-tanned leather that gradually turns dark brown with use, she browsed leatherworking forums all over. At first, because she mistakenly bought very stiff American carving leather, she held a rather unsteady small folding knife and spent ages trying without success to cut through that thick layer of leather.

After stumbling a few times, she finally figured out the ropes: punching holes hammer by hammer with a 3mm pricking iron and polishing the edges with a wooden stick, she hand-sewed an entire series of genuine leather bags on her apartment desk.

Cat Wu hand-punching holes with a pricking iron on her apartment desk (left) and the finished leather bags she hand-sewed (right)

After three years of reviewing projects in venture capital, she still felt that building products with her own hands was more interesting, so she decided to return to a tech company.

After not writing a single line of code for over ten hours, she switched roles on the spot and joined Claude Code

In August 2024, Cat Wu joined Anthropic, initially as a product manager on the research team, responsible for turning frontier model capabilities into customer-facing products.

That autumn, the Claude Code built by Boris Cherny had just opened for internal trials at the company, and the earliest demo posted on internal Slack even received only two emoji reactions.

Cat first used it to write a Streamlit app to analyze user feedback, and then tried something bold: having Claude Code help her build a reinforcement learning environment (RL env) from scratch.

In her farewell letter, she recalled that she had Claude run continuously for over ten hours that time, without writing a single line of code herself, and in the end it actually built the entire environment.

That experience completely stunned her, and she immediately decided to switch roles and join the just-getting-started Claude Code team as founding product manager.

She and Boris both jumped ship to Cursor, only to make a lightning-fast return after just two weeks

As Claude Code took off in May 2025, she and Boris Cherny became Silicon Valley's most sought-after duo, and even their top competitor could not sit still.

On July 1, 2025, Anysphere, the parent company of Cursor, suddenly made its move and poached both Boris Cherny and Cat Wu, entrusting them with the important roles of chief architect and head of engineering, and head of product, respectively. The news set the whole of Silicon Valley abuzz.

Unexpectedly, things soon took a major turn.

After only about two weeks, before business cards printed with their new titles had probably even come out, the two were back at Anthropic together, setting one of the shortest job-hopping records in the history of Silicon Valley's AI circle.

After returning, the two went on to build Claude Code into a core business with billions of dollars in annualized revenue.

While looking at daily data, she also discovered that within the company, even colleagues in legal, finance, and operations were learning to use Claude Code in the terminal, and the proportion of non-technical users exceeded 20%.

She immediately realized that ordinary people who do not write code also need this kind of tool, so she led the team to build Claude Cowork and Tag for the general office worker population.

No long documents, no half-year plans: how she built the Claude Code team

A small team of a few dozen people was able to simultaneously build the command line, IDE plugin, desktop app, web version, and Cowork in just over a year, largely thanks to Cat Wu's unique leadership style.

In her article "Building Products in the Age of AI Exponential Growth," published in March this year, and in multiple public talks, she shared three lessons from the team's daily work.

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

No reporting stand-ups; building the desktop version through "afternoon tea side quests"

In traditional software companies, product managers often have to spend months talking to customers and writing lengthy requirements documents before engineers even start writing code.

But on Cat's team, almost all product managers come from full-time engineering backgrounds, and the boundaries between product, design, and engineering are very blurred.

Cat Wu's diagram of role evolution: product, design, and engineering went from a one-way pipeline to deep integration

The team abolished traditional reporting stand-ups and instead spent half a day letting Claude Code directly build runnable prototypes for discussion.

She even wrote herself a set of testing tools that let Claude automatically open local applications and simulate real users clicking through the interface to validate new features; when errors came up, there was no need to ask anyone—she simply had Claude read the team's Slack chat logs to see whether a colleague had already encountered the same problem.

At the same time, she encouraged everyone to run "Side Quests" outside their main line of work, taking an afternoon to freely try new ideas.

Today's popular Claude Code desktop app (the work of Anthony, who left 10 days ago, and his partner Alice), the AskUserQuestion follow-up tool, and the Todo checklist all came from these afternoon side-quest experiments.

The reason the team cares so much about speed is that the underlying model capabilities are improving too fast.

A test question witnesses a 41-fold improvement; when the model upgrades, 20% of the prompt is deleted

Starting with Sonnet 3.5 in October 2024, whenever a new model was trained internally, Cat would test the same question: have Claude Code add a "table tool" to the open-source whiteboard tool Excalidraw.

Early models could never do it; by Opus 4 in June 2025, it began to occasionally work, and the team hurriedly recorded a video to use as a launch demo.

Less than a year later, Opus 4.6 was stable enough to demo live in front of thousands of developers.

Evaluation data from the third-party organization METR also confirms this change.

When Claude Code was first launched, the model could only complete software tasks that took a human 21 minutes; 16 months later, Opus 4.6 could independently complete engineering tasks that take a human nearly 12 hours, an increase in task duration of about 41 times.

So she established a "three-month obsolescence principle": even if she knew a feature would be eliminated by a new model in three months, as long as it made the product better today, she would immediately build and ship it.

Once the next-generation model came with that capability built in, the team would simply delete the patch code from before. When upgrading to Opus 4.6 alone, they deleted 20% of the system prompt, leaving cleaner context for the model.

She also never let the team stare at competitors and copy their homework, saying bluntly, "If you're always thinking about competitors, your execution speed will always be two weeks to a month slower than your own limit."

Since they do not stare at competitors, what to do next depends entirely on real users' daily usage habits.

Discovering that "99% of pop-ups go unread," and video-calling users who complained to investigate bugs

In its early days, Claude Code required a pop-up confirmation for every operation for safety, but Cat soon discovered a problem.

She observed that when users, out of habit, clicked "Agree" on 99% of pop-ups without even looking, the permission pop-ups lost their original safety role, and truly dangerous operations were instead more likely to slip through.

So she and the team launched "auto mode," adding a lightweight safety classification model underneath that automatically allows the vast majority of routine operations and only alerts users when there is a real risk.

Today, enabling Auto-approve has become the default usage for most developers.

She also pays close attention to users' complaints and tinkering. When she found that developers kept manually switching to the Chrome plugin to copy and paste for testing after writing web pages, she made the call to connect the two sides and built the built-in Claude Code with Chrome.

When she saw people in the community complaining that tokens were being consumed too quickly, she did not fob them off with platitudes; instead, she proactively contacted several users who had posted and video-called them, going through their local conversation logs together to investigate exactly where too many tokens were being used.

From perfectionist to someone who learned to "let go"

Looking back at Cat Wu's two years at Anthropic, her own mindset also underwent a considerable change.

When she used to hand-sew leather bags in her apartment, she was a perfectionist who would repeatedly adjust even a few millimeters of error in hole spacing, accustomed to holding every detail in her own hands.

In her personal Substack bio, for five years she had only three English words: "Not sure yet."

On the eve of Anthropic's sprint toward an IPO, the people who built Claude Code with their own hands are waving goodbye one by one.

References:

This article comes from the WeChat public account "Xinzhiyuan," author: ASI Revelations; editor: Moses, published by 36Kr with authorization.