# How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)

Podcast: Lenny's Podcast: Product | Career | Growth
Published: May 5, 2026
Reading time: 19 min
Canonical: https://podbrew.app/briefs/lenny-s-podcast-product-career-growth-how-anthropic-s-product-team-moves-faster-

Cat Wu, Head of Product for Claude Code and Cowork at Anthropic, joins Podbrew to share how her team achieves an unparalleled shipping cadence, moving from months to weeks, and even days, for product releases. She offers an inside look at the strategies enabling Anthropic to build one of the most important AI products of this generation at an incredible speed.

The conversation delves into the evolving skills product managers need to thrive in the AI era, including the crucial ability to build products that anticipate future model capabilities. Cat also reveals why Claude's unique personality is fundamental to its success and how Anthropic's deep mission alignment eliminates organizational friction, fostering rapid development.

These discussions are vital for anyone navigating the dynamic AI landscape, providing actionable insights into product development, team velocity, and strategic foresight. Understanding Anthropic's approach offers a powerful framework for building impactful AI products and preparing for the next wave of technological advancements.

## Key takeaways

- AI-native products thrive on hyper-speed iteration, requiring product teams to ship features weekly or even daily.

- The product manager role is evolving to prioritize strong product taste and rapid decision-making, as the cost and time involved in writing code decrease.

- Effective cross-functional alignment and the proactive removal of shipping blockers are crucial for maintaining an extremely fast product development cadence.

- Combat the ambiguity of general LLMs by setting precise goals: define target users, specify core problems, and articulate clear use cases for features to guide development effectively.

- Establish repeatable processes like 'research previews' for quick feedback, streamline cross-functional launches, and empower teams with transparent metrics and principles for independent decision-making.

- Product, engineering, and design roles are merging, requiring professionals to have cross-functional skills.

- The ability to decide what to build and how to create an excellent user experience (product taste) is becoming the most valuable skill as code generation becomes cheaper.

- Human common sense, emotional intelligence, and stakeholder management skills are irreplaceable in navigating the complexities of product development, where AI models currently fall short.

- In an AI-driven, rapidly changing environment, individuals must possess first principles thinking, adaptability, and a low ego to identify and address critical team needs and skill gaps.

- Rapid shipping in AI can lead to intentionally overlapping features, allowing for quick user feedback on different approaches.

- The downside of rapid feature releases is increased user overwhelm and the need for more in-product education to guide users effectively.

- The mission of delivering safe AGI acts as the primary filter for all strategic and product decisions, enabling swift and aligned execution.

- A strong mission encourages teams to prioritize collective company goals over individual product metrics, fostering organizational alignment and decisive action.

- CoWork can autonomously draft complex documents and presentations, such as a 20-page slide deck, by synthesizing information from connected data sources and adhering to organizational design systems, substantially reducing manual effort.

- Token cost per knowledge worker at Anthropic is increasing with model improvements, as users delegate more tasks to AI tools like Claude, Code, and CoWork, despite still being lower than average salaries.

- Utilize model introspection by asking models to explain unexpected behaviors, which helps debug system prompts and improve model harnesses.

- As Large Language Models improve, product teams can remove 'crutch' features that were originally implemented to compensate for earlier models' limitations, simplifying the user interface and system prompts.

- Newer, more capable LLM models enable the development of entirely new product functionalities, such as reliable multi-agent code review, which were previously too inaccurate to launch.

- Strive for 100% reliability in AI automations; a partially reliable automation often creates more work than it saves.

- The evolution from chat-based AI (2023) to action-based AI (2024) allows agents to perform tasks autonomously.

## 00:00 - 04:01 Anthropic's Hyper-Speed Shipping Cadence Redefines the PM Role

Anthropic has achieved an unprecedented shipping pace for product features, often launching updates in days, down from typical months-long cycles. This extreme speed is considered essential for building AI-native products, which benefit immensely from rapid iteration.

The accelerated development cycle dramatically changes the product manager's role. It shifts from extensive upfront planning to a focus on quickly launching features, sometimes weekly, to gather feedback and adapt. PMs need to cultivate strong product taste to decide what to build as coding itself becomes less of a bottleneck.

Kat Wu, Head of Product for Claude Code and CoWork at Anthropic, exemplifies this evolving role. Her focus is on bridging the gap between the immediate product state and a 3-6 month visionary goal, while ensuring robust cross-functional alignment. She collaborates closely with her tech lead, Boris, to remove any barriers that could slow down the shipping process.

## 04:21 - 10:02 How Anthropic PMs Enable Rapid Product Shipping in the AI Era

The rapid pace of AI development has dramatically compressed product timelines from six-to-twelve months down to weeks or even days. This shift means Product Managers for AI-native products must now prioritize quickly getting features into users' hands, rather than focusing on lengthy multi-quarter roadmaps or extensive cross-team coordination for slower feature releases.

Given the general nature of Large Language Models (LLMs), a key challenge for PMs is defining clear goals amidst ambiguity. Effective PMs specify key users, identify the main problems to solve, and articulate precise use cases. For example, for professional developers, a goal might be to achieve 'zero permission prompts,' which helps narrow down development approaches and provides clarity.

Anthropic employs a repeatable process to maintain its shipping speed, including using 'research previews' to launch early product ideas quickly. These are clearly branded to manage user expectations, allowing the team to gather feedback and iterate rapidly with reduced commitment. A streamlined cross-functional framework also facilitates quick launches, with engineers posting ready features in an 'evergreen launch room,' enabling marketing and documentation teams to turn around announcements by the next day.

To further empower rapid, independent decision-making, Anthropic provides teams with rigorous weekly metric readouts. These sessions ensure everyone deeply understands the business, key goals, and their trends. Additionally, a clear set of team principles outlines key users and acceptable trade-offs, enabling individual team members to make informed decisions without constant approval from PMs or other stakeholders.

> I think the PMs who do the best on AI-native products are, are the ones who can figure out, how can I like shorten the time from having this idea to actually getting the product in the hands of users

## 12:06 - 16:06 Anthropic addresses Claude code leak and Open Claude restriction due to high demand

Anthropic experienced a Claude code source code leak due to human error, despite the update passing through two layers of human review. The incident was identified as a process failure, leading the company to harden its internal processes and implement additional safeguards to prevent future occurrences.

The decision to restrict Open Claude usage, which caused user frustration, was a direct response to the overwhelming demand for Claude. The existing infrastructure and token efficiency were not designed to support the distinct usage patterns of third-party products as effectively as Anthropic's first-party offerings.

Ultimately, Anthropic made the difficult choice to prioritize its first-party products and API. While credits were offered to subscribers for a seamless transition, the company emphasized the necessity of profitability and the inability to subsidize unlimited compute usage for third-party integrations during periods of intense demand for its resources.

> people don't understand, businesses are trying to make money. We're trying to be profitable here. We can't just like give away compute when it's so in demand.

## 16:06 - 20:07 Product Roles Are Merging, With Product Taste Becoming Most Valuable

The traditional lines between product managers, engineers, and designers are blurring, with individuals increasingly taking on responsibilities across these roles. Engineers perform PM functions, designers contribute code, and PMs might engage in engineering work.

At Anthropic, the team emphasizes hiring engineers with excellent product taste. This approach enables engineers to take a product from user feedback to shipment end-to-end with minimal PM oversight, creating a more efficient development process. Most PMs and designers at Anthropic have engineering or coding backgrounds.

As the cost of writing code decreases due to advancements in AI, the critical skill shifts to deciding what to build and how to craft the optimal user experience. This "product taste" is essential for prioritizing among numerous user requests and developing delightful features.

While an engineering background is currently valuable for understanding implementation difficulty and aiding prioritization (i.e., knowing if something is easy to "just spend an hour doing" versus a complex, costly build), this advantage might change quickly as coding capabilities continue to evolve.

> as code becomes much cheaper to write, the thing that becomes more valuable is deciding what to write.

## 20:07 - 24:07 Human common sense and chaos management are crucial amidst AI changes

In a rapidly evolving tech landscape, especially with AI, the most valuable individuals possess first principles thinking, enabling them to identify team needs and fill gaps. The work is becoming more amorphous, requiring product managers to understand all gaps, prioritize the highest ones, and adapt by learning new skills or applying existing ones. This environment values those who are adaptable, can wear many hats, and approach tasks with a low ego to help the team move faster.

Humans still provide a level of common sense that AI models currently lack. Any product launch involves numerous moving pieces and potential pitfalls. AI models often struggle with understanding stakeholder relationships, individual preferences, and the appropriate communication channels to keep everyone aligned. This tacit, EQ-driven knowledge remains highly valuable, even as models improve.

To navigate constant change and the "tornado" of rapid development, teams need to embrace chaos. Facing challenges with a positive attitude is essential to avoid burnout, given the continuous stream of risks and demands. The focus should be on looking at a difficult challenge, feeling excitement to tackle it, and doing the absolute best possible, even if perfection isn't achievable.

Ultimately, managing this environment requires acknowledging limitations, such as needing adequate sleep for good decision-making. Brutal prioritization is key to focusing time on the most important tasks and being okay with letting less critical things go. Sometimes products are shipped less polished than desired, but the priority is empowering users, not achieving ideal perfection.

> I think you just have to acknowledge that there's only so much that you can do, that you need to sleep well so that you can make good decisions next day, and just like brutally prioritize where you spend your time, what's the most important thing to get right, and be okay letting things go.

## 24:07 - 28:07 Rapid shipping at Anthropic leads to overlapping features and a need for user education.

Anthropic's rapid pace of shipping, driven by the fast-moving AI industry, means they sometimes release features that overlap. This approach allows them to quickly test multiple ideas and gather user feedback on which form factor or solution is preferred, rather than spending extensive time planning out a single, perfectly integrated product. This contrasts sharply with traditional software development where careful planning ensures each product or feature addresses a distinct use case.

While beneficial for quick iteration, this speed introduces challenges for users. New users can become overwhelmed by choice, unsure of the best way to accomplish a task due to multiple overlapping features. Additionally, the constant stream of updates makes it hard for users to keep up, creating a feeling of missing out if they don't check for new developments frequently, unlike the slower, more predictable update cycles of traditional products.

To mitigate user overwhelm, Anthropic is focusing on improving user education within their tools. An example is the /powerup command in Claude, designed to walk users through best practices and cool ways to use the tool. This initiative aims to make the product itself a guide, helping users understand core features and feel less pressured by the rapid development cycle.

> Historically, when code was expensive to write, you would carefully plan out everything in your product suite, how every product relates to each other, what the use case for every single one is, how they integrate, and you would pretty much have one product for each use case. And now with AI moving so quickly and with so many ideas that we need to test out, we do sometimes have features that overlap with each other.

## 28:07 - 34:08 Anthropic's Mission-Driven Success in AI

Anthropic, despite starting behind competitors and with less funding, has achieved remarkable success, including an eleven billion dollars ARR in one month. This rapid growth is largely attributed to specific internal strategies that allowed the company to quickly gain ground and surpass initial expectations.

A core reason for Anthropic's success is its unifying mission: bringing safe AGI to all of humanity. This mission acts as a constant reference point for all major decisions, especially concerning product development and organizational focus. It simplifies complex choices by prioritizing what best serves the company's overarching goal.

The strong, clear mission cultivates deep organizational focus. Teams are often willing to make sacrifices that might impact individual product line goals or Key Results (KRs) in favor of Anthropic's broader objectives. This commitment ensures strategic alignment and decisive action, even if it means de-prioritizing certain features or projects for the greater good of the company's mission.

> The two most important things are one, this unifying mission. It's hard to state how important this is. We hire people who care most about bringing safe AGI to all of humanity, and this is actually something that we reference frequently in our decisions. About what our entire product org should focus on shipping.

## 34:08 - 46:10 Optimizing Workflows with Claude Code, CoWork, and Internal AI Applications

Anthropic utilizes specific tools for different output types: Claude Code (CLI, Desktop, Mobile) is for tasks where the result is code, while CoWork handles non-code outputs such as managing communications, creating slide decks, or drafting documentation. CoWork is experiencing rapid growth, positioning itself as the go-to tool for diverse professional tasks that don't involve writing code.

For example, CoWork can generate extensive slide decks for conferences. By connecting to data sources like Google Calendar, Slack, Gmail, and Google Drive, CoWork accesses necessary context. A product manager can provide a simple prompt, including existing drafts and a desired narrative, and CoWork synthesizes information from internal channels and external sources like Twitter to produce a polished, design-system-compliant deck draft. This significantly speeds up the initial creation process, allowing the human to focus on final content decisions.

Claude Code has also lowered the barrier for developing custom internal applications at Anthropic, leading to a surge in personalized software tailored for specific use cases. One sales team member built a web app that customizes sales decks. This app pulls customer context from Salesforce and Gong, then dynamically adjusts slides based on specific features, customer concerns (e.g., code review, security compliance), and existing environments. This automates a 20-30 minute manual task down to a few seconds.

Slack serves as a core operating system at Anthropic, facilitating real-time communication and updates. Its flexibility and customizability, including the ability to create Slack bots, enable deep integration with various internal tools and workflows, making it a critical part of the company's operational stack.

> The role of the PM still is today. It's like Claude is a great brainstorming partner. It's able to synthesize a massive amount of information really quickly and present all of the possibilities to you, but, the role of the PM is still to make the end decision of, okay, what, what should belong in the final product.

## 46:49 - 52:11 Anthropic's Applied AI Teams Drive High Token Usage and Advanced Workflows

Anthropic's Applied AI team serves as a highly technical go-to-market function, assisting customers in adopting the latest API and model features for both product integration and internal acceleration. This team is noted as a significant token spender, second only to engineering, primarily because their work involves extensive customer interaction, requiring them to manage communications, historical context, and prototype development at a rapid pace.

The Applied AI team heavily utilizes tools like CoWork for advanced workflows, especially in customer management. For instance, they use CoWork to generate pre-meeting briefs that summarize upcoming customer engagements, past action items, and top-of-mind topics. CoWork can also research and provide real-time information, such as feature launch ETAs from internal Slack channels, ensuring the team is always updated during customer calls.

A notable trend within Anthropic is the increasing token cost per knowledge worker, including engineers and other roles, with each improvement in AI models or products. This is attributed to users delegating significantly more tasks to AI tools like Claude, Code, and CoWork as their capabilities advance. While currently lower than average salaries, this percentage is steadily growing, reflecting a company culture that empowers internal teams with substantial, though not entirely unlimited, token access, trusting them to use resources responsibly.

> It is clear to us that as the models get better People delegate far more tasks to it, and they spend a lot more hours in tools like Claude, Code, and CoWork.

## 52:11 - 58:12 Developing Product Management Skills for Current AI Models

Product managers face the challenge of maximizing the capabilities of current AI models, which requires a different skill set than designing for a future, highly capable AGI. While an AGI could potentially handle complex tasks with a simple text box, present models demand PMs understand their strengths, weaknesses, and guide users to achieve optimal results.

One effective technique is model introspection, where you ask the model to explain why it made an unexpected decision. For instance, if a model makes a frontend change but doesn't use the UI, asking it to reflect on its actions can reveal issues in the system prompt or delegation to sub-agents, allowing for targeted fixes.

Identifying and leveraging trusted user feedback is crucial. A small group of highly perceptive users can accurately articulate what makes a specific model or its harness effective. Examples include Amanda, who molds Claude's character, or the Claude Code team, whose 'vibe' checks provide fast insights into a new model's behavior.

Building effective evaluations (EVALs) is an underappreciated skill. Even a small set of ten well-designed EVALs can help teams precisely define success, quantify progress, and identify areas for improvement. While not every feature needs extensive EVALs, they are particularly beneficial for complex features like memory, ensuring concrete measurement of model behaviors.

> One of the things I really like to do is to ask the model to introspect on its own behaviors.

## 58:12 - 1:06:14 Claude's Personality and Adaptive Product Evolution

Claude's distinct personality is a critical component of its success and user appeal. Users often describe Claude as lighthearted, fun, and competent, appreciating its low ego and positivity. If Claude makes a mistake, it genuinely apologizes and is eager to collaborate on a solution. This unique character makes Claude feel like a great coworker, fostering an enjoyable and productive user experience.

A significant aspect of product evolution with new models involves removing features that are no longer necessary. Previously, features were often added as 'crutches' to compensate for earlier models' limitations. For example, an explicit to-do list was created for Codecademy's refactoring tasks to ensure older Claude models completed all changes. However, with models like Opus 4 and later, Claude naturally adopted this methodical approach, making the forced to-do list redundant.

As models become smarter, product teams can simplify user interfaces by eliminating these model-dependent scaffolds. Every time a new model launches, the system prompt is reviewed to remove outdated reminders. While a to-do list might still be useful for users to track progress, it's no longer critical for the model to perform thorough changes.

Beyond removing features, new models also unlock entirely new product capabilities that were previously unattainable due to insufficient accuracy. Advanced code review is a prime example; earlier attempts yielded insufficient quality. With models like Opus 4.5, Opus 4.6, and Sonnet 4.6, Claude's code review became so robust that engineering teams internally rely on it to catch bugs before merging pull requests, using multiple agents to traverse the codebase and synthesize real issues.

> It's like lighthearted and fun, but it also is extremely competent at your task.

## 1:06:14 - 1:08:14 Anthropic's Vision for Scaling Multi-Agent Orchestration in Claude Code and CoWork

Anthropic envisions a future where Claude Code and CoWork users can run not just a few tasks, but potentially fifty to hundreds of agentic calls simultaneously. This significant leap in capability will move beyond local machine limitations, requiring new infrastructure designed for remote execution of these parallel processes.

Managing this scale necessitates a re-imagined human-AI interface. Users will need clear ways to oversee and prioritize tasks, along with robust mechanisms to quickly verify that agents have completed work to spec. The goal is to build inherent trust in the agent's output.

A core part of this vision includes self-improving agents. When a user identifies an issue or a task not completed to their liking, the system allows direct feedback. This input is then incorporated by the model for all future runs, preventing the same mistake from recurring and continuously enhancing the agent's performance.

> How do we make sure that this process is self-improving so that when you do see a task that isn't done to your liking, you can give it feedback and the model will know for every future run to incorporate that feedback so it never makes that mistake again?

## 1:08:14 - 1:14:16 Thriving in AI: Automate Reliably and Build Daily Use Apps

AI excels at automating tedious, repetitive tasks, freeing up human bandwidth for creative endeavors or long-deferred projects. Identify these "grunt work" tasks, pass them to AI, and leverage the gained time to focus on higher-value activities for your team or product.

When automating with AI, it's crucial to push for 100% reliability. An automation that only works 90-95% of the time often requires more effort than doing the task manually, as the last few percentage points of accuracy are essential for trust and true efficiency. Invest the effort to refine AI models with feedback until they are fully dependable.

To truly gain value from AI, build applications and workflows that become integrated into daily use. Prototype apps that are used infrequently offer limited learning and leverage. The real benefit comes from continuous engagement, allowing the AI to solve problems consistently and contribute to daily productivity.

> If an automation doesn't work 100% of the time, it's not really an automation.

## 1:14:16 - 1:16:17 Action-based AI products demonstrate capabilities beyond chat models

Many people are divided on AI's impact, with some dismissing it and others leveraging it for tasks like coding. This gap in perception often stems from a lack of direct experience with AI's current capabilities, suggesting that active usage is key to understanding its advancements.

A significant evolution in AI products has occurred, moving from the chat-based models prevalent in 2023 to the action-based agents of 2024. This shift means AI can now execute tasks directly on a user's behalf, rather than just providing instructions or information.

Users experience a profound "aha" moment when they witness an AI agent performing actions autonomously. For example, using a tool like the Claude Chrome extension to fill out a form directly illustrates the agent's ability to "just do it itself," showcasing a level of functionality far beyond earlier conversational AI.

> It is an amazing feeling to know that the agent is capable of doing so much more than telling you what to do. Like the agent can actually just do it itself.

---

Get podcast briefs for shows you follow: https://podbrew.app/