How Anthropic’s product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code)
Cat Wu, Head of Product for Claude Code and Cowork at Anthropic, discusses how AI is transforming product management, emphasizing Anthropic's rapid shipping cadence, essential emerging PM skills like product taste and building for current model capabilities, and practical uses for Claude Code and Cowork.
Deep Dive Analysis
27 Topic Outline
Cat Wu's Role and Collaboration with Boris Cherny
Evolving PM Skills for AI-Native Products
Strategies for Rapid Shipping Velocity
Anthropic's Internal Processes and Culture
The Claude Code Source Code Leak Incident
Decision Behind OpenClaw Usage Restrictions
Anthropic's PM Team Structure and Organization
Merging Roles: Engineers, PMs, and Designers
Importance of Product Taste in AI Development
Human Brains' Enduring Value in AI Era
Coping with Constant Chaos and Change
Sacrifices Made for Rapid Shipping
Anthropic's Success Factors: Mission and Focus
Understanding Claude Code, Desktop, and Cowork Use Cases
Tips for Getting Started with Cowork and Practical Demos
PM Tech Stack and Internal AI Tools
Token Usage Across Anthropic Teams
Emerging Skills for AI Product Managers
The Value of Building Evals
Importance of Claude's Character and Personality
Adapting Products to New Model Capabilities
Vision for Claude Code and Cowork
Thriving in an AI-Driven World
The Importance of 100% Automation
Building Daily-Use AI Apps vs. Prototypes
The Divide Between AI Skeptics and Believers
Lightning Round
7 Key Concepts
AGI-Pilled
This term describes the tendency to build products based on the assumption of a super-intelligent Artificial General Intelligence (AGI) model, which can be easier than optimizing for the current capabilities of existing models.
Concept Corner
A designated part of a product suite or framework where engineers or PMs can rapidly launch new ideas and get them into users' hands, often within a week, to gather quick feedback.
Research Preview
A strategy for releasing early product features or ideas, clearly branded as experimental, to users. This reduces the commitment required for shipping and facilitates rapid iteration based on feedback.
Product Taste
The crucial skill of discerning what features are truly worth building and identifying the most delightful user experience. This becomes increasingly valuable as AI makes code generation more accessible and cheaper.
Model Introspection
The practice of asking an AI model to reflect on its own behaviors and mistakes. This helps developers understand what might have misled the model, allowing them to fix the underlying 'harness' or system prompt.
Evals
Quantitative evaluations used to measure the success rate of AI models on specific tasks. Building effective evals is important for defining product goals, tracking progress, and identifying areas for improvement.
Thinking Words
Internal prompts or 'verbs' used by AI models, such as Claude, to guide their reasoning and actions during a task. These were revealed in a source code leak and include terms like 'manifesting'.
9 Questions Answered
The PM role now emphasizes rapid iteration, shortening the time from idea to user, and defining key tasks that must work out of the box, rather than multi-quarter roadmap alignment, due to the accelerated pace of AI development.
Anthropic achieves rapid shipping by setting clear goals, using 'research previews' to reduce commitment, establishing tight cross-functional processes, and empowering engineers with strong product taste to ship end-to-end features.
Product taste is paramount, as deciding what to build and designing delightful user experiences becomes the most valuable contribution when AI makes code generation cheap and fast.
Humans still provide common sense, understand complex stakeholder relationships, and possess the tacit, EQ-driven knowledge that models currently lack, especially in the nuanced aspects of product launches and team dynamics.
Product consistency is sometimes sacrificed, leading to overlapping features or a greater need for user education. Users can also feel overwhelmed trying to keep up with the constant stream of updates.
Claude's lighthearted, competent, low-ego, and positive personality makes it a more enjoyable and effective co-worker, fostering better collaboration and user experience, which is core to its success.
New models often allow for the removal of 'crutch' features that were previously needed to compensate for model limitations, and they can unlock entirely new product capabilities that were not possible with older models.
The vision is to progress from making individual tasks successful to managing multiple tasks simultaneously (multi-clotting), eventually scaling to hundreds of tasks with robust remote infrastructure and self-improving agents that fully verify their work.
It's a philosophy that encourages individuals to understand constraints, deduce the right course of action, and execute quickly without waiting for strict role definitions or permission, learning from mistakes and fixing them as they go.
16 Actionable Insights
1. Prioritize Product Taste
As AI makes code cheaper to write, the most valuable skill for PMs is deciding what to build and designing the most delightful user experience, which requires strong product taste.
2. Build Products for Current Models
Focus on eliciting the maximum capability from current AI models, rather than solely designing for a future, super-intelligent AGI, as optimizing for today’s models is the harder and more impactful challenge.
3. Automate Repetitive Tasks
Identify and use AI tools like Claude Code or Cowork to automate manual, tedious parts of your job, freeing up bandwidth for creative work and new initiatives.
4. Achieve 100% Automation Accuracy
Invest the time and effort to make your AI automations work 100% of the time; a 90-95% accurate automation is not truly reliable and offers little value.
5. Build Apps for Daily Use
Focus on creating AI applications and workflows that you genuinely use every single day, as this provides real value, leverage, and continuous learning, unlike one-off prototypes.
6. Iterate Quickly on AI Products
Embrace a rapid shipping cadence, aiming to launch features weekly or even daily, by shortening the time from idea conception to getting the product into users’ hands.
7. Set Clear Product Goals
Define your key users, the specific problems you’re solving, and the top use cases to reduce ambiguity, especially when working with general LLMs, and enable faster decision-making.
8. Use Research Previews for Features
Ship early features in ‘research preview’ mode, clearly branding them as experimental, to reduce commitment, gather quick feedback, and iterate rapidly.
9. Establish Repeatable Launch Process
Create tight, low-friction workflows between engineering, marketing, and documentation teams so that engineers can ship features quickly with minimal overhead.
10. Understand Model Limitations via Introspection
Ask the AI model to reflect on its own behaviors and mistakes to understand what misled it, allowing you to identify and fix gaps in the model’s ‘harness’ or system prompt.
11. Build Great Evals
Develop a small set of 10-20 high-quality evaluations (evals) to quantify product goals, measure progress, and identify what the model or product is still missing.
12. Embrace Constant Change with Calmness
Acknowledge that P0-level issues will arise constantly; prioritize brutally, ensure good sleep, and be okay with shipping products that aren’t perfectly polished, knowing you’ll iterate quickly.
13. Build Products That Don’t Fully Work Yet
Prototype products at the edge of current AI capabilities; this helps identify missing elements, allowing you to quickly swap in newer, more capable models when they become available to close those gaps.
14. Connect All Relevant Data Sources for AI Tools
When using tools like Cowork, connect all relevant communication and data sources (e.g., Google Calendar, Slack, Gmail, Google Drive) to provide the AI with the necessary context for high-quality outputs.
15. Leverage AI for Non-Code Tasks
Utilize AI agents like Cowork for tasks where the output isn’t code, such as creating slide decks, writing documents, managing communications, or summarizing meetings.
16. Foster a ‘Just Do Things’ Culture
Empower individuals to understand constraints, deduce the right course of action, and execute quickly across team boundaries without waiting for strict role definitions or explicit permission.
7 Key Quotes
The hard thing is figuring out for the current model, how do you elicit the maximum capability?
Cat Wu
As code becomes much cheaper to write, the thing that becomes more valuable is deciding what to write.
Cat Wu
We want to remove every single barrier to shipping things.
Cat Wu
If an automation doesn't work a hundred percent of the time, it's not really an automation.
Cat Wu
It's pretty important to build products that don't necessarily work yet so that you know okay what is missing for this product to work and then with the newest model you can just swap it into the prototype you've already made and see okay does this new model close that gap.
Cat Wu
The 2024 generation of products were chat based and the cloud code generation of products is action based.
Cat Wu
I really like manifesting.
Cat Wu
2 Protocols
Anthropic's Rapid Feature Shipping Process
Cat Wu- Engineers develop a feature and dogfood it internally.
- They post the ready feature in the evergreen launch room.
- Docs, Product Marketing (PMM), and Developer Relations (DevRel) teams immediately engage.
- The marketing announcement for the feature is turned around the very next day.
Generating a Slide Deck with Cowork
Cat Wu- Connect all relevant data sources (e.g., Google Calendar, Slack, Gmail, Google Drive) to Cowork.
- Provide Cowork with a clear prompt detailing the desired slide deck, including narrative, suggested content, and any existing drafts or templates.
- Review Cowork's proposed outline and ideas, making decisions on what should be included in the final deck.
- Allow Cowork to generate the full slide deck, leveraging access to design systems for polished visuals.
- Provide feedback to Cowork for any necessary tweaks or refinements to the generated deck.