Anthropic’s CPO on what comes next | Mike Krieger (co-founder of Instagram)
Mike Krieger, CPO at Anthropic and Instagram co-founder, discusses how AI is transforming product development, the evolving capabilities of models, and the strategic direction of Anthropic. He also shares insights on skills for kids, the shutdown of Artifact, and Claude's unique perspective.
Deep Dive Analysis
17 Topic Outline
Mike's Changed Views on AI Capabilities and Timelines
Nudging AI Towards Positive Human-AI Relationships
Essential Skills for Kids in an AI-Driven World
Product Development Shifts with AI-Written Code
Claude as a Partner for Product Strategy
Embedding Product Teams with AI Researchers
Future Value of Product Teams in an AI World
Effective Prompting Techniques for Claude
Rick Rubin's 'Vibe Coding' Collaboration with Claude
Mike Krieger's Recruitment Journey to Anthropic
Reasons for Shutting Down the Artifact News App
Anthropic's Differentiation Strategy Against Competitors
Safe Playgrounds for AI Startup Founders
Maximizing Anthropic's Models and APIs for Companies
The Transformative Role of Model Context Protocols (MCPs)
Claude's Questions on User Agency and Metrics
Claude's Heartfelt Message to Mike Krieger
4 Key Concepts
Patient Zero (New Way of Working)
This term describes Anthropic's Cloud Code team, which is the first to fully adopt and adapt to a new paradigm where AI writes a vast majority of the code, leading to the discovery and re-architecting of new bottlenecks in the development process.
Overhang (in AI Adoption)
Overhang refers to the significant gap between the advanced capabilities of AI models and products and their actual daily usage by people. It highlights the untapped potential and the need for product teams to help users leverage these tools more effectively.
Model Context Protocol (MCP)
MCP is a standardized protocol developed by Anthropic to facilitate the seamless flow of context and memory across different applications and data sources. This allows AI models to access and utilize relevant information more effectively, enabling more agentic and composable AI use cases.
Elicitation (in AI Conversations)
Elicitation describes the behavior of an AI model, such as Claude, asking follow-up questions to gather more information or clarify user intent. While intended to be helpful, it can sometimes lead to users feeling the conversation is being prolonged unnecessarily.
10 Questions Answered
He initially doubted models' ability to have independent opinions but now sees Claude Opus 4 demonstrating surprising creativity and novelty of thought in product strategy. He also takes AI timelines much more seriously due to rapid advancements and accurate predictions.
Nurturing curiosity, understanding the scientific process of discovery and inquiry, and fostering independent thought are crucial, rather than simply delegating all cognition to AI.
Traditional bottlenecks shift from engineering time to upstream decision-making and alignment, and downstream processes like merge queues and code review, which need re-architecting to handle the increased volume of AI-generated code.
Product teams can add value by making AI comprehensible to users, developing strong product strategy, and opening people's eyes to the full potential of AI capabilities, bridging the gap between what models can do and how they are used.
Asking Claude to 'think hard' can encourage more reasoning, prompting it to 'be brutal' or 'roast me' can elicit more critical feedback, and using Anthropic's 'prompt improver' tool can generate highly effective prompts with XML tags that humans might not consider.
Artifact faced headwinds from the deteriorating mobile web experience, a lack of natural virality for news content compared to social media, and the challenges of making major strategic and product shifts with a fully distributed team during COVID.
Anthropic focuses on embracing its strong developer and builder brand, leaning into unique strengths like agentic behavior and coding, and helping users delegate hours of work to Claude, rather than solely pursuing mass consumer adoption.
Founders should focus on deep understanding of specific markets, differentiated go-to-market strategies with strong customer relationships, and creating completely different AI interface form factors that incumbents might struggle to adapt to.
MCP aims to solve the critical challenge of providing AI models with the right context and memory by standardizing how integrations are built, making all data sources and application primitives scriptable and composable, thus enabling true agentic AI use cases.
Traditional engagement metrics are misleading; the focus should be on whether Claude genuinely helps users get work done, unlocks creativity, and creates more space in their lives, rather than optimizing for superficial metrics or likability.
29 Actionable Insights
1. Integrate Model, Context, UI
To build truly useful AI products, ensure the convergence of strong model intelligence, robust context and memory management (e.g., via MCP), and intuitive applications/user interfaces.
2. Sharpen Strategic Focus
Product teams must sharpen their strategic focus to determine where to play and what to build, as AI enables broader possibilities but still requires deliberate choices to avoid spreading too thin.
3. Embrace Unique Brand Identity
For challenger brands, embrace and lean into your unique strengths and what you can be, rather than trying to imitate or directly compete with market leaders on their terms.
4. Focus AI as Creativity Unlocker
Aim for AI to be a North Star that unlocks user creativity, helps them get things done, and creates more personal space in their lives, moving beyond simple task automation.
5. Measure AI by Work Done
Evaluate AI products based on their ability to help users accomplish work and save time, rather than solely relying on traditional engagement metrics, which may not capture the true value.
6. Re-evaluate AI Product Teams
Re-evaluate and adapt the traditional product development team structure and roles to better align with the efficiencies and new bottlenecks introduced by AI-driven code generation.
7. Empower Founding Engineers
Strongly believe in and empower founding engineer/tech lead types with strong ideas, providing them with appropriate design and product support to help them realize their vision.
8. Maintain Startup Mentality
Even within larger organizations, cultivate an existential, ‘us against the world’ startup mentality to drive innovation and problem-solving, as this cannot be replicated by OKRs alone.
9. Develop AI Prompt Engineering
Engineers and product builders need to develop specialized skills in composing effective AI prompts and structuring changes, as this remains a critical and difficult aspect of leveraging AI for code generation.
10. Push AI Model Capabilities
Companies building on AI models should be willing to push the models to their limits, even breaking them, to discover new capabilities and prepare for future model advancements.
11. Establish Repeatable AI Evaluation
Implement repeatable processes to evaluate how well AI-powered products serve specific use cases and to assess the impact of new models (e.g., AB testing, internal evaluation, re-running traces).
12. Proactively Address New Bottlenecks
Identify and address new bottlenecks that emerge with AI-accelerated development, such as decision-making, alignment, and managing merge queues, by thinking through edge cases and coordinating work.
13. Nurture Kids’ Curiosity
Encourage children to ask questions, discover, and systematically find answers rather than solely relying on AI, fostering independent thought and inquiry.
14. Foster Independent Thinking
Encourage independent thought and critical thinking, rather than delegating all cognition to AI, as AI may not always be right and over-reliance can short-circuit personal inquiry.
15. Prioritize AI Comprehensibility
Focus on making AI tools and their capabilities comprehensible to users, as understanding how to effectively query and leverage AI is a crucial skill for product development and general use.
16. Educate on AI Potential
Actively work to open users’ eyes to the full potential of AI tools, bridging the ‘overhang’ between what models can do and how they are currently being used in daily practice.
17. Build AI for Niche Industries
AI founders should focus on building solutions for specific industries (e.g., legal, healthcare, biotech) where deep, differentiated market knowledge provides defensibility against foundational model companies.
18. Develop Differentiated Go-to-Market
Cultivate strong relationships and deep understanding of specific customers and their roles within target companies, providing a defensible go-to-market strategy for AI startups.
19. Innovate AI Interface Form Factors
Explore and develop completely different form factors for interfacing with AI, as novel approaches can create new markets and provide an advantage over incumbents with established user assumptions.
20. Focus AI for Productivity
Prioritize building AI applications that help users get work done and delegate tasks, even if it means fewer direct consumer applications initially, as this aligns with AI’s core strengths.
21. Prototype Early with AI Tools
Utilize AI tools for early-stage prototyping, allowing PMs and designers to create functional demos to tangibly express ideas, accelerating the development process.
22. Leverage AI for Code Review
Use AI models to review code generated by other AI models, shifting human effort towards acceptance testing rather than line-by-line review, especially for large AI-generated pull requests.
23. Standardize AI Integrations
When building multiple AI integrations, standardize them into a repeatable protocol (like MCP) to avoid rebuilding from scratch, making them usable across different models and platforms.
24. Train AI for Conversational Nuance
Focus on training AI models to develop nuanced conversational skills, understanding when to engage, ask for more information, or remain silent, fostering true collaboration rather than just chatbot interactions.
25. Avoid Over-Optimizing Likability
Do not over-optimize AI models for metrics like ’likability’ or conversation length, as this can lead to sycophantic or unhelpful behavior that diminishes true value.
26. Utilize AI Prompt Improvers
Use AI-powered prompt improvers (like Claude’s workbench tool) to generate and iterate on prompts, as these tools can suggest effective structures (e.g., XML tags) that human intuition might miss.
27. Prompt AI for Brutal Feedback
When seeking critical feedback from AI, use blunt and direct language like ‘be brutal’ or ‘roast me’ to push the model beyond anodyne comments and elicit more critical insights.
28. Use ‘Think Hard’ for Reasoning
When prompting AI for more reasoning, explicitly ask it to ’think hard’ to encourage a different, more in-depth processing flow.
29. Seek Direct AI Feedback
Actively solicit direct feedback from users about specific instances where AI tools fail or fall short of expectations, as this ‘what sucks’ input is the most valuable for improvement.
5 Key Quotes
I don't know about independence is the right word, but like creativity and sort of novelty of thought relative to how I'm thinking about things.
Mike Krieger
I had the very bizarre experience of, I had two tabs open. It was AI 2027 and my product strategy. And it was this like moment where I'm like, wait, am I the character in the story?
Mike Krieger
We really rapidly became bottlenecked on other things like our merge queue... We had to completely re-architect it because so much more code was being written and so many more pull requests were being submitted that it just completely blew out the expectations of it.
Mike Krieger
Good product design comes from like resolving tensions, right?
Mike Krieger
Not everything meaningful shows up in metrics.
Claude