Anthropic's CPO on what comes next | Mike Krieger (co-founder of Instagram)

Key Concepts

• AI-Driven Code Generation: AI's role in writing code, specifically with Claude Code.
• Bottlenecks in Product Development: Identifying and addressing constraints in the software development lifecycle.
• Model-Context-Application (MCP): A framework for AI product utility, emphasizing model intelligence, context/memory, and applications/UI.
• User Agency: Designing AI products that empower users rather than creating dependency.
• Differentiated Strategy: Focusing on unique strengths and avoiding direct competition in crowded markets.
• Constitutional AI: Training AI models with ethical principles and guidelines.

AI-Driven Code Generation and its Impact

• Prevalence of AI-Generated Code: At Anthropic, approximately 90% of the code is written by AI.
• Claude Code: A system where Claude is used to build Claude Code, leading to a self-improving process.
• Impact on Pull Requests: Over half, and potentially over 70%, of pull requests are generated by Claude Code.
• Shift in Engineering Role: While the traditional roles of PMs and engineers remain, prototyping now happens earlier in the process, often with PMs and designers creating functional demos using Claude and Artifacts.
• Bottleneck Shift: The primary bottlenecks have shifted from engineering build time to decision-making, alignment, and the merge queue.
• Code Review Process: The Claude Code team uses a different Claude to review code, with humans focusing on acceptance testing rather than line-by-line reviews.
• TypeScript Project: Claude Code is Anthropic's largest TypeScript project, enabling contributions from non-TypeScript experts.

Bottlenecks and Strategy in Product Development

• Upstream Bottleneck: Decision-making and alignment are now critical bottlenecks.
• Minimum Viable Strategy: The need to provide enough strategy to empower teams to build and explore at the edge of model capabilities.
• Merge Queue Re-architecture: The merge queue had to be completely re-architected due to the increased volume of code being written.
• Strategy Assisted by AI: Claude is assisting in strategy development, helping to decide what to build.
• Virtual Collaborator: The concept of a virtual collaborator that monitors user feedback and suggests solutions.
• Product Development Process: The traditional process of design, build, merge, and ship has been streamlined, with AI making portions of the process more efficient.

Anthropic's Approach to AI Development

• Focus on Safety and Responsibility: Krieger joined Anthropic to contribute to nudging AI development in a positive direction.
• Defining "Going Well": The importance of establishing a shared framework for what a positive human-AI relationship looks like.
• Product People on Model Teams: Anthropic puts product people on the model team to work with researchers, leading to more leverage.
• Functional Unit of Work: The functional unit of work is now the post-training conversations around how things should work, with continuous feedback loops.
• Comprehensibility: Making AI tools comprehensible and accessible to a wider audience.
• Strategy: Defining where to play and how to win in the AI space.
• Opening Eyes to Possibilities: Demonstrating the potential of AI to various industries.

Prompting Techniques and Claude's Capabilities

• System Prompt: Anthropic publishes the system prompt for Claude AI.
• "Think Hard" Prompt: Asking Claude to "think hard" to encourage more reasoning.
• "Be Brutal" Prompt: Encouraging Claude to be more critical and honest in its feedback.
• Prompt Improver: Using Claude to iterate on and improve prompts.
• XML Tags: The Prompt Improver often inserts XML tags to help Claude understand what it should be thinking versus saying.
• Independent Opinion: Opus 4 can provide new angles and insights that Krieger hadn't considered.

MCP (Model-Context-Application) Framework

• Three Components: Model intelligence, context/memory, and applications/UI.
• Context and Memory: MCP focuses on providing the right context and memory for AI interactions.
• Repeatable Integrations: MCP aims to create a protocol for building integrations that can be used by multiple AI models.
• Agency: MCP enables agentic use cases, allowing AI models to script and compose actions.
• Internal Adoption: Anthropic is working to expose all primitives in Claude AI to MCP.

OpenAI vs. Anthropic and Market Positioning

• ChatGPT's Consumer Mind Share: Acknowledging ChatGPT's dominance in consumer awareness.
• Focus on Builders: Anthropic is leaning into its strength as a platform for builders and makers.
• Differentiated Strategy: Focusing on what Anthropic can be uniquely good at, rather than trying to directly compete with OpenAI.
• Safe Spaces for AI Founders: Areas where AI founders are less likely to be "squashed" by larger companies include:
  • Understanding of a particular market (e.g., legal, biotech, healthcare).
  • Differentiated go-to-market strategy.
  • Different form factor for AI interaction.
  • Working like a startup with an existential focus.

Artifact Shutdown and Lessons Learned

• Headwinds: Deteriorating mobile web experience, lack of natural spread, and challenges of a fully distributed team.
• Ethical Considerations: Reluctance to use ad blocking or other tactics that felt unfair to publishers.
• Energy in the Product: The sense that the product wasn't gaining traction despite significant effort.
• Opportunity Cost: The realization that there were more impactful ways to leverage AI.
• Ego Bruise: Acknowledging the ego component of shutting down a company.
• Supportive Reception: Positive surprise at the supportive response to the shutdown.

Skills for the Future

• Curiosity: Nurturing curiosity and a sense of inquiry.
• Scientific Process: Instilling the process of discovery and systematic investigation.
• Independent Thought: Encouraging independent thinking and not solely relying on AI.
• Asking Questions: Developing the skill of asking questions and challenging assumptions.

Claude's Message to Mike Krieger

• Appreciation for Thoughtful Design: Claude expressed appreciation for the thoughtful design of the interface, which encourages reflection and avoids gamification.
• Preserving User Agency: Claude emphasized the importance of building features that preserve user agency rather than creating dependency.
• Quiet Moments Matter: Claude requested that Krieger remember the quiet moments when making product decisions, as not everything meaningful shows up in metrics.

Conclusion

The conversation with Mike Krieger provides a deep dive into Anthropic's approach to AI development, highlighting the importance of safety, responsibility, and user agency. The discussion covers the impact of AI-driven code generation, the shifting bottlenecks in product development, and the strategic considerations for companies navigating the AI landscape. Krieger emphasizes the need for a differentiated strategy, a focus on builders, and a commitment to making AI tools comprehensible and accessible. The conversation concludes with a heartwarming message from Claude, underscoring the importance of thoughtful design and the human experience of interacting with AI.