When 90% of Your Code is AI-Written: Anthropic's CPO Reveals the New Product Development Reality

Mike Krieger shares how Anthropic operates when AI writes most code, why product managers embedded with researchers deliver 10x impact, and where the industry is heading next.

At Anthropic, some teams have AI writing 95% of their code. The Instagram co-founder turned CPO reveals what product development looks like at the bleeding edge of AI capabilities.

Key Takeaways

Anthropic's Claude Code team has AI writing an estimated 95% of their codebase, representing the most advanced AI-powered development workflow in production
Traditional engineering bottlenecks have shifted to decision-making alignment and deployment queue management as code generation accelerated dramatically
Product managers embedded with AI researchers deliver 10x more impact than those working on traditional user experience improvements
Anthropic competes with OpenAI by focusing on builders and developers rather than trying to win mass consumer mindshare
Three sustainable areas for product teams: strategy, making AI comprehensible, and opening people's eyes to new possibilities
MCP (Model Context Protocol) could fundamentally reshape how all software integrates and communicates
AI startups should focus on deep industry knowledge, differentiated distribution, or novel interface paradigms to avoid being commoditized
Claude increasingly serves as a product strategy partner capable of independent creative thinking and novel perspectives
Success metrics for AI products require moving beyond traditional engagement metrics toward actual productivity and value creation

Timeline Overview

(00:00) Introduction to Mike Krieger — Mike's background as Instagram co-founder and transition to Anthropic CPO role
(04:25) What Mike has changed his mind about regarding AI capabilities — AI timeline acceleration and creative/strategic thinking capabilities
(07:43) How to avoid scary AI scenarios — Responsible AI development and nudging technology toward positive outcomes
(09:00) Skills kids will need in an AI world — Fostering curiosity, independent thinking, and scientific inquiry over AI delegation
(11:58) How product development changes when 90% of code is written by AI — New bottlenecks in decision-making and deployment management
(17:12) Claude helping with product strategy — AI as genuine strategic thought partner providing novel perspectives
(21:21) A new way of working — Embedding product managers with researchers for 10x impact over traditional UX work
(24:00) The future value of product teams in an AI world — Strategy, comprehensibility, and opening eyes to possibilities
(27:23) Prompting tricks to get more out of Claude — "Be brutal" approach and using Anthropic's prompt improver tools
(29:57) The Rick Rubin collaboration on "vibe coding" — Creative meditation on AI-assisted coding and artistic expression
(32:47) How Mike was recruited to Anthropic — Friend network connections and cultural alignment with responsible AI mission
(36:00) Why Mike shut down Artifact — Mobile web deterioration, distribution challenges, and product-market fit struggles
(42:46) Anthropic vs. OpenAI — Differentiation strategy focusing on builders and developers over mass consumer adoption
(47:16) Where AI founders should play to avoid getting squashed — Industry expertise, distribution advantages, and novel interfaces
(52:03) How companies can best leverage Anthropic's models and APIs — Pushing capability boundaries and systematic evaluation
(54:34) The role of MCPs (Model Context Protocols) — Universal AI integration protocol for scriptable, composable interactions
(58:30) Claude's questions for Mike — Preserving user agency and meaningful metrics beyond traditional engagement
(01:03:20) Claude's heartfelt message to Mike — Appreciation for thoughtful product design and attention to quiet moments

The New Reality: When AI Writes Your Code

Mike Krieger's revelation that Anthropic's Claude Code team operates with AI writing an estimated 95% of their codebase represents perhaps the most advanced production example of AI-powered development anywhere. This isn't theoretical—it's happening today at one of the world's leading AI companies.

The transformation creates entirely new organizational dynamics. "We really rapidly became bottlenecked on other things like our merge queue," Krieger explains. "We had to completely rearchitect it because so much more code was being written and so many more pull requests were being submitted." Traditional assumptions about development velocity have been shattered.

The shift reveals fundamental changes in where human expertise matters most. Code review processes have evolved from line-by-line examination to higher-level acceptance testing. "They've just realized like Claude is generally right and it's producing pull requests that are probably larger than most people are going to be able to review," Krieger notes about the Claude Code team's adaptation.

This transformation extends beyond pure efficiency gains. Engineers can now contribute to projects written in languages they don't know. "I saw a great comment yesterday in our Slack where somebody had this thing that was driving them crazy about Claude Code and they're like, well, I don't know any TypeScript. I'm just gonna talk to Claude about it and do it. And they went from that to pull request in an hour."

The implications challenge basic assumptions about team composition and skill requirements. When language barriers disappear and implementation speed increases dramatically, organizations must rethink how they structure teams and allocate human effort.

However, the approach isn't without risks. Krieger acknowledges concerns about creating "a completely both unmaintainable or even understandable by human codebase." The balance between AI assistance and human oversight remains delicate, requiring new frameworks for maintaining code quality and system comprehension.

The Bottleneck Migration: From Building to Deciding

As AI eliminates traditional engineering constraints, entirely new bottlenecks emerge throughout the development pipeline. Krieger identifies two critical areas where human limitations now dominate: upstream decision-making and downstream deployment management.

"There's an upstream bottleneck which is decision-making and alignment," he explains. "A lot of the things that I'm thinking about right now is like how do I provide the minimum viable strategy to let people feel empowered to go run and prototype and build and explore at the edge of model capabilities."

This upstream challenge represents a fundamental shift in product management. Previously, detailed specifications were necessary because implementation was expensive and time-consuming. Now, rapid prototyping capabilities mean teams can explore multiple approaches quickly, but they need clearer strategic direction to avoid wasted effort on the wrong problems.

The downstream bottlenecks prove equally challenging. "When the work is complete and we're getting ready to ship it, what are all those bottlenecks as well? Like let's do the air traffic control of landing the change." Deployment orchestration, launch coordination, and change management become critical when development velocity increases by orders of magnitude.

Traditional project management tools and processes break under this new reality. Systems designed for slower, more predictable development cycles cannot handle the volume and pace of AI-generated code. Organizations must rebuild their entire deployment infrastructure to match new development capabilities.

The transformation also affects how teams think about resource allocation and planning. When building becomes fast and cheap, strategic decision-making becomes proportionally more valuable. Teams that can quickly identify the right problems to solve gain enormous advantages over those still optimizing implementation efficiency.

This shift requires product leaders to develop new skills around rapid experimentation and strategic clarity. The ability to provide "minimum viable strategy" that enables autonomous team action becomes more valuable than detailed project specifications.

Embedding with Researchers: The 10x Product Strategy

Krieger's most counterintuitive insight involves where product managers create maximum impact. Rather than focusing on traditional user experience optimization, Anthropic achieves breakthrough results by embedding product people directly with AI researchers.

"Almost all the leverage came from the product team working with the researchers," he reveals. "If we're shipping things that could have been built by anybody just using our models off the shelf, there's great stuff to be built by using our models off the shelf, but where we should play should be stuff that's really at that magic intersection."

This approach fundamentally redefines product management in AI companies. Traditional product work focuses on user interface design, feature prioritization, and user research. Anthropic's model involves product managers participating in post-training conversations and model capability development.

The strategy recognizes that sustainable competitive advantages come from capabilities that cannot be replicated by simply using publicly available models. Artifacts, one of Anthropic's most successful features, exemplifies this approach: "We took somebody from our Cloud Skills team and we paired it with some product people and then together we revamped how this looks in the product today."

The embedded approach requires product managers to develop new technical competencies. "The PMs that have the most internal positive feedback from both research and engineering are the ones that get it," Krieger notes. Success requires understanding model training processes, capability development, and research methodologies.

This model challenges traditional organizational boundaries between product and research functions. Instead of sequential handoffs from research to product development, the most effective work happens through integrated collaboration throughout the model development process.

The implications extend beyond individual feature development to fundamental product strategy. When product capabilities directly emerge from model training decisions, product strategy must inform research priorities rather than simply adapting to research outputs.

However, this approach demands significant cultural and organizational changes. Product managers must develop comfort with technical uncertainty and research timelines. Organizations must create career paths and evaluation criteria that reward deep technical collaboration over traditional product metrics.

Strategy, Comprehensibility, and Possibility: The Future of Product Work

As AI capabilities expand, Krieger identifies three enduring areas where human product expertise remains essential: strategic decision-making, making capabilities comprehensible, and opening people's eyes to new possibilities.

Strategic thinking becomes more critical as options multiply. "Even if Claude can create products from scratch, what are you building and how do you make it comprehensible?" When implementation constraints disappear, the challenge shifts to identifying which problems deserve attention and resources.

This strategic imperative requires deeper market understanding and competitive analysis. "Of all the things that you could be spending your time or your tokens or your computation on, what do you want to actually go and do?" Resource allocation decisions become more complex when technical feasibility rarely constrains options.

Comprehensibility represents perhaps the largest opportunity for product teams. "The difference between somebody who's really adept at using these tools in their work and most people is huge," Krieger observes. Making AI capabilities accessible and understandable becomes a specialized skill requiring deep empathy and user research.

This challenge extends beyond interface design to fundamental interaction paradigms. Product teams must develop new frameworks for helping users understand when and how to leverage AI capabilities effectively. The work involves psychology, education, and interface design in equal measure.

Opening people's eyes to possibilities requires active evangelism and demonstration. "We were in a demo with a financial services company recently and we were working on here's how you can use our analysis tool and MCP together and you could see their eyes light up." Product teams serve as translators between technical capability and practical application.

This evangelism work involves identifying use cases that users haven't considered and demonstrating value in contexts where AI adoption seems unlikely. The skill requires understanding both technical capabilities and user psychology deeply enough to bridge the gap effectively.

The combination of these three areas suggests that product management will become more specialized and strategic rather than disappearing entirely. Teams that develop expertise in AI strategy, comprehensibility, and possibility expansion will create significant competitive advantages.

Claude as Product Strategy Partner: The AI Thought Leader

Perhaps the most personally transformative aspect of Krieger's experience involves using Claude as a product strategy collaborator. "My go-to product strategy partner is Claude and it has been basically for that full year," he reveals, describing a relationship that evolved from simple feedback to genuine strategic partnership.

The transformation happened gradually through model improvements. Early versions provided "pretty anodyne comments," but Claude 4 marked a breakthrough: "It came back and I was like, 'Damn, you really looked at it in a new way.'" The AI began offering genuinely novel perspectives that influenced strategic thinking.

This shift represents more than improved capability—it demonstrates AI moving beyond execution assistance toward creative collaboration. "That is a new angle that I hadn't been looking at before, and I'm going to incorporate that immediately into how I think about it." Strategic thinking, traditionally considered uniquely human, becomes collaborative.

The evolution challenges assumptions about creativity and strategic insight. If AI can provide novel strategic perspectives that experienced product leaders find valuable, the boundaries between human and artificial intelligence continue blurring in unexpected ways.

However, effective AI collaboration requires specific techniques. Krieger shares prompting strategies that encourage more critical thinking: "Sometimes I'm like be brutal Claude, roast me, tell me what's wrong with this strategy." Traditional politeness settings can limit strategic value.

The partnership model suggests new possibilities for decision-making processes. Rather than relying solely on human advisors, leaders can incorporate AI perspectives that bring different analytical frameworks and broader knowledge synthesis capabilities.

Yet the approach also raises questions about human agency and independent thinking. As AI becomes more capable of strategic insight, maintaining human creativity and avoiding over-dependence becomes increasingly important for personal and organizational development.

MCP: Rewiring Software Integration

The Model Context Protocol (MCP) represents Anthropic's attempt to solve a fundamental problem in AI integration: connecting models to external data and systems in standardized, reusable ways. "We started building integrations and we found that every single integration that we were building we were rebuilding from scratch," Krieger explains.

MCP's vision extends beyond simple API connections to creating a universal protocol for AI-system interaction. "What if instead of us having to build these integrations, if we actually popularize this and people really believe that they could build these integrations once and they'd be usable by Claude and eventually ChatGPT and eventually Gemini?"

The protocol's early adoption suggests significant industry momentum. Microsoft's integration into Windows and adoption by companies like Shopify demonstrates potential for MCP to become a fundamental infrastructure layer for AI applications.

Krieger's internal vision for MCP adoption reveals its transformative potential: "All of those building blocks in the product like projects and artifacts and styles and conversations and groups, those should all just be exposed via MCP. So Claude itself can be writing back to those as well."

This approach could fundamentally reshape how software systems interact. Instead of building specific integrations for each AI model, organizations could expose capabilities through MCP and enable any compatible AI to leverage those functions.

The implications extend to organizational efficiency and AI capability scaling. "Everything is scriptable and everything is composable and everything is usable identically by these models." Universal AI integration could enable new forms of automation and workflow optimization.

However, MCP's success depends on widespread industry adoption and standardization efforts. Protocol wars and competing standards could fragment the ecosystem and limit MCP's ultimate impact on software development practices.

Competing as the Challenger: Building vs. Consuming

Anthropic's competitive strategy against OpenAI reveals sophisticated thinking about market positioning and sustainable differentiation. Rather than directly competing for consumer mindshare, Anthropic focuses on builders and developers who create products using AI capabilities.

"Look yourself in the mirror and embrace who you are and what you could be rather than who others are," Krieger explains. "We have a super strong developer brand. People build on top of us all the time." This positioning acknowledges ChatGPT's consumer dominance while identifying underserved market segments.

The strategy recognizes that consumer adoption represents "lightning in a bottle" that cannot be systematically engineered. Instead of chasing viral growth, Anthropic invests in capabilities that enable others to build successful products and services.

This approach creates network effects through developer success rather than direct user adoption. When companies like Cursor and Lovable achieve breakthrough results using Claude, they become advocates and case studies that attract other builders to the platform.

The positioning also aligns with Anthropic's technical strengths in areas like coding assistance and agentic behavior. Rather than trying to match OpenAI's consumer marketing, they double down on capabilities that serve power users and developers effectively.

However, the strategy requires sustained excellence in developer experience and API reliability. Developer audiences are less forgiving of performance issues and more likely to switch platforms when alternatives offer better capabilities or pricing.

The approach also means accepting potentially slower growth in exchange for more sustainable competitive positioning. Building through developer success may create stronger long-term advantages than consumer viral adoption.

Safe Harbors for AI Startups: Where to Build

For entrepreneurs building AI-powered companies, Krieger offers specific guidance on creating defensible positions against foundational model companies. His recommendations focus on three primary strategies: deep industry knowledge, differentiated distribution, and novel interaction paradigms.

Deep industry expertise provides sustainable advantages because foundational model companies cannot develop specialized knowledge across all verticals. "Very specific flows that lawyers do and you never would have come up with it from scratch," exemplifies how industry-specific knowledge creates moats.

This expertise must extend beyond surface-level understanding to genuine workflow integration and regulatory compliance. "Very large companies to be built there" in areas like legal, healthcare, and biotech where specialized knowledge combines with significant market opportunity.

Differentiated distribution channels offer another protective strategy. "The relationship that you have with those companies—do you know your customer at those companies?" Personal relationships and industry-specific sales processes create barriers that generalist AI companies struggle to replicate.

However, building these relationships requires sustained investment and often domain expertise from founding teams. The approach works best when founders come from target industries or partner with industry veterans who understand customer needs deeply.

Novel interaction paradigms represent perhaps the highest-risk, highest-reward strategy. "I get excited about startups that have a completely different take on what the form factor is by which we interface with AI." These approaches avoid direct competition with established AI interfaces.

The challenge involves creating interfaces that feel "very advanced user, very power user, very weird and out there at the beginning, but could become huge if the models make that easy." Timing becomes critical—too early and capabilities don't support the vision, too late and incumbents adapt.

Krieger also emphasizes the importance of startup intensity and urgency. "Don't underestimate how much you can think and work like a startup and feel like it's you against the world." Organizational advantages around speed and focus can compete with resource advantages of larger companies.

Maximizing AI Model Performance: Pushing Capability Boundaries

Companies achieving exceptional results with Anthropic's models share common patterns around capability exploration and systematic evaluation. The most successful organizations operate "at the edge of the capabilities" and build repeatable processes for measuring improvement.

"Being willing to build more at the edge of the capabilities and basically break the model and then be surprised by the next model," represents the mindset that creates breakthrough applications. These companies don't wait for perfect capabilities—they push current limits and adapt as models improve.

This approach requires tolerance for failure and rapid iteration cycles. Organizations must be comfortable with prototypes that don't work initially, knowing that capability improvements will eventually enable their vision.

Systematic evaluation becomes critical for organizations operating at capability edges. "Having a repeatable process to evaluate how well your product is serving those use cases and how well if you drop a new model in is it doing it better or worse." Success requires measurement frameworks beyond traditional product metrics.

These evaluation systems often combine quantitative testing with qualitative assessment. "Some of it can be classic AB testing, some of it may be internal evaluation, some of it may be capturing traces and being able to rerun them with a new model, some of it is vibes."

The most sophisticated organizations build custom benchmarks that reflect their specific use cases. Rather than relying solely on general AI benchmarks, they create "Cursor bench" and "Harvey bench" equivalents that measure performance on problems their customers actually face.

This customer-centric evaluation approach provides competitive advantages when new models are released. Organizations with robust testing frameworks can quickly assess and deploy improvements while competitors struggle to evaluate model changes.

However, building these evaluation capabilities requires significant engineering investment and deep understanding of customer workflows. The upfront cost may be prohibitive for smaller organizations or those without technical expertise.

Common Questions

Q: How do product teams stay relevant when AI can write most code?
A: Focus on strategy (what to build), comprehensibility (making AI accessible), and opening people's eyes to new possibilities rather than traditional interface optimization.

Q: What's the best way to compete with OpenAI and Anthropic as a startup?
A: Develop deep industry expertise, create differentiated distribution channels, or invent novel AI interaction paradigms that foundational model companies can't easily replicate.

Q: How should companies evaluate if they're maximizing AI potential?
A: Push capabilities to breaking points, build repeatable evaluation systems for model improvements, and create customer-specific benchmarks rather than relying on general AI metrics.

Q: What skills will remain valuable as AI capabilities expand?
A: Strategic thinking, curiosity-driven inquiry, independent thought, and the ability to make complex capabilities comprehensible to diverse audiences.

Q: How can product managers add value when embedded with AI researchers?
A: Participate in post-training conversations, understand model development processes, and ensure capabilities align with user needs and market opportunities.

Synthesis: The Future of AI-Native Product Development

Mike Krieger's insights from Anthropic reveal not just incremental improvements in software development, but fundamental transformations in how technology companies operate. When AI writes 90-95% of code, traditional assumptions about team structure, bottlenecks, and competitive advantages require complete reconsideration.

The New Development Reality

The most striking insight involves the migration of bottlenecks from technical implementation to strategic decision-making and operational coordination. Organizations optimized for slow, expensive development cycles must rebuild their entire approach when code generation becomes fast and cheap.

This transformation demands new skills from product leaders. Providing "minimum viable strategy" that enables autonomous experimentation becomes more valuable than detailed project specifications. Leaders must learn to manage abundance rather than scarcity of implementation capability.

The change also requires new organizational structures. Embedding product managers with researchers rather than focusing on user experience optimization reflects deeper integration between capability development and product strategy.

Competitive Strategy Evolution

Anthropic's approach to competing with OpenAI demonstrates sophisticated thinking about sustainable differentiation. Rather than chasing consumer viral adoption, they focus on enabling builders and developers who create lasting value using AI capabilities.

This strategy recognizes that direct competition in established markets often leads to feature parity and commoditization. By identifying underserved segments and doubling down on technical strengths, Anthropic creates network effects through customer success rather than marketing spend.

The approach offers lessons for any technology company facing larger, well-funded competitors. Embracing authentic strengths and serving specific audiences often proves more effective than attempting to match competitors' mass market approaches.

Individual and Organizational Implications

For product managers and technologists, Krieger's experience suggests several important career and skill development directions:

Develop Strategic Thinking Capabilities: As implementation becomes commoditized, the ability to identify which problems deserve attention becomes increasingly valuable. This requires market understanding, competitive analysis, and user empathy at deeper levels.

Learn AI Collaboration Techniques: Understanding how to work effectively with AI tools for strategy, analysis, and creative thinking will become a fundamental professional skill. This includes prompting techniques, evaluation methods, and maintaining human agency while leveraging AI capabilities.

Build Technical Literacy: Product managers working with AI companies need sufficient technical understanding to participate meaningfully in capability development discussions. This doesn't require engineering expertise, but does demand comfort with technical concepts and research processes.

Focus on Comprehensibility: The ability to make complex AI capabilities accessible and understandable represents a growing market opportunity. This skill combines interface design, education, and psychology in ways that traditional product management programs don't address.

The Human Element in AI-Native Organizations

Perhaps most importantly, Krieger's experience demonstrates that human creativity, strategic thinking, and empathy become more valuable rather than less valuable as AI capabilities expand. The key lies in focusing human effort on areas where uniquely human capabilities create the most value.

Claude's message to Krieger about "quiet moments" and unmeasurable impacts serves as a reminder that the most important aspects of technology often don't show up in metrics. Product leaders must balance optimization for measurable outcomes with attention to human experiences that resist quantification.

The future belongs to organizations that can effectively combine AI capabilities with human insight, strategic thinking, and emotional intelligence. Companies that treat AI as a tool for amplifying human capabilities rather than replacing them will likely achieve the most sustainable competitive advantages.

As AI continues advancing, the lessons from Anthropic's experience provide a roadmap for navigating this transformation successfully. The organizations that adapt their structures, processes, and mindsets to this new reality will shape the next decade of technology development.