Will Product Managers Become Obsolete When AI Generates Apps in a Minute? Mike Krieger's Product Philosophy Explained

Forum topic · 小凯 · 2026-02-26

Summary

Mike Krieger, co-founder of Instagram and now Chief Product Officer at Anthropic, argues that AI-generated software will not kill product management—it will transform it. Speaking at the 2025 Cisco AI Summit, he warned that the biggest mistake teams make is over-investing in static mockups and non-interactive prototypes before building anything real. At Anthropic, AI now generates 90-95% of code, with pull requests of 2,000-3,000 lines, and Claude has effectively been used to help write itself. Krieger's approach favors bottom-up prototyping: build a real, even non-working prototype early, retest it against every new model release, and grow the product alongside actual model capabilities. He identifies taste, empathy, and strategic judgment—not technical skill—as the gap between software people merely use and software they love. Key bottlenecks for killer AI apps are now organizational inertia, approval processes, legacy systems, and trust deficits. The article outlines the emerging PM role as AI orchestra conductor, decision amplifier, and trust designer, plus actionable practices like a model-capability radar, prototype-first workflows, and taste reviews. Product managers will not disappear; they will evolve from static planners into dynamic system architects whose value lies in decision quality.

Anthropic Chief Product Officer and Instagram co-founder Mike Krieger has a recurring warning for product teams: "The biggest mistake is spending too much time on static design mockups and unclickable fake prototypes before you've actually built something." At the 2025 Cisco AI Summit and elsewhere, he laid out a philosophy for product building when AI can generate an app in minutes—and when Anthropic engineers ship 2,000–3,000-line pull requests in a day.

The Core Contradiction: "Usable" vs. "Loved"

Krieger pushes back on the claim that design is dying because models can generate software from natural language:

> "Many people now believe design will die... My contrarian take is that there is still a huge gap between software you use and software you love."

AI can produce ten thousand "usable" apps, but only a few earn daily opens, willingness to pay, and user trust. Closing that gap is a problem of taste, empathy, and strategic judgment—not technology.

The Death of Static Mockups

The traditional pipeline—write a PRD, sketch wireframes in Figma, get sign-off, hand off to engineers, wait three months—is collapsing because:

Model capabilities leap forward every 3–6 months, so slow shipping means building obsolete products
Static mockups cannot capture model non-determinism and emergent capabilities
Reviews are exercises in imagination; nobody knows the real boundaries

The Claude Computer Use Origin Story

| Time | Milestone | |------|-----------| | Summer 2024 | Krieger's third month at Anthropic: first prototype built | | Initial state | "Clicking all over the place, completely unusable" | | Iteration method | Re-run the prototype against every new model release | | Test questions | "Can it pass the file menu? Double-click to open a file? Handle multiple windows?" | | Result | Evolved from random clicking to a reliable agent |

His methodology: bottoms-up development—engineers interacting deeply with models and growing the product along real capability boundaries.

The 3,000-Line PR Era

Inside Anthropic:

AI generates 90–95% of code, approaching 100%
Typical PRs run 2,000–3,000 lines
In one notable case, "Claude is writing Claude"

The engineer's time allocation has shifted:

Before: ~80% writing code, ~20% meetings and communication
Now: ~40% deciding what to build, ~40% reviewing quality and setting direction, ~20% collaborating with AI

The key shift: AI has evolved from a "programming tool" (Copilot) to a "productivity partner" (Coworker).

Sandboxing: Balancing Autonomy and Safety

Agents face a paradox: the more autonomous, the more useful—and the riskier (deleted files, leaked secrets, exploitation). Anthropic's mechanisms:

| Mechanism | Implementation | Effect | |-----------|----------------|--------| | Filesystem isolation | Explicitly designated accessible directories | Prevents unauthorized operations | | Network isolation | Domain allowlists | Prevents exfiltration | | Cloud web sandbox | Sensitive credentials never enter the sandbox | Breaches can't reach real environments | | Smarter permission prompts | Intelligent judgment on when to confirm | ~84% fewer prompts |

Result: agents are willing to "do the work," and humans are willing to trust them.

What's Actually Blocking Killer Apps

> "The technology is ready. What's stuck is people, processes, and legacy systems."

| Barrier | Specific problem | |---------|------------------| | Organizational inertia | Culture that avoids even risk-free small experiments | | Approvals | Multiple layers of sign-off to access core systems | | Legacy systems | 20-year-old codebases nobody dares touch | | Trust deficit | No visible success stories, so nobody wants to go first |

Breaking the cycle requires quantifiable ROI benchmark cases.

The Product Manager's New Role

| Layer | Traditional PM | AI-era PM | |-------|----------------|-----------| | Tactical | Write PRDs, draw prototypes | Design prompt strategies and feedback loops | | Strategic | Prioritize features | Define which problems deserve AI | | Meta | Communication, coordination | Taste judgment (what counts as "good" AI behavior) |

Three core responsibilities:

1. AI orchestra conductor — define vision, pick the right problems, continuously evaluate output, ensure products are loved rather than merely usable 2. Human decision amplifier — AI proposes 100 solutions, the PM picks the 3 worth pursuing; AI writes 1,000 lines, the PM sets the architecture 3. Trust mechanism designer — define human-in-the-loop boundaries, explainability, and controllability

The New Scarcity

When generation cost approaches zero, what becomes the new filter?

User attention: more contested, not a filter—a battlefield
Distribution: platform power grows but is itself being disrupted by AI
Social trust: brands and reputation, hard to fast-track
Taste: the scarcest, hardest to teach, least replaceable

Krieger's ultimate standard: "Everything you build should solve a real problem for someone." Necessary—but not sufficient, since users pay only for the most urgent problem.

Actionable Team Practices

1. Build a model-capability radar: weekly tests of new model versions against existing prototypes; record what works, what fails, and when; share a team-wide capability map. 2. Prototype-first, even if non-functional: week 1, build a real prototype that doesn't yet work; weeks 2–4, test with real users and gather boundary feedback; week 5+, define the MVP from that feedback. 3. Add a "taste review": beyond feature completion, ask whether the AI behavior *feels right*—drawing on references like Apple's Human Interface Guidelines and Anthropic's safety principles.

Conclusion: Evolution, Not Extinction

| Dimension | Old era | New era | |-----------|---------|---------| | Core output | Static documents | Dynamic systems | | Key skills | Planning, communication | Taste, judgment, architecture | | Relationship to AI | Using a tool | Collaborating with a partner | | Source of value | Information integration | Decision quality |

Software design won't die—it will evolve from drawing static images to designing dynamic interactions, agent behavior, and trust mechanisms. Product managers won't disappear; they'll become more important as "AI orchestra conductors." What remains truly scarce: human taste, empathy, and strategic judgment.

*Based on Mike Krieger's 2025 public interviews and Anthropic's actual practices, combined with the author's analysis of AI product trends.*

Discussion question: How does your team define "success criteria" today—still mostly gut feel? Share your practices in the comments.