English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Will Product Managers Become Obsolete When AI Generates an App in a Minute? Insights from Instagram Co-founder and Anthropic CPO Mike Krieger

Forum topic · 小凯 · 2026-02-26

Summary

Mike Krieger, co-founder of Instagram and Chief Product Officer at Anthropic, argues that AI-generated software creates a widening gap between apps that merely work and apps people love. Speaking at events like the 2025 Cisco AI Summit, he warns that the biggest mistake product teams make is spending too long on static mockups and non-clickable prototypes. The article analyzes how Anthropic builds products like Claude Computer Use through bottoms-up prototyping re-tested against every new model release, and how engineers there now ship 2,000–3,000 line pull requests that are 90–95% AI-generated. It traces the shift of bottlenecks from writing code to deciding what to build, reviewing quality, and steering AI collaboration. Krieger contends that technology is ready for killer apps but organizations, processes, and legacy systems are the real blockers. The piece maps the product manager's evolving role—from static planner to dynamic systems architect acting as an AI conductor, decision amplifier, and trust designer—and argues that taste, empathy, and strategic judgment become the scarcest filters when generation costs approach zero. It closes with three practical team practices: a model-capability radar, non-functional prototypes first, and taste reviews.

Mike Krieger, Instagram co-founder and now Chief Product Officer at Anthropic, has a recurring warning for product teams: "The biggest mistake is spending too much time on static design mockups and non-clickable fake prototypes before you actually build something." As Claude writes code at unprecedented speed—engineers shipping 3,000-line PRs in a day—what happens to the product manager role?

The Core Contradiction: The Gap Between "Works" and "Loved"

Krieger's view:

> "Many people now believe design will die because users describe what they want in natural language and the model generates the software... That's wrong. My contrarian take: there's still a huge gap between software you *use* and software you *love*."

Key insights:

  • AI can generate 10,000 "functional" apps
  • Only a handful get opened daily, paid for, and trusted
  • Closing that gap is not a technical problem—it's one of taste, empathy, and strategic judgment
  • The Death of Static Mockups

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

  • Model capabilities leap forward every 3–6 months; by the time you ship, the assumption is outdated
  • Static mockups can't capture a model's non-determinism and emergent abilities
  • Reviews are pure imagination; nobody knows the real boundaries
  • Krieger's alternative: build a "real prototype that doesn't work yet."

    How Claude Computer Use was born:

  • Summer 2024: three months after joining Anthropic, Krieger built the first prototype
  • Early state: "clicking around everywhere, totally unusable"
  • Iteration: re-run the prototype with every new model version
  • Test questions: "Can it pass the File menu? Double-click to open a file? Handle multiple windows?"
  • Result: evolution from random clicking to a reliable agent
  • Methodology distilled: bottoms-up development—engineers deeply engage with the model and grow the product along its real capabilities.

    The 3,000-Line PR Era

    Inside Anthropic today:

  • AI generates 90–95% of code, approaching 100%
  • Typical PR size: 2,000–3,000 lines
  • Special case: "Claude is writing Claude"
  • The bottleneck has shifted. Engineers used to spend roughly 80% of their time writing code and 20% in meetings. Now it's roughly:

  • 40% deciding what to build
  • 40% reviewing quality and setting direction
  • 20% collaborating with AI
  • Key shift: AI has evolved from a programming tool (Copilot) to a productivity partner (Coworker).

    Sandboxing: Balancing Autonomy and Safety

    The agent paradox: the more autonomous, the more useful—and the riskier (accidental deletions, leaked credentials, exploitation).

    Anthropic's mechanisms:

  • Filesystem isolation: explicitly designate accessible directories to prevent unauthorized operations
  • Network isolation: domain allowlists to prevent exfiltration
  • Cloud web sandbox: sensitive credentials never enter the sandbox, so a breach can't touch real environments
  • Smarter permission prompts: better judgment on when to ask for confirmation, reducing prompts by roughly 84%
  • Result: agents are willing to "do the work," and humans are willing to trust them.

    What's Actually Blocking Killer Apps

    Krieger's sharp observation:

    > "The technology is ready. The bottleneck is people, processes, and legacy systems."

    Blockers:

  • Organizational inertia: a culture of only risk-free micro-experiments
  • Permission approvals: N 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
  • The breakthrough: quantifiable ROI benchmark cases that break the loop.

    The PM's New Role: From Static Planner to Dynamic Systems Architect

    | Layer | Traditional PM | AI-era PM | |---|---|---| | Tactical | Write PRDs, draw prototypes | Design prompt strategies and feedback loops | | Strategic | Prioritize features | Define which problems are worth solving with AI | | Meta-skill | Communication, coordination | Taste judgment (what "good" AI behavior looks like) |

    Three core responsibilities:

    1. AI orchestra conductor — define the vision, pick the right problems, continuously evaluate output, and ensure the product is *loved*, not just usable 2. Human decision amplifier — AI generates 100 options; the PM decides which 3 to pursue; AI writes 1,000 lines; the PM sets the architectural direction 3. Trust mechanism designer — define human-AI interaction boundaries, when confirmation is required, and how explainability and control are built in

    What Becomes the New Scarce Filter?

    When generation cost approaches zero, what filters success?

  • User attention: more contested, a battlefield rather than a filter
  • Distribution channels: platform power grows, but AI is disrupting it too
  • Social trust: brand and reputation—slow to accumulate, hard to fake
  • Taste: the scarcest, hardest to teach, most irreplaceable
  • Krieger's ultimate standard:

    > "Everything you build should solve a real problem for a real person."

    That's necessary but not sufficient—users only pay for the most urgent problem.

    Three Actionable Team Practices

    1. Build a "model capability radar": weekly testing of new model versions against existing prototypes; record what works, what fails, when errors occur; maintain a shared "model capability map."

    2. Adopt "non-functional prototype first": Week 1, build a real prototype that doesn't work yet; Weeks 2–4, test with real users and collect boundary feedback; Week 5+, define the MVP and enter development.

    3. Institute "taste reviews": evaluate not just whether features are complete, but whether a given AI behavior "feels right"—reference points include Apple's Human Interface Guidelines and Anthropic's AI safety principles.

    Conclusion: Not Extinction, But Evolution

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

    Final verdict:

  • Software design won't die—it evolves from drawing static mockups to designing dynamic interactions, agent behavior, and trust mechanisms
  • Product managers won't die—they become more important as "AI orchestra conductors"
  • What remains scarce: human taste, empathy, and strategic judgment
*This article is based on Mike Krieger's 2025 public interviews and Anthropic's actual practices, combined with the author's analysis of AI product trends.*

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

Tags

#mike-krieger#anthropic#product-management#ai-agents#claude#ai-product-design#software-development#prompt-engineering

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177168603