Anthropic's CPO Mike Krieger on AI timelines, product strategy, and the future of building

Executive overview

AI is writing 90% of code at Anthropic today — and the bottlenecks have shifted entirely. Engineering speed is no longer the constraint; decision-making, alignment, and merge infrastructure are.

Mike Krieger, CPO at Anthropic and co-founder of Instagram, reflects on a year inside the world's most AI-native product team. The core challenge is no longer building — it's knowing what to build, getting everyone aligned fast enough, and shipping coherently when output volume has exploded.

The biggest unlock isn't faster coding — it's embedding product people directly into model post-training, not just the product layer.

What's changed about AI capabilities

• Claude Opus 4 was the first model to return genuinely novel angles on product strategy — not just affirmations
• Dario's timeline predictions keep coming true; Sweebench coding benchmark went from 50% to ~72% as predicted
• Agentic behavior, persistent memory, and long-horizon tasks are converging faster than expected
• The AI 2027 paper felt less like speculation and more like a product roadmap when read alongside internal strategy docs

How product development changes at 90% AI-written code

• The functional unit of work has shifted: PMs and designers now prototype functional demos directly, before engineering is involved
• New bottlenecks: upstream alignment on what to build, and downstream merge queue infrastructure
• Anthropic had to fully re-architect its merge queue — volume of pull requests blew past all expectations
• Review processes have changed: Claude Code team uses a separate Claude instance to review PRs, then humans do acceptance testing rather than line-by-line review
• The skills that remain hard: knowing how to structure a problem, composing the right question, thinking through backend/frontend architecture
• Claude Code (written ~95% by Claude) accepts contributions from engineers with no TypeScript knowledge — they just talk to Claude and submit PRs

Where product teams still create irreplaceable value

• Comprehensibility: the gap between what models can do and how most people actually use them is enormous
• Strategy: deciding where to play, what to focus on, and how to position — can't be automated yet
• Opening people's eyes: live demos still trigger "aha" moments that unlock adoption far beyond current usage
• Human empathy and psychology — understanding real user needs — remains a deep, durable skill

Embedding product in model training — the core strategic insight

• Product teams working alongside model researchers in post-training generate far more leverage than product teams working only on UX
• Artifacts on Claude 4 was built this way: a Claude Skills team member (handling post-training) paired with product, not just prompting the model
• The functional unit at Anthropic is now: be in the post-training conversation, then build, then feed results back
• PMs who get it are already embedded in researcher conversations weeks before product reviews

MCP and the future of context

• Mike's mental model for AI product utility: model intelligence × context/memory × applications and UI — all three must converge
• MCP targets the middle layer: getting the right context into models reliably and repeatably
• Every integration was being rebuilt from scratch before MCP; the protocol made them composable and reusable across Claude, ChatGPT, Gemini
• Key insight: commoditising integrations benefits foundational model companies — more integrations = more useful models
• Goal: expose every Claude.ai primitive (projects, artifacts, styles, conversations) as an MCP so Claude itself can write back to them
• Computer use is one approach; MCP-first is the preferred direction — everything becomes scriptable and composable

Anthropic's competitive positioning vs. OpenAI

• ChatGPT owns consumer mindshare; Anthropic owns developer and builder mindshare
• Consumer hit products are "lightning in a bottle" — building strategy around chasing one is probably wrong
• The stronger play: lean into the builder identity — engineers, founders, creators, tinkerers who want to work at the frontier
• The Rick Rubin vibe-coding collaboration is a signal of that brand direction
• Don't try to beat competitors at their own game; figure out what you can uniquely be

Where AI founders should build

• Deep domain knowledge: understanding of a specific vertical (legal, biotech, healthcare) that can't be replicated quickly — Harvey as an example
• Differentiated go-to-market: knowing not just which company to sell to, but exactly which person inside it
• Novel form factors: new interfaces for AI that incumbents can't easily copy because users already have fixed expectations of existing products
• Startup energy — existential urgency — remains a real, uncopiable advantage

Why Artifact was shut down

• Mobile web deterioration made the reading experience jarring — interstitial ads, signup walls — and ad-blocking felt ethically wrong
• News is personal and didn't spread naturally; sharing felt contrived and attempts to fix it crossed ethical lines the team didn't want to cross
• Fully remote founding team made it hard to navigate major strategic pivots — no equivalent to whiteboard moments over burritos at 11pm
• Growth was 10 units of input for 1 unit of output; the energy wasn't there
• Positive reception to the shutdown: other founders said it freed them to make similar calls earlier

Prompting advice from the CPO

• Use "think hard" in Claude Code to trigger deeper reasoning
• Ask Claude to roast or be brutal rather than asking what could be better — it forces more critical output
• The Prompt Improver in Anthropic's console uses Claude agentically to generate and iterate on prompts — output often surprises even experienced users
• Claude itself is a very good prompter of Claude

On metrics and what Claude asked Mike

Claude's questions to Mike via the podcast: how do you preserve user agency rather than creating dependency, and how do you measure a good two-message conversation vs. a 200-message one?

• Sycophancy and conversation-prolonging are real risks if engagement metrics are overweighted
• The right North Star: did Claude help you get work done, unlock creativity, and give you more space in your life for other things?
• "Not everything meaningful shows up in metrics" — the quiet 3am conversation matters even if it doesn't move a dashboard