90% of Code Written by AI: What Product Changes

Mike Krieger

Chief Product Officer, Anthropic
Co-founder, Instagram

JUN 5 2025

The Shift

From 0% to 90% AI-Generated Code

"The team that works in the most futuristic way is the Claude Code team. They're using Claude Code to build Claude Code in a very self-improving kind of way. Probably over 70% of pull requests are Claude Code generated now."

From zero to 70% AI-written code at Anthropic in months
Claude Code team uses Claude to build Claude, self-improving architecture
Over half of all pull requests are now AI-generated
The codebase itself became the experimental ground for AI development

New Bottlenecks

What Changes When Code Writes Itself

Decision bottleneck: Deciding what to build and aligning everyone faster than engineers can code
Merge bottleneck: Infrastructure can't keep up with volume of pull requests
Review complexity: Line-by-line reviews impossible; need Claude-to-Claude review instead

70%+

of Anthropic PRs are Claude Code-generated

95%+

of Claude Code itself is written by Claude Code

The key insight

Engineers are no longer the constraint. Decision-making, alignment, and infrastructure are. The entire development process must be re-architected.

What Changed

PMs, Designers & Engineers in the Age of AI Coding

Designers now prototype: PMs and designers use Claude + Artifacts to build functional demos before engineering even starts
Earlier visibility: What used to take weeks to validate takes hours, then review happens much faster
Role shift: Engineers focus on architecture, prompt composition, backend/frontend decisions, not typing code
Review reimagined: Claude Code team uses Claude to review PRs instead of line-by-line human review

The hidden skill

Knowing what to ask the AI, how to compose a question, how to structure a change, these are now the specialized skills that still require experienced engineers.

New priority: strategy

"I need to provide the minimum viable strategy to let people feel empowered to go run and type and build at the edge of model capabilities.", Mike Krieger

Playbook

Build for AI Velocity

Empower with minimal strategy, not detailed specs, let AI and teams explore
Re-architect your merge queue and deployment infrastructure immediately
Use Claude-to-Claude for code review, human acceptance testing for validation
Hire founding engineers with strong opinions, not just individual contributors
Change your conception of software building every 6–12 months; don't hold on to old processes

The future playbookA year from now, the way we conceive of building and shipping software will be radically different. Trying to do this the old way will become very painful.

Contrarian

AI + Engineering Myths

✗AI means engineers do less workINSTEAD →✓ Engineers shift to harder problems: architecture, decision-making, and guiding AI. The work gets deeper, not lighter.

✗Prompting is a junior skillINSTEAD →✓ Composing the right question for AI is a highly specialized, senior engineering skill. It directly determines quality.

✗You need way more engineersINSTEAD →✓ You need fewer, but stronger founding engineers paired with great product and design. One strong person + AI is more powerful than before.

✗Product process stays the sameINSTEAD →✓ Everything changes: how you spec, how you review, how you deploy, who decides what to build. It's a complete reset.

Based on Mike Krieger's episode on Lenny's Podcast. All ideas on this page are from the episode.Watch on YouTube Follow @mikeyk on X

GO DEEPER IN THE EPISODE

04:25 What Mike has changed his mind about regarding AI capabilities 07:43 How to avoid scary AI scenarios 09:00 Skills kids will need in an AI world 11:58 How product development changes when 90% of code is written by AI 17:12 Claude helping with product strategy 21:21 A new way of working 24:00 The future value of product teams in an AI world 27:23 Prompting tricks to get more out of Claude 29:57 The Rick Rubin collaboration on “vibe coding”32:47 How Mike was recruited to Anthropic 36:00 Why Mike shut down Artifact 42:46 Anthropic vs. OpenAI 47:16 Where AI founders should play to avoid getting squashed 52:03 How companies can best leverage Anthropic’s models and APIs 54:34 The role of MCPs (Model Context Protocols)58:30 Claude’s questions for Mike 01:03:20 Claude’s heartfelt message to Mike

Every sketch here began as a transcript.

Send us yours and we’ll draw it. $1 a transcript.

Sketch mine →

0:00