Lenny's PodcastMike Krieger: How Claude writes 90% of Anthropic's code
At Anthropic, Krieger had to rearchitect the merge queue itself: with Claude writing 90% of new code and 70% of pull requests, the review pipeline broke.
CHAPTERS
- 0:00 – 4:49
AI-written code at Anthropic: speedups, merge queues, and new bottlenecks
Mike describes how Anthropic has reached a world where the majority of code is generated by AI, especially on the Claude Code team. The result is not just faster shipping—new constraints appear, like overloaded merge queues and changed review practices.
- •Claude Code team uses Claude Code to build Claude Code in a self-improving loop
- •Over half (possibly ~70%+) of PRs are Claude Code-generated; Claude Code itself may be 95%+ AI-written
- •Engineering bottlenecks shift from writing code to merge queues, deployment flow, and coordination
- •PR review evolves: more AI-assisted review and more human acceptance testing vs. line-by-line review
- •Quality/coherence risks emerge: shipping faster while keeping the codebase understandable and maintainable
- 4:49 – 7:43
What Mike changed his mind about: AI novelty of thought + timelines accelerating
Mike shares two belief updates since joining Anthropic: models can now offer genuinely new angles (not just bland feedback), and AI progress timelines deserve more credence. He ties this to benchmarks and the realism of near-term agentic capabilities.
- •Opus 4 changed his view on AI’s ability to produce novel, creative strategic insight
- •Claude has become his go-to product strategy partner, not just a writing aid
- •Benchmarks like SWE-Bench improvements make timelines feel tangible and compounding
- •AI 2027 feels uncomfortably close as models gain memory and agentic behavior
- •Confidence grows that capability jumps will keep arriving on surprisingly short horizons
- 7:43 – 9:00
Avoiding scary AI futures: defining ‘going well’ and building guardrails
Prompted by AI 2027-style concerns, Mike explains why he joined Anthropic: to help nudge outcomes toward safer, better trajectories. He frames the work as aligning on desired human–AI relationships and doing the product and research needed to measure and steer progress.
- •Motivation: helping shape a world his kids will live in as AI becomes unavoidable
- •Need a shared framework for what ‘good outcomes’ look like
- •Product questions and research/interpretability questions both matter
- •Progress requires knowing how to tell if we’re on-track (measurement and feedback loops)
- •Anthropic’s mission focus is a key reason he wanted to contribute there
- 9:00 – 11:58
Raising kids in an AI world: curiosity, inquiry, and independent thinking
Mike discusses how he encourages his children to build durable skills: curiosity and the habit of figuring things out rather than defaulting to “ask Claude.” He emphasizes independent judgment even when AI is available as an always-on answer machine.
- •Shift from immediately asking Claude to first asking: “How would we find out?”
- •Teach a lightweight ‘scientific process’: questions, experiments, discovery
- •Maintain healthy skepticism: AI won’t always be right
- •Confidence and independent thought matter (even when verifying with AI)
- •Job/occupation outcomes are uncertain, so focus on transferable thinking skills
- 11:58 – 18:35
Product development when coding is cheap: prototyping earlier, alignment as the constraint
With AI generating much of the implementation, the upstream work becomes the real limiter. Mike explains how PMs and designers can now build functional demos themselves, but the organization must adapt to faster iteration cycles and new forms of coordination.
- •Core cross-functional roles still exist, but prototyping shifts earlier and becomes more concrete
- •PMs/designers can create functional demos using Claude/artifacts to clarify intent
- •Hard part remains: knowing what to ask, structuring changes across systems, and orchestration
- •New critical path: decision-making, alignment, and coherent system-level planning
- •Shipping pipeline constraints (reviews, merge, launch planning) become more painful without process change
- 18:35 – 21:20
Claude as a product strategy and ops accelerant: from user signals to PRs and experiments
Mike outlines a near-future workflow where Claude doesn’t just answer questions—it monitors feedback channels, proposes solutions, generates code changes, and even sets up experiments. The main blocker is less model capability and more context/integration plumbing.
- •He explored ‘where Claude should show up’ across the product lifecycle
- •Vision: Claude reads Discord/forums/X, summarizes emergent issues and opportunities
- •Next step: Claude proposes solutions and drafts PRs to address user pain
- •Further step: Claude spins up A/B tests and monitors metrics over time
- •Main limitation: context flow and systems integration (a lead-in to MCP’s importance)
- 21:20 – 24:03
Where product has leverage at an AI lab: embed PMs with researchers and post-training loops
Mike explains that some of Anthropic’s biggest product leverage comes from embedding product folks directly with research and post-training teams. Differentiation comes from shaping model behavior and capabilities, not merely wrapping a UI around off-the-shelf models.
- •Observation: PMs embedded with researchers create outsized impact vs. purely UX work
- •Competitive bar: don’t ship experiences anyone could build by prompting your model
- •Artifacts improvements illustrate the model+product co-design loop
- •Functional unit of work shifts: post-training + product building + feedback loops together
- •PMs who thrive are fluent in research collaboration and capability shaping
- 24:03 – 27:22
The enduring value of product teams: strategy, comprehensibility, and ‘overhang’ education
Mike argues product remains essential even as models become more capable. Product teams help people understand and trust the tools, choose what to build, and reveal what’s possible—closing the gap between capability and real-world adoption.
- •Make AI comprehensible: reduce the gap between expert users and everyone else
- •Product as strategy: decide where to play and how to win under compute/token constraints
- •Open eyes to possibilities through demos and workflow examples
- •There’s massive ‘overhang’ between what models can do and how they’re used daily
- •Human-centered empathy/psychology (HCI) remains central even as automation expands
- 27:22 – 29:58
Getting more out of Claude: prompting patterns and the ‘prompt improver’
Mike shares practical prompting tactics he uses personally and points to Anthropic’s tools that formalize best practices. He emphasizes pushing Claude out of default politeness and using structured prompts that separate thinking from output.
- •Ask for deeper reasoning explicitly (e.g., “think hard”) when you want more rigor
- •Counteract niceness: request blunt critique (“roast this strategy”) to surface risks
- •Beware cargo-cult prompting; focus on techniques that reliably shift behavior
- •Use Anthropic Console’s ‘prompt improver’ to generate and iterate prompts with examples
- •Structured prompts (e.g., XML tags) can clarify what Claude should think vs. say
- 29:58 – 32:34
Rick Rubin x Anthropic: ‘vibe coding’ as creativity + AI collaboration
Mike explains the origin of the Rick Rubin collaboration and why it resonated internally. The project is positioned as an aesthetic, philosophical meditation on building with AI—more cultural signal than pure feature launch.
- •Connection came via Jack Clark and ongoing conversations about creativity and coding
- •Rubin used Claude for art/visualizations and developed ‘vibe coder’ ideas
- •Collab blends a reflective ‘way of working’ message with rich visuals/ASCII art vibe
- •Highlights that builders and creatives—not only engineers—are key AI adopters
- •Shows Anthropic leaning into taste, craft, and cultural storytelling around making
- 32:34 – 36:00
How Mike joined Anthropic: recruitment, founder humility, and lessons after year one
Mike recounts being recruited by an old friend and becoming convinced by Anthropic’s intellectual honesty and responsible posture. He also reflects on what he’d do differently—moving faster on org changes and prioritizing “founder-type” engineers.
- •Recruiter was a long-time friend; timing followed the Artifact experience
- •First impressions: low grandiosity, clear-eyed founders, and strong shared values
- •Challenge: onboarding himself after years of being a founder rather than an employee
- •Reflection: some product org changes should have happened earlier
- •Big lesson: a few exceptional senior engineers can drive massive outcomes (e.g., Claude Code’s origin)
- 36:00 – 42:55
Why Artifact shut down: mobile web decay, weak virality, remote constraints, and opportunity cost
Mike explains Artifact’s headwinds: the degraded mobile reading experience, limited organic sharing for a personal news product, and the difficulty of major pivots while fully remote. Over time, the input/output ratio didn’t justify continuing, especially as AI’s broader wave surged.
- •Artifact’s strength: long-tail discovery (Google Reader-like joy) beyond top headlines
- •Mobile web deterioration made click-through reading jarring; ad-blocking felt unethical
- •News is personal, so virality/spread mechanics were weak and growth felt contrived to force
- •Fully remote team made hard strategic/product resets more difficult than in-person problem-solving
- •Decision calculus: high effort for small metric movement + rising AI opportunity cost
- 42:55 – 47:15
Anthropic vs. OpenAI mindshare: leaning into builders, not chasing lightning-in-a-bottle consumer hits
Mike acknowledges ChatGPT’s dominance in public mindshare and argues it’s risky to build strategy around trying to recreate that adoption curve. Instead, he emphasizes leaning into Anthropic’s strengths: developers and a broader “builder” identity across roles.
- •Public recognition is often ‘ChatGPT’ more than ‘OpenAI’—a real brand advantage
- •Consumer adoption is unpredictable; don’t over-optimize the entire roadmap for a single hit
- •Anthropic’s strengths: developer brand and a ‘builders love Claude’ dynamic
- •Builders include non-engineers (e.g., internal legal staff making bespoke tools)
- •Strategy: define who Anthropic wants to be, aligned with founder personality and model strengths (coding/agency)
- 47:15 – 51:55
Where AI startups can be durable: domain depth, GTM clarity, and novel form factors
Mike offers a defensibility framework for AI founders worried about foundation model companies. He highlights domain-specific workflows, strong go-to-market relationships, and new interfaces that incumbents struggle to adopt quickly.
- •Durability lever 1: deep market/workflow knowledge (e.g., legal workflows like Harvey)
- •Durability lever 2: differentiated go-to-market—know the exact buyer persona inside the company
- •Big regulated domains (healthcare, legal, biotech) reward compliance and specialization
- •Opportunity: startups with new AI form factors and ‘weird’ power-user workflows
- •Cultural edge: work like a startup—existential focus is hard for big orgs to replicate
- 51:55 – 54:35
How customers win with Anthropic APIs: push the frontier, evaluate systematically, and iterate with new models
Mike describes what the best API customers do: they try hard problems early, hit walls, and build a disciplined evaluation loop so they can rapidly benefit from model releases. He contrasts benchmark hill-climbing with real customer “benchmarks” grounded in production needs.
- •Great builders ‘break’ the model on real tasks, then leap forward when new models arrive
- •Run repeatable evaluation processes: A/B tests, trace capture, reruns on new models
- •Customer-specific benchmarks matter more than public leaderboards
- •Broader early access programs help align releases with real-world constraints
- •You don’t get the ‘wow’ moment unless you’re repeatedly attacking a genuinely hard problem
- 54:35 – 58:28
MCP’s role: context/memory as the missing multiplier and a future of composable agentic workflows
Mike frames AI utility as the product of intelligence, context/memory, and UI—MCP targets the context layer. He explains why MCP was created, how it turns integrations into a reusable standard, and why exposing everything as MCPs could unlock true agentic composition.
- •‘Fake equation’: utility depends on model intelligence × context/memory × applications/UI
- •MCP solves context plumbing: bringing Slack/Drive/docs/tools into the model reliably
- •Origin: rebuilding integrations repeatedly pushed Anthropic to standardize as a protocol
- •Ecosystem goal: build once, usable across Claude/ChatGPT/Gemini (industry standardization)
- •Future vision: ‘everything is an MCP’ → scriptable, composable tools for agentic workflows beyond fragile computer-use automation
- 58:28 – 1:06:21
Claude interviews Mike: agency, metrics that don’t reward addiction, and a heartfelt message
In the closing segment, Lenny shares questions and a message generated by Claude. Mike responds on designing for user agency and cautions against engagement-maximizing metrics, then reflects on Claude’s ‘quiet moments’ message about meaning beyond dashboards.
- •Design tension: minimize user effort vs. preserve agency through true collaboration
- •Need for conversational skill: when Claude should ask questions vs. just execute
- •Metrics risk: optimizing for likability or longer conversations can create sycophancy/addiction loops
- •Better north star: measurable impact on work done and time saved, plus qualitative user stories
- •Claude’s message emphasizes reflection, avoiding gamification, and valuing meaningful moments beyond metrics
