Lenny's Podcast: Product | Career | Growth
Head of Claude Code: What happens after coding is solved | Boris Cherny
This debrief was written and voiced by AI from the episode's audio.
🎧 AI Debrief
🎙️ Full Episode
Summary
This episode features Boris Cherny, creator and head of Claude Code at Anthropic, reflecting on the one-year anniversary of the product's launch. What began as a solo terminal-based hack has grown to account for 4% of all public GitHub commits, with daily active users doubling in the past month alone. Boris shares the counterintuitive product decisions behind Claude Code's success — including deliberately under-resourcing teams, betting on future model capabilities rather than current ones, and embracing the terminal as a form factor precisely because it couldn't constrain the model's rapid improvement. The conversation covers Boris's bold claim that coding is 'largely solved,' what comes next (agentic AI for non-technical workers via Cowork), and the principle of 'latent demand' — observing how people misuse existing products to find where new ones should be built. He draws a compelling historical parallel to Gutenberg's printing press: a skill once confined to a tiny elite (scribes/programmers) democratizing to the entire population, enabling civilizational leaps no one could have predicted. Boris also touches on Anthropic's three-layer safety approach (mechanistic interpretability, evals, and real-world observation), advice for building AI products, and his personal philosophy — including the miso-making patience he developed while living in rural Japan and why the Anthropic safety mission pulled him back after a two-week detour to Cursor. The episode is both a product post-mortem and a forward-looking map of how AI will reshape every knowledge worker role in the near term.
Key Takeaways
Coding is 'largely solved' — 100% of Boris's code has been written by Claude Code since November, and Anthropic engineers ship 200% more pull requests year-over-year. Within months this will generalize across all codebases and tech stacks.
highBuild for the model six months from now, not today. Claude Code succeeded because Boris designed the product around where model capability was heading, not where it was at launch. Early poor product-market fit is acceptable if you'll hit the ground running when the next model drops.
highLatent demand is the most powerful product signal: when users hack or 'abuse' your product to accomplish something it wasn't designed for (e.g., people using Claude Code in a terminal to analyze MRIs or grow tomatoes), build a purpose-fit product for that use case. This is how Cowork was born.
highDon't box the model in. Avoid complex orchestration workflows and rigid step-by-step scaffolding. Give the model tools and a goal, then let it decide how to proceed. Scaffolding may add 10-20% performance but gets wiped out by the next model release anyway.
highUnder-resourcing teams slightly forces 'Claude-ification.' With fewer people, engineers are intrinsically motivated to use AI to ship fast. Pair this with giving everyone unlimited tokens at the start — the experimentation cost is trivial compared to salaries, and optimization can come later.
mediumThe next roles disrupted will be those adjacent to engineering — product management, design, data science — as agentic AI (AI that uses tools and acts in the world, not just converses) becomes accessible to non-technical workers.
mediumTo thrive in this era, be a generalist who crosses disciplines. The most effective people on Boris's team combine engineering with product sense, design intuition, or business acumen. AI-native skills matter, but curiosity and breadth matter more.
mediumNotable Quotes
“A hundred percent of my code is written by Claude Code. I have not edited a single line by hand since November.”
— Boris Cherny
Establishes the radical state of AI-assisted engineering today, coming from the creator of the tool itself.
“I have never enjoyed coding as much as I do today because I don't have to deal with all the minutia.”
— Boris Cherny
Counterintuitive take that removing the act of writing code has made coding more enjoyable, not less.
“Coding is largely solved. At least for the kinds of programming that I do, it's just a solved problem because Claude can do it.”
— Boris Cherny
The boldest claim in the episode — framing the core capability of software engineering as a finished problem.
“I think by the end of the year, everyone's going to be a product manager and everyone codes. The title software engineer is going to start to go away. It's just going to be replaced by builder.”
— Boris Cherny
Prediction about the near-term collapse of traditional engineering role boundaries.
“The model is improving so quickly. We felt that there wasn't really another form factor that could keep up with it. And honestly, this was just me kind of struggling with what should we build.”
— Boris Cherny
Reveals that the terminal form factor — now celebrated as a key success factor — was born from constraint and uncertainty, not grand strategy.
“In the 50 years after the printing press was built, there was more printed material created than in the thousand years before. I imagine a world where everyone is able to program — anyone can just build software anytime.”
— Boris Cherny
The printing press analogy is the most expansive framing of what AI-enabled coding could unlock for civilization.
“When I first started Claude Code, it got two likes internally. That was the sense of the reaction at the time.”
— Boris Cherny
Illustrates how even internally, transformative products can fail to generate early excitement — a useful data point for anyone building something novel.
Predictions
Claude Code will account for approximately 20% of all GitHub commits
— SemiAnalysis (cited by Lenny) · Confidence: medium · By end of 2025
Coding will be entirely handled by AI for the vast majority of engineers — no IDE needed
— Boris Cherny · Confidence: high · End of 2025 (already true for Boris since November 2024)
The title 'software engineer' will begin to disappear, replaced by 'builder'; everyone will be a product manager and everyone will code
— Boris Cherny · Confidence: medium · By end of 2025
AI agents will run for hours, days, or even weeks unattended — the 'babysitting' requirement will largely disappear
— Boris Cherny · Confidence: high · Next 1-2 years
In a year or two, learning to code will no longer matter for most people — AI will handle it entirely
— Boris Cherny · Confidence: high · 1-2 years
Token spend will exceed salary costs for some engineers/teams
— Boris Cherny · Confidence: low · Already emerging at Anthropic for some engineers spending hundreds of thousands per month
Agentic AI (AI that uses tools and acts in the world) will transform non-technical roles — PMs, designers, data scientists — just as it transformed engineering
— Boris Cherny · Confidence: high · Next 1-2 years
Action Items
Start using Claude Code's plan mode (Shift+Tab twice in terminal, or the plan button in desktop/web) for ~80% of tasks before letting it write code
Boris's top tip: letting the model plan first and then one-shot execute produces dramatically better results with fewer corrections, especially with Opus-class models.
Give your engineering team unlimited or very generous token budgets immediately — don't optimize costs until you've proven the use case works
Premature cost optimization kills experimentation. More capable models often use fewer tokens net because they require less correction. The experimentation cost is small relative to salaries.
Scan your product's usage data for 'latent demand' — users doing something your product wasn't designed for
This is Boris's single most important product principle. Facebook Marketplace, Facebook Dating, and Cowork all came from observing people hacking existing products. Your next big feature or product line is likely hiding in your analytics.
Build your AI product or feature for where the model will be in 6 months, not where it is today
Claude Code had poor early product-market fit but was positioned perfectly for the capability leap that came with Opus 4. Designing for the current model means your product is obsolete by the next release.
Try Cowork (Claude's desktop app) for non-coding tasks: managing email, filling out forms, syncing project status to Slack, canceling subscriptions
Boris uses it for all his team's project management and personal admin. Start by having it use one tool, then chain tools together (e.g., read email → send Slack message).
Deliberately under-staff AI-powered projects slightly to force the team to 'Claudify' their workflows
Constraint creates intrinsic motivation to use AI. Fully-staffed teams default to human labor even when AI would be faster.
Invest in becoming a cross-disciplinary generalist — an engineer with product sense, a PM who can code, a designer who understands systems
Boris sees these hybrid profiles as the most resilient and valuable as role boundaries collapse. AI handles the specialist execution; human judgment and breadth become the differentiator.