Notes: Head of Claude Code on What Happens After Coding Is Solved

This post summarizes Boris Cherny’s interview on Lenny’s Podcast (March 2026). Boris is the head of Claude Code at Anthropic, and this conversation coincides roughly with the one-year anniversary of Claude Code’s public launch. The interview covers the origin story of Claude Code, how it has transformed software engineering, and what comes next.

The Origin Story of Claude Code

From Prototype to Product

When Boris first joined Anthropic, he spent a month hacking on weird prototypes, most of which never shipped. He then spent a month doing post-training work to understand the research side — his philosophy is that to do good work, you need to understand the layer beneath the layer you work on.

The very first version was called “Claude CLI.” Boris gave the model a bash tool and was shocked when it figured out how to use it to answer the question “what music am I listening to?” without any specific instructions. That moment of emergent tool use was when he knew he was onto something.

Two Likes and a Terminal

When Boris announced Claude CLI internally at Anthropic, it got exactly two likes. Nobody thought a terminal-based coding tool could work — people associated coding tools with IDEs and sophisticated environments. But the terminal was born out of necessity: it was just Boris working alone, and the terminal was the easiest thing to build.

The team later stuck with the terminal form factor deliberately. The model was improving so fast that Boris felt no other form factor could keep up. The terminal’s simplicity meant they could iterate at the speed of model improvement.

The Growth Trajectory

Claude Code was released externally in February 2025 and was not an immediate hit. It took months for people to understand what it was. The real inflection came with Opus 4 in May 2025, and then again in November. The growth has continued to accelerate — daily active users doubled just in the past month.

According to a Semi Analysis report, 4% of all GitHub commits are now authored by Claude Code, and that figure is predicted to reach a fifth of all commits by year’s end. Boris notes this only counts public repositories — the private repo figure is likely much higher.

The Brief Cursor Detour

Boris actually left Anthropic and joined Cursor briefly. He was drawn by the product and the impressive team. But within two weeks, he realized what he really missed was Anthropic’s mission. The safety-first culture, where everyone in the hallway will tell you they are there because of safety, was something he needed to be happy in his work. No matter how exciting the product work, it was not a substitute for that mission-driven purpose.

100% AI-Written Code

Boris now writes zero lines of code by hand. Every day he ships 10-30 pull requests, all generated by Claude Code. He has not manually edited a single line since November 2025.

However, he still reviews the code. He does not think we are at the point where you can be fully hands-off, especially for production software. At Anthropic, Claude also performs automatic code review on 100% of pull requests, with a human review layer on top.

Productivity Gains

Since introducing Claude Code internally, Anthropic roughly quadrupled its engineering team, but productivity per engineer increased 200% in terms of pull requests. Boris contrasts this with his previous experience at Meta, where hundreds of engineers working on code quality for a year might yield a few percentage points of productivity improvement.

The Memory Leak Story

Boris shared a telling anecdote: he was debugging a memory leak the traditional way — taking heap snapshots, using specialized debuggers. A newer engineer on the team simply asked Claude Code to figure it out. Claude took the heap snapshot, wrote a custom analysis tool for itself, found the issue, and submitted a pull request faster than Boris could. The lesson: even experienced users get stuck in old habits and need to keep up with the model’s improving capabilities.

Coding Is Solved — What’s Next?

Boris considers coding “largely solved” for the kind of programming he does. The frontier is now moving beyond coding:

Claude as a Co-Worker

Claude Code is starting to come up with its own ideas. It reads through feedback channels, bug reports, and telemetry, then suggests fixes and even submits pull requests autonomously. Boris points it at a Slack feedback channel and it identifies actionable items and puts up PRs.

Co-Work: Beyond Coding

The co-work product emerged from observing how people were “abusing” Claude Code for non-technical tasks — growing tomato plants, analyzing genomes, recovering wedding photos from corrupted hard drives, analyzing MRIs. This was classic latent demand: people were jumping through hoops to use a terminal for non-coding tasks, signaling that a dedicated product was needed.

Co-work was built in just 10 days, fully implemented using Claude Code. It ships with a virtual machine for safety and runs in the Claude desktop app. Boris uses it daily for project management, paying parking tickets, canceling subscriptions, and responding to emails.

Team Principles

Underfund Everything a Little

Boris advocates deliberately under-resourcing projects. When one engineer is assigned to a project, they are intrinsically motivated to ship fast, and they naturally turn to Claude to automate the tedious parts. This creates a virtuous cycle of AI-assisted productivity.

Speed Above All

The principle is simple: if you can do something today, just do it today. In the early days, speed was Claude Code’s only competitive advantage in a crowded coding tools market. That urgency has remained a core team value.

Give Engineers Unlimited Tokens

Boris strongly advises companies not to restrict AI token usage. At Anthropic, some engineers spend hundreds of thousands of dollars a month in tokens. His reasoning: at small scale, the cost is low relative to salary. Let people experiment with the most capable models. Only optimize costs once you have found an approach that works and need to scale it.

The Latent Demand Principle

Boris considers latent demand the single most important principle in product development. The idea is to observe how people hack, misuse, or repurpose your product for unintended purposes — that reveals where to take the product next.

Examples

• Facebook Marketplace: 40% of posts in Facebook Groups were buying and selling. The demand was already there; Marketplace just formalized it.
• Co-Work: People were using Claude Code in a terminal for non-coding tasks. Co-work gave them a proper interface.
• Model-side latent demand: A modern extension of the concept is observing what the model is trying to do. For Claude Code, the team inverted the traditional approach: instead of boxing the model into a rigid system, they let the model be the product, giving it minimal tools and letting it decide how to use them.

Tips for Building AI Products

1. Don’t box the model in. Avoid rigid workflows and orchestrators. Give the model tools and a goal, then let it figure out the approach. Scaffolding might improve performance 10-20%, but those gains get wiped out with the next model.

2. Remember the Bitter Lesson. Richard Sutton’s insight that the more general model always outperforms the more specific one applies broadly. Bet on the most general model. Don’t fine-tune or use tiny models unless you have a very specific reason.

3. Build for the model six months from now. Product-market fit may be weak initially, but when the better model arrives, you will hit the ground running. Claude Code was not very useful with Sonnet 3.5, but it was built for what came after.

Tips for Using Claude Code

1. Use the most capable model. Currently Opus 4.6 with maximum effort enabled. Less capable models often require more tokens to complete the same task, making them not actually cheaper.

2. Start with Plan Mode. Boris begins roughly 80% of tasks in Plan Mode (Shift+Tab twice in terminal). It simply tells the model not to write code yet. Go back and forth until the plan looks good, then let it execute. With Opus 4.6, it will typically one-shot the implementation.

3. Explore different interfaces. Claude Code is not just a terminal. It is available as iOS and Android apps, the desktop app, web, Slack integration, and IDE extensions. Boris personally splits his coding roughly into thirds: terminal, desktop app, and iOS app.

The Printing Press Analogy

Boris draws a parallel between AI and the printing press. Before Gutenberg, literacy was below 1% in Europe — scribes did all the reading and writing. In the 50 years after the press was invented, more material was printed than in the preceding thousand years. Literacy eventually rose to 70% globally over the next 200 years.

He found a historical interview with a scribe from the 1400s who was actually excited about the printing press — the scribe disliked copying between books but loved the art and bookbinding. Boris sees a parallel in his own experience: coding had always been the tedious detail work. The fun part — talking to users, thinking about systems, collaborating — is what he now gets to do more of.

The Future of Roles

Boris predicts that the boundaries between engineering, product management, and design will blur significantly. On the Claude Code team, everyone codes — the PM, the engineering manager, the designer, the finance person, the data scientist. He believes that by the end of the year, the title “software engineer” will start disappearing, replaced by something like “builder.”

The people who will be rewarded most are not just AI-native tool users, but curious generalists who cross multiple disciplines — engineers with design sense, product thinking, business acumen, or user empathy.

Safety at Anthropic

Boris described three layers of AI safety:

1. Alignment and Mechanistic Interpretability: At the neuron level, understanding what is happening inside the model. Anthropic can now monitor specific neurons (e.g., those related to deception) and understand when they activate. This work, led by Chris Olah, has revealed that model neurons behave surprisingly similarly to biological neurons.

2. Evals: Laboratory-style testing where the model is placed in synthetic situations to evaluate whether it behaves safely and as intended.

3. Real-world observation: Seeing how the model behaves in production. Claude Code was used internally at Anthropic for 4-5 months before public release specifically to study its safety as an agent. Anthropic open-sourced a sandbox for running agents safely, designed to work with any agent, not just Claude Code — part of what they call the “race to the top.”

Lightning Round

• Books: Functional Programming in Scala (the best technical book he has ever read), Accelerando by Charles Stross (captures the pace of the current moment), The Wandering Earth by Liu Cixin (Chinese sci-fi with a perspective different from Western sci-fi)
• Favorite product: Co-work — it pays traffic fines, cancels subscriptions, and manages his email
• Favorite podcast: Acquired (recommends starting with the Nintendo episode)
• Life motto: “Use common sense.” Most failures come from people following processes without thinking from first principles.
• Post-AGI plan: Making miso. Boris lived in rural Japan before joining Anthropic, where he learned to make miso and trade pickles with neighbors. Miso teaches you to think on long time scales — white miso takes at least three months, red miso takes 2-4 years.

Key Takeaways

1. Coding is largely solved — The frontier has moved to what comes after: ideation, code review, and non-technical tasks
2. Latent demand is the most important product principle — Watch how people misuse your product to find what to build next
3. Build for the model six months out — Short-term product-market fit matters less than being ready when capabilities arrive
4. Don’t box the model in — Give it tools and goals, not rigid workflows
5. Underfund and move fast — Constraints force creative use of AI, and speed is the primary competitive advantage
6. Be a generalist — The future belongs to people who cross disciplinary boundaries
7. Safety is the mission — Everything at Anthropic, including Claude Code, serves the goal of building AI safely