Head of Claude Code: What Happens After Coding Is Solved

Coding is solved. Now what?

Boris Cherny, creator and head of Claude Code at Anthropic, hasn’t edited a single line of code by hand since November. Every day he ships 10 to 30 pull requests, with five agents running in parallel while recording this podcast. Productivity per engineer has increased 200%. In his framing, coding as a bottleneck is largely over. The question is no longer whether AI can write code, but what happens when everyone can build software.

This conversation with Lenny Rachitsky traces the full arc of Claude Code: from an accidental prototype to a multi-billion dollar business responsible for 4% of public GitHub commits, with daily active users doubling month over month. But the real substance is in the product philosophy and the vision of what comes after coding.

The Two-Week Detour to Cursor

Boris briefly left Anthropic for Cursor, drawn by curiosity about a startup building the next editor. He lasted two weeks. The reason he came back was simple: at Anthropic, the feedback loop between product and model is uniquely tight. When he finds a limitation, he can walk to the model team and say “this is what users need.” That bidirectional loop between product insights and model training doesn’t exist at a company that consumes models as an API. This wasn’t about Cursor being bad; it was about the structural advantage of building product inside the lab.

From Terminal Hack to 4% of GitHub

Claude Code started as Boris’s personal productivity tool. He was already at Anthropic, needed something to help him code faster, and built a terminal-based agent in a few days. The original version was simple: Claude in a terminal with the ability to read and write files.

The early days were rough. Sonnet 3.5 could run for maybe 15 to 30 seconds before going off the rails. The product-market fit was poor. But the bet was deliberate: build for the model six months from now, not the model of today. When Opus 4 shipped, everything clicked. Growth went exponential and stayed there.

The progression was revealing: Sonnet 3.5 needed constant hand-holding. Opus 4.6 runs 10 to 30 minutes unattended, sometimes hours or days. Some tasks have run for weeks. The trajectory is clear: models that run autonomously for longer and longer periods.

Underfund Teams, Give Unlimited Tokens

One of the most counterintuitive product principles Boris describes: Claude Code’s team is deliberately kept small and underfunded. The logic is that a small team forces ruthless prioritization, while unlimited access to the best models enables massive individual leverage. Every team member uses Claude Code to build Claude Code.

The practical effect: instead of hiring more engineers, they give each engineer the most capable model with no token limits. Boris estimates this makes each person dramatically more productive than hiring would.

The Bitter Lesson, Applied to Products

Boris keeps Rich Sutton’s “Bitter Lesson” as required reading for his team. The core insight: the more general model always outperforms the more specific one. For product building, this translates into three principles:

Give the model tools, not instructions. Don’t over-curate context or put the model in a box. Give it tools to get what it needs. This consistently outperforms hand-crafted scaffolding.

Scaffolding gains are temporary. Custom workflows might improve performance 10-20%, but those gains get wiped out with the next model generation. It’s almost better to just wait for the next model.

Build for the model six months out. This is uncomfortable because product-market fit will be poor initially. But when the better model arrives, you’re already positioned. Boris gives this advice to startups frequently: bet on capability trajectories, not current limitations.

Cowork: The Next Surface

The conversation reveals how Anthropic thinks about product expansion. The sequence is intentional: coding first, then tool use, then computer use. Each step teaches the company about safety and capability boundaries.

Cowork is Claude Code’s sibling product: the same agent, but operating across your entire computer rather than just code. Boris uses it constantly. It has paid traffic fines for him, canceled subscriptions, responded to emails, and manages his team’s weekly status updates by messaging engineers on Slack who haven’t filled out their spreadsheet row.

The Chrome integration lets Claude observe and act within any browser context. Boris describes the adoption curve as similar to Claude Code a year ago: most people don’t understand what it can do until they see specific examples.

“We tried this experiment like a year ago and it didn’t really work cuz the model wasn’t ready, but now it actually just works.”

Three Principles for Every New Team Member

Boris shares three principles he gives every person joining his team:

Speed is everything. Ship fast, iterate fast, don’t let perfection slow you down. The pace at which they address user feedback surprises people, and that speed is itself a competitive advantage.

Talk to users obsessively. Boris describes his ideal day as talking to users and making the product better for them. Feature development is almost entirely feedback-driven. He actively searches Twitter and Threads for bugs and feature requests, often fixing them within minutes using Claude Code.

Use common sense. His single most repeated piece of advice. Most failures he sees in work environments come from people following process without thinking, working on products that aren’t good ideas, or following momentum instead of first-principles reasoning.

Pro Tips for Using Claude Code

Boris offers practical advice, with the caveat that there’s no one right way:

Use the most capable model. Currently Opus 4.6 with maximum effort enabled. The counterintuitive insight: less capable models often cost more total tokens because they need more correction and hand-holding.

Start in plan mode. Boris starts roughly 80% of tasks in plan mode, which is technically just one sentence injected into the prompt: “please don’t write any code yet.” Once the plan looks good, auto-accept edits because with Opus 4.6, it one-shots the implementation almost every time.

Explore different interfaces. Claude Code isn’t just a terminal. It runs on iOS, Android, desktop apps, Slack, and web. The same agent powers all surfaces.

On Codex and Competition

Boris hasn’t deeply used OpenAI’s Codex, noting it looked a lot like Claude Code, which he found flattering. His stance on competition is deliberately detached: the Claude Code team doesn’t spend much time looking at competing products. The focus is entirely on solving user problems.

“We don’t spend a lot of time looking at competing products. We don’t really try the other products. I love talking to users. I love making the product better.”

The Title “Software Engineer” Is Going Away

Boris makes a sharp prediction about the near future of professional identity in tech. If everyone codes, then the title “software engineer” loses meaning. It gets replaced by “builder.” Product managers should be sweating too: by the end of the year, everyone’s going to be a product manager and everyone codes.

The implication is that the value shifts from writing code to taste, judgment, and the ability to direct AI toward the right problems. What Boris calls “ask not what the model can do for you,” meaning: think about how to give the model the tools to do things, rather than trying to do things yourself.

Post-AGI: Miso

Before Anthropic, Boris lived in rural Japan as the only engineer and English speaker in his town. The social fabric was organized around seasons and fermentation. White miso takes three months. Red miso takes two to four years. The pace was the complete opposite of San Francisco.

It was during this period, reading science fiction and thinking on long timescales, that Boris decided he needed to contribute to AI going well. That’s what brought him to Anthropic, partly through Ben Mann’s influence.

His post-AGI plan? Go back to making miso.

A Few Observations

This conversation is valuable not for any single revelation, but for the coherence of Boris’s worldview. A few threads worth pulling on:

• The “build for the model six months out” principle is the most actionable insight for anyone building AI products. It explains why so many AI startups feel brittle: they’re optimizing for today’s model instead of riding the capability curve.
• The deliberate small-team-plus-unlimited-tokens approach is a genuine organizational innovation. Most companies respond to opportunity by hiring. Anthropic’s Claude Code team responds by giving each person more leverage.
• Boris’s journey from rural Japan to heading a multi-billion dollar product line is the kind of story that only makes sense in a moment of genuine discontinuity. A year ago, Claude Code didn’t exist. Now it accounts for 4% of GitHub commits.
• The safety framing is subtle but important: coding → tool use → computer use isn’t just a product roadmap. It’s a safety research methodology. Each step teaches Anthropic about the boundaries of autonomous AI in increasingly unconstrained environments.
• The miso metaphor lands: you mix it up, then wait. Sometimes the most important skill is patience on long timescales while everything around you accelerates.