Summary: "Boris Cherny: How We Built Claude Code" (YC Light Cone Podcast)
Boris Cherny, the creator/lead engineer of Claude Code, joins the Y Combinator Light Cone podcast to share the origin story, design philosophy, and future vision of Claude Code.
Origin Story
- Boris joined Anthropic's "Labs" team, which produced three products: Claude Code, MCP, and the Desktop app
- It started as an accident — he built a simple terminal chat app just to learn the Anthropic API. Nobody asked him to build a CLI.
- He gave it the bash tool (literally copied from the docs), asked it to read a file, then asked "what music am I listening to?" — it wrote AppleScript to query his Mac. That was his first "feel the AGI" moment: "The model just wants to use tools. That's all it wants."
- 2 days after the prototype, teammates were already using it to code. Internal adoption went vertical — Dario asked if engineers were being forced to use it. They weren't.
Key Design Principles
-
Build for the model 6 months from now, not today. Early on, Claude Code couldn't really write code well. Boris wrote ~90% of his code by hand in Feb 2024. But they bet the model would get good at it — and it did.
-
Latent demand is everything. Every feature came from watching what users were already trying to do:
- CLAUDE.md — users were already writing markdown instruction files for the model
- Plan mode — users were already telling Claude "think about this but don't code yet." Boris built it in 30 minutes on a Sunday night, shipped Monday morning. "All plan mode does is add one sentence to the prompt: 'please don't code.'"
-
The Bitter Lesson — They have a framed copy of Rich Sutton's essay on the wall. Never bet against the model. Any "scaffolding" you build to improve performance 10-20% will be wiped out by the next model. So they constantly rewrite: "There is no part of Claude Code that was around 6 months ago."
-
The terminal was an accident — chosen because it was the cheapest UI to build. Boris expected it to last 3 months. He's been wrong — it's still the primary interface.
How the Team Works
- Boris considers himself an "average engineer" — uses VS Code, not Vim
- He's been 100% Claude Code, no IDE since Opus 4.5, landing ~20 PRs/day
- His personal CLAUDE.md is just 2 lines: (1) enable automerge on PRs, (2) post PRs to the team Slack channel. Everything else lives in the repo-level CLAUDE.md that the whole team maintains.
- Advice on CLAUDE.md: delete it and start fresh if it gets bloated. Add back only what you need. With each model, you need less.
Productivity Numbers
- Anthropic team doubled in size, but productivity per engineer grew 150% since Claude Code launched (measured by PRs + cross-checked with commits)
- Company-wide, 70-90% of code is written by Claude, many individuals at 100%
- At Meta, Boris's previous job, a 2% productivity gain took hundreds of people a year of work
Sub-agents & Claude Teams
- The plugins feature was entirely built by a swarm of agents over a weekend with minimal human intervention — one agent created an Asana board, spawned sub-agents that picked up tasks
- Boris uses plan mode for ~80% of sessions, running multiple tabs in parallel
- Plan mode has a "limited lifespan" — the model is learning to enter plan mode on its own, and soon may not need it at all ("Maybe in a month")
Co-work (Desktop App)
- Again, latent demand: people were using Claude Code for non-coding tasks (monitoring tomato plants, recovering wedding photos, finance). Designers, data scientists, and the finance team were all jumping through hoops to use a terminal tool.
- Co-work was built in ~10 days, 100% written by Claude Code. It's just Claude Code in a GUI wrapper with a VM sandbox and extra guardrails for non-technical users.
Advice for Founders
- Build for the model 6 months out — otherwise you'll find PMF today and get leapfrogged tomorrow
- Follow latent demand — make the thing people are already trying to do easier
- Never bet against the model (The Bitter Lesson)
- Hire for beginner's mindset — ask candidates "when were you wrong?" The best engineers can admit mistakes and think from first principles
- Expect code to be disposable — shelf life is now ~2 months
Bold Predictions
- Coding will be "generally solved for everyone" soon
- The title "software engineer" will evolve toward "builder" or "product manager"
- Everyone will code — PMs, designers, finance people (already happening at Anthropic)
- The scarier upper bound: ASL-4 (recursively self-improving models) and potential for catastrophic misuse, which Anthropic is actively working to prevent
