Boris Cherny on What Grew His Career and Building at Anthropic
Boris Cherny, creator of Claude Code and former Meta principal engineer, shares how side projects, generalist thinking, and cross-org navigation shaped his career—plus insights on how AI tools are reshaping engineering productivity.
Key Takeaways
Career Growth Through Side Projects
Boris credits side projects as major career accelerators. His work on Undux (a simpler Redux alternative) led to 20-30 tech talks across Meta, building relationships and driving adoption. His TypeScript book came from noticing a gap—no one was writing about the language's advanced features.
Side projects reveal patterns: when you hit a problem, others likely share it. Look for signals in support forums and team struggles.
The Generalist Advantage
Boris prioritizes hiring generalists—engineers who code but also do product work, design, and talk to users. On the Claude Code team, product managers code, data scientists code, and everyone wears multiple hats.
At early-stage Facebook, he'd take prototypes to cafeteria workers for guerrilla user research. This generalist mindset compounds: you see problems others miss and ship faster.
See also: the-age-of-the-generalist
Latent Demand: The Core Product Principle
The single most important principle in product: find what users already do, then build for it.
Examples from Meta:
- Marketplace: 40% of Facebook Groups posts were buy/sell. Users hacked groups for commerce → Marketplace was the natural next step.
- Dating: 60% of profile views were opposite-gender non-friends. The intent existed → build the product.
You can never get people to do something they don't yet do. Find the intent and make it easier.
Cross-Org Collaboration is Hard
Working between Facebook and Messenger revealed deep cultural friction. Facebook wanted to ship fast; Messenger optimized for reliability and performance. Different goals (DAU vs SLAs), different org structures, different incentives.
His advice: find shared hypotheses both teams want to test. For deep integration, the common manager can't be at the VP level—you need closer organizational alignment.
Claude Code's Impact on Engineering
Even though Anthropic has tripled in size, productivity per engineer grew 70% because of Claude Code. Boris estimates 80-90% of Claude Code is now written by Claude Code itself.
A 12-engineer, 2-year project (like the Facebook Groups rewrite) would now take 5 engineers for 6 months. In 6 months, it might be one engineer.
His advice: don't build for today's models. Build for where models will be in 6 months.
See also: 2025-the-year-in-llms, ai-codes-better-than-me-now-what
Think in Types
The book that changed how Boris codes: Functional Programming in Scala. It teaches a different way to think about problems.
His core principle: type signatures matter more than the code itself. Get the types right, and clean code follows. Even in Python, thinking in types leads to better design.
Being Underleveled is an Advantage
Boris came into Meta underleveled and considers it lucky. Lower expectations gave him space to explore and build cool stuff without the pressure of checking promotion boxes. The momentum from exceeding expectations built his reputation.
How Boris Actually Uses Claude Code
In January 2026, Boris shared his personal workflow on X. Developers expected elaborate tooling—what they got was "surprisingly vanilla" but profoundly effective.
Parallel Sessions Strategy: Boris runs 5 Claude instances in numbered terminal tabs (1-5) using iTerm 2's system notifications to alert when any session needs input. On top of that, he runs 5-10 additional sessions on claude.ai/code. That's 15+ parallel Claude sessions working simultaneously.
He frequently teleports sessions between web and terminal using claude --teleport <session-id>, and kicks off morning sessions from his iPhone via the Claude iOS app.
Opus 4.5 with Thinking Mode for Everything: Boris's model choice is counterintuitive—he uses Opus 4.5 (the slowest model) for everything. His reasoning: "even though it's bigger and slower than Sonnet, since you have to steer it less and it's better at tool use, it is almost always faster than using a smaller model in the end."
The math: Sonnet might take 5 fast iterations (5 min total), while Opus takes 1 iteration (2 min total). The bottleneck is human steering time, not AI processing.
Shared CLAUDE.md Philosophy: The Claude Code team maintains a single CLAUDE.md checked into git. Every team member contributes multiple times per week. When Claude makes a mistake, they add the pattern to CLAUDE.md. Next session, Claude avoids it. This creates compounding knowledge—every mistake becomes a permanent lesson.
GitHub Actions for Continuous Learning: During code review, Boris tags @claude on coworkers' PRs to add learnings to CLAUDE.md as part of the PR. The learning loop extends into the review process itself.
His core philosophy: "There is no one correct way to use Claude Code. Each person on the Claude Code team uses it very differently."
Timestamps
- 0:00 - Intro and senior engineer promotion at Meta
- 3:30 - Cafeteria user research and generalist culture
- 6:00 - Latent demand explained (Marketplace, Dating)
- 10:00 - Cross-org collaboration nightmares
- 14:00 - Side projects: Undux and TypeScript book
- 22:00 - The TypeScript meetup and meeting "famous" devs
- 27:00 - Breaking arms, learning Haskell, thinking in types
- 32:00 - Being underleveled as an advantage
- 36:00 - Staff promotion via Comet rewrite
- 45:00 - Claude Code impact on engineering productivity