🎥 Watch the full episode here

📆 Published: 17-02-2026 🕒

Estimated Reading Time: 5 mins. Time saved: 45 mins! 🔥

Boris Cherny is the creator and lead engineer of Claude Code at Anthropic. Before joining, he led code quality across all of Meta's products - Facebook, Instagram, WhatsApp - where a 2% productivity gain represented a year's work by hundreds of engineers. He's also the author of "Programming TypeScript" (O'Reilly), a book he wrote before TypeScript was cool, and previously lived in rural Japan before deciding he wanted to be as close to frontier AI development as possible. At Anthropic, his team launched Claude Code, MCP (Model Context Protocol), and the desktop app - all from a single team called Anthropic Labs.

Speaking on Y Combinator's The Lightcone podcast, Boris drops a jaw-dropping stat: since Claude Code launched, productivity per engineer at Anthropic has grown 150%. Meanwhile, Semi Analysis reports that 4% of all public commits globally are now made by Claude Code, and Mercury's data shows 70% of startups choose Claude as their model of choice.

Why this matters now: Claude Code isn't just another AI coding tool - it's reshaping what "software engineer and product manager" even means. Boris's insights on building for the future model (not today's), the death of scaffolding, and the rise of agent swarms represent the sharpest available window into where product development is heading in 2026. Whether you're a founder, PM, or engineer, this interview has implications for how your role changes in the next 6-12 months.

"No one asked me to build a CLI," Boris explains. He needed to learn the API, so he built a little terminal chat app. Then tool use came out. He gave the model Bash. "I asked it, 'What music am I listening to?' It wrote some AppleScript to script my Mac. That was my first ever 'feel the AGI' moment. The model just wants to use tools. That's all it wants."

Two days later, engineers were using it without permission. When Dario asked if Boris was forcing adoption, the answer was no - "I just posted about it and they've been telling each other about it." It's a radically different bet from Windsurf's $3B IDE play (Pod Shot #93) or Replit's multiplayer approach (#67) - Boris built less UI while they built more, and all three worked.

Key Takeaways:

"We don't build for the model of today. We build for the model six months from now." When Claude Code launched, it wrote maybe 10% of Boris's code. He bet the model would improve - and it did exponentially. Today: "I uninstalled my IDE. It's just 100% Claude Code and Opus."

The uncomfortable flip side for founders: "You find PMF for the product right now, and then you're just going to get leapfrogged - because they're building for the next model." Pod Shot #115 (Why 'Vibe Coding' Will Kill SaaS) takes this to its extreme - if scaffolding gets wiped out every model cycle, the entire SaaS model gets wiped out too.

Key Takeaways:

In the Claude Code area at Anthropic, there's a framed copy of Richard Sutton's "The Bitter Lesson" on the wall. The core argument: general-purpose computation always beats specialised human engineering. In practice, this means scaffolding - the product engineering built around the model - has a limited shelf life.

"We could build a feature. Or we could wait a couple of months and the model can probably just do the thing instead." The result: "There is no part of Claude Code that was around six months ago. 80% of the codebase is less than a couple months old." That's a sharp contrast to LaunchDarkly's Zach Davis (Pod Shot #97), who argues institutional knowledge must be systematically preserved. Boris can afford constant rewrites at a 500-person company - enterprise reality is different.

Key Takeaways:

Boris drops "latent demand" repeatedly throughout the interview. "People will only do a thing that they already do. You can't get people to do a new thing." Every major Claude Code feature came from watching users hack their own solutions. Plan mode? "People were already telling Claude 'don't code yet.' I wrote it in 30 minutes on a Sunday night." CLAUDE.md? "People started writing markdown files for themselves and having the model read them."

The non-developer adoption was the biggest surprise. "There was that one guy using Claude Code to monitor his tomato plants. Another person recovering wedding photos off a corrupted hard drive." Internally, every designer was using it. The entire finance team was using it. "People were jumping over hoops to install a thing in the terminal" - which is what led to Co-work, Claude's non-developer interface. It's the same philosophy that built Samsara into a $50B+ company (Pod Shot #118) - extreme customer proximity as product strategy.

Key Takeaways:

The moment that shocked the Lightcone hosts - Interviewer: "Maybe in six months you won't need to prompt that explicitly?" Boris: "Maybe in a month." The technical reality is disarmingly simple: "All it does is add one sentence to the prompt - 'please don't code.' That's all it is."

The progression is clear: six months ago you had to babysit before and after the plan. Now it's just before. "So maybe the next thing is you just won't have to babysit at all." The irony? Even as plan mode dies as a feature, the discipline of planning becomes more important. Boris himself starts 80% of his sessions in plan mode despite declaring its death.

Key Takeaways:

Productivity at Anthropic has grown 150% per engineer since Claude Code launched. For context, at Meta, a 2% gain was a year's work by hundreds. Boris personally lands ~20 PRs per day, 100% written by Claude Code. His prediction: "The title 'software engineer' is going to go away. It's going to be maybe builder, maybe product manager." LinkedIn's already moving this direction (Pod Shot #114) - killing the APM programme in favour of "full stack builders."

The most sci-fi example: Claude Code's plugins feature was built entirely by an agent swarm over a weekend. "An engineer gave Claude a spec, told Claude to use an Asana board. Claude put up tickets, spawned agents, and they started picking up tasks." They call the orchestrating agent "Mama Claude." Boris calibrates sub-agents based on task difficulty - three, five, even ten agents researching in parallel.

It goes further than coding too: "A really common pattern is I ask Claude to build something. It'll look in the codebase, see some engineer touched something in the git blame, then it'll message that engineer on Slack asking a clarifying question. Once it gets the answer back, it keeps going." The Claudes talk to each other, talk to users on Slack, and Boris's Claude even tweets occasionally - "but I delete it because it's a little cheesy." Though remember agents aren’t always the right answer - as Pod Shot #109 argues, most organisations asking for agents actually just need structured workflows first.

Key Takeaways:

Boris sees a "very bimodal" distribution at Anthropic. Hyper specialists like the Bun team "understand JavaScript runtime systems better than anyone else." Hyper generalists span product and engineering, design and research. "Every single function on our team codes. Our PMs code, our designers code, our finance guy codes."

His hiring filter: "I really like to see people that just do weird stuff." He tells the story of Daisy, who didn't just add a feature - she built a tool that lets Claude Code test arbitrary tools, then had Claude write its own tool. And the interview question that reveals the most: "What's an example of when you were wrong?" Boris values people who can claim credit for mistakes over people with strong opinions. As Carl Vellotti explored in Pod Shot #111, the barrier to using Claude Code isn't technical skill - it's willingness to engage.

Key Takeaways:

Whether you're building AI products, managing teams, or just trying to keep up:

🔗 Links Referenced:

That’s a wrap.

As always, the journey doesn't end here!

Please share and let us know what you liked or want changing! 🚀👋

Alastair 🍽️.