How Claude Code Was Built: From Rejected PR to 80% of Anthropic's Code

Developer at terminal with Claude AI generating code on multiple monitors

In a fascinating interview on The Pragmatic Engineer podcast, Boris Cherny — the creator and engineering lead behind Claude Code — reveals the origin story of one of the fastest-growing developer tools in the world. The conversation covers how a rejected pull request led to a paradigm shift, why Anthropic almost kept Claude Code internal, and what it means when AI writes 80% of your company’s code.

The PR That Got Rejected

When Boris joined Anthropic after seven years at Meta (where he was a Principal Engineer leading code quality across Instagram, Facebook, WhatsApp, and Messenger), he did what any engineer would do: he wrote his first pull request by hand. It got rejected.

"Adam rejected my PR and he was like, 'Actually, you should use this Clyde thing for it instead.' It took me half a day to figure out how to use the tool, but then it spat out a working PR. It just one-shotted it."

That moment — discovering that the AI model could produce a working PR faster than he could write one manually — was Boris’s first “wow moment” at Anthropic.

From Chatbot to Agentic Tool

Claude Code didn’t start as a coding agent. It started as a simple terminal chatbot that Boris built to learn the Anthropic API. The breakthrough came when he gave it a single tool: the Bash command.

"I asked it what music I was listening to, and it just wrote a little AppleScript program to query my music player and one-shotted it with Sonnet 3.5. The model just wants to use tools — that’s what I realized."

This insight — that you shouldn’t put the model in a box but let it use whatever tools it needs — became the founding philosophy of Claude Code. First Bash, then file editing, then more tools.

The Internal Debate: Keep It or Release It?

After Claude Code spread virally inside Anthropic (internal adoption chart was “just vertical”), there was a genuine debate about whether to release it externally or keep the competitive advantage internal.

"Dario had this question: 'How did it grow this fast? Are you like forcing people to use it?' And I was like, no, we offer this tool, people vote with their feet."

The decision to release came down to safety: Anthropic could study how the model behaves in the wild, learning from real-world usage patterns to make it safer. In hindsight, Boris says it was “totally the right decision.”

80% AI-Written Code and Zero Handwritten PRs

Today, Claude Code writes approximately 80% of the code at Anthropic. Boris personally ships 20-30 pull requests per day, with Opus 4.5 and Claude Code writing 100% of every single one. He doesn’t edit a single line manually.

This raises the obvious question: what does code review look like when AI writes everything? Boris explains that code review has shifted from reviewing code to reviewing intent — making sure the AI-generated changes align with the desired outcome and architectural decisions.

The Printing Press Metaphor

Boris compares the current moment to the invention of the printing press in the 1400s:

"There was a group of scribes that knew how to write. Some of the kings were illiterate who were employing the scribes. If you think about what happened to the scribes, they ceased to become scribes, but now there’s a category of writers and authors. The market for literature just expanded."

His argument: the skills that matter are changing, but the market for software is about to expand dramatically — just as literacy expanded the market for written content far beyond what scribes could have imagined.

Boris Cherny’s Daily Workflow

Perhaps the most practical takeaway from the interview is how Boris actually uses Claude Code day-to-day. His setup, which he describes as "surprisingly vanilla," would blow most developers’ minds:

10-15 Parallel Sessions

Boris runs 5 Claude Code sessions simultaneously in iTerm2 terminal tabs (numbered 1-5), with system notifications pinging him when a session finishes or needs input. On top of that, he runs 5-10 more sessions on the web, handing off local sessions to the browser and pulling them back with --teleport. He even starts sessions from his phone every morning.

Plan Mode First, Always

Every session starts in Plan Mode (Shift+Tab twice) — a read-only mode where Claude explores the codebase and reasons through the problem without making changes. Boris iterates on the plan until it looks right, then flips to auto-accept and lets Claude build it in one go.

MCP Integrations for Everything

Through MCP (Model Context Protocol) servers, Boris connects Claude Code to Slack (search and post), BigQuery (analytics queries), and Sentry (error logs). The team’s MCP configuration is checked into a shared .mcp.json file so everyone gets the same integrations automatically.

CLAUDE.md as a Living Rulebook

Boris’s team maintains a shared CLAUDE.md file that Claude reads at the start of every session. The rule is simple: any time Claude makes a mistake, they add a note so it never repeats it. The file compounds over time, making every future session smarter. The team contributes to it multiple times per week.

Additional Tips

  • Prefers Opus over Sonnet for complex tasks
  • Uses subagents for parallel workloads
  • Leverages slash commands and PostToolUse hooks for automation
  • Every mistake becomes a rule in CLAUDE.md — "correct once, never repeat"

The Bottom Line

This interview is required watching for any software engineer trying to understand where the profession is heading. The creator of Claude Code doesn’t write code anymore. He writes 30 PRs a day without touching a single line. His first handwritten PR at Anthropic got rejected. And the tool he built as a weekend hack now writes 80% of one of the world’s most advanced AI company’s codebase. Whether that’s inspiring or terrifying depends on whether you see yourself as the scribe or the author in Boris’s printing press metaphor.