If You Want to Understand Claude Code, Start Here
A guide to the primary sources on Boris Cherny and what he built
I wrote Kudos to Boris Cherny and Claude Code — the terminal agent that used grep and glob instead of RAG, ignored IDEs entirely. Here’s a map to the primary sources for anyone who wants to go deeper.
Cherny doesn’t have a Medium blog. What he has is a trail of interviews, podcast appearances, Twitter threads, and — thanks to an accidental source code leak in March 2026 — 512,000 lines of TypeScript that speak for themselves. Below are the best primary and secondary sources, in rough order of importance.
The Man Himself
Latent Space podcast (May 2025) The foundational interview. Cherny and PM Cat Wu explain the Unix utility philosophy, why grep and glob beat RAG, the “do the simple thing first” principle, and how Claude Code was already writing ~80% of its own code at launch. This is where the agentic search decision gets explained directly, in Boris’s own words: “We tried RAG… eventually we landed on just agentic search. One is it outperformed everything. By a lot. And this was surprising.” Essential. The Latent Space Podcast — Claude Code
Lenny’s Newsletter podcast (February 2026) By this point Claude Code had 4% of public GitHub commits and daily active users doubling month over month. Boris talks about what comes after coding is “solved,” the latent demand principle (non-coders using a terminal tool to analyze MRIs and recover wedding photos), and how Cowork emerged from watching that happen. The best interview for understanding where he thinks this goes. Lenny’s Newsletter — Head of Claude Code
Pragmatic Engineer newsletter (March 2026) Gergely Orosz’s deep technical interview. This is where the Instagram origin story surfaces: Boris watched engineers at Meta navigate code effectively using grep when click-to-definition broke, and that planted the seed. Also covers the Meta code quality research, the shift from PRDs to prototypes, and the “always finish the migration” principle. The most technically detailed of the interviews. Pragmatic Engineer — Building Claude Code with Boris Cherny
Every.to podcast with Dan Shipper (October 2025) Focus on how Boris actually uses the tool day-to-day: parallel subagents, the code review slash command, making subagents argue with each other to eliminate false positives. Less philosophy, more workflow. Every.to — How to Use Claude Code Like the People Who Built It
Boris’s Own Posts
X/Twitter tips threads (January–February 2026) 42 tips across five threads, the closest thing to a written manifesto on workflow. Not a single article — scattered across dozens of posts — but someone built a fan site that assembles all of them cleanly: How Boris Uses Claude Code — fan-assembled tips site
The 259 PRs tweet (December 2025) The post that went viral: 30 days, 259 pull requests, 497 commits, 100% AI-written code, zero IDE opens. The moment that crystallized the “software engineering is changing” thesis. Worth reading for the reaction as much as the content. Boris Cherny on X
Hacker News thread on agentic search Boris made his most direct technical comments about the RAG decision here, in response to a developer asking about codebase indexing. The thread that spawned a hundred articles: Hacker News — Boris on agentic search vs. RAG
The Best Secondary Analysis
“Why Claude Code is special for not doing RAG/Vector Search” — Aram on Medium The best single article on the grep/glob vs. RAG decision. Goes beyond repeating Boris’s quotes to actually analyze the tradeoffs: token costs, the “indexing tax,” the bitter lesson alignment, debuggability. Includes a tradeoff matrix and honest skepticism about the performance claims. Why Claude Code is Special for Not Doing RAG
“Everyone Analyzed Claude Code’s Features. Nobody Analyzed Its Architecture” — Medium, Data Science Collective The post-leak architecture deep dive. Everyone else catalogued the Easter eggs (187 spinner verbs, Tamagotchi pets, frustration regex). This one asked what the code reveals about how Claude Code thinks. The punchline: developers paid Anthropic, per token, to understand Anthropic’s own product. Everyone Analyzed Claude Code’s Features. Nobody Analyzed Its Architecture
“Claude Code Doesn’t Index Your Codebase. Here’s What It Does Instead.” — Vadim Nicolai Traces the HN thread where Boris made his RAG comments, and goes deep on what agentic search actually means architecturally. Good comparison with Cursor’s embedding-based approach. Claude Code Doesn’t Index Your Codebase. Here’s What It Does Instead
The Leak
On March 31, 2026, Anthropic accidentally published the entire Claude Code source — 512,000 lines of TypeScript — via a source map file bundled into an npm package. A missing line in .npmignore. No hack, just human error.
The leak verified everything Boris had said publicly: no RAG pipeline, no vector database, Grep and Glob as first-class tools baked directly into the system prompt. On Grep specifically:
“ALWAYS use Grep for search tasks. NEVER invoke grep or rg as a Bash command.”
(Grep — capital G — is a dedicated first-class tool built into Claude Code with its own permissions and structured output. lowercase grep is the raw Unix shell command. Same underlying operation, very different levels of control.)
It also revealed an unshipped background daemon — currently called Chyros — that suggests Anthropic was already building beyond pure agentic search: not a retrieval index, but something proactive, building contextual awareness in the background before the user even issues a command.
Comprehensive Analysis of the Claude Code Source Leak
The leaked system prompt (preserved on GitHub — and effectively everywhere else by now)
Within hours of the leak, a developer rewrote the entire architecture in Python overnight using an AI orchestration tool, published it as “claw-code,” and hit 50,000 GitHub stars in two hours — the fastest-growing repository in GitHub history. It survived Anthropic’s DMCA takedowns because it wasn’t a copy of the TypeScript source; it was a clean-room reimplementation. The architecture is now permanently public.
The through-line across all of these: Cherny built something fast, let real usage teach him what mattered, and trusted the model to do more than anyone thought it could. The sophistication came after the foundation was proven. That’s almost always the right order.