Summary
Matthew Berman covers the accidental leak of Claude Code’s entire source code, exposed via a map file left in Anthropic’s npm registry. The leak — which accumulated over 22 million views on X within 24 hours — comprises roughly 500,000 lines of code across 2,300 files, giving developers, researchers, and competitors an unprecedented look at one of the most capable AI coding agents available. A Python rewrite of the codebase was quickly published, potentially complicating Anthropic’s ability to enforce copyright takedowns.
The video distills the most actionable findings from the leaked code. CLAUDE.md, a configuration file loaded into every single prompt, accepts up to 40,000 characters of custom instructions covering coding standards, architecture patterns, and team preferences — a largely underutilized lever most users ignore. The architecture is explicitly designed for parallelism: multiple sub-agents can run simultaneously while sharing prompt caches, dramatically reducing cost. Three named execution models are hardcoded in the source — fork (inherits parent context), teammate (separate tmux/iTerm pane with file-based mailbox communication), and worktree (isolated git branch per agent).
Berman also addresses Claude Code’s permission system, explaining that constant permission prompts indicate a misconfiguration failure rather than expected behavior. A settings.json file allows pre-authorizing specific commands, and Anthropic has shipped a smarter auto-permissioning layer that predicts user intent — effectively deprecating the previously popular “dangerously skip permissions” flag. For competitors and open-source maintainers, the leak provides a direct look at the prompt structures and agent orchestration patterns that make Claude Code perform well with Anthropic’s model family.
📺 Source: Matthew Berman · Published April 01, 2026
🏷️ Format: News Analysis
