Claude Code’s source code just leaked through a public NPM package (via a sourcemap pointing to readable TypeScript on Anthropic’s servers), and I went through what people found so I could pull out the stuff that actually changes how you use the tool. The big mindset shift is this: Claude Code isn’t “Claude in your terminal.” It’s a full agent runtime—tool system, command system, memory, permissions engine, task manager, multi-agent coordinator, plus MCP client/server—wired into one execution pipeline. If you’re using it like a chatbot, you’re leaving most of the leverage on the table. Then I break down 8 practical insights. I cover the slash-command layer (there are ~85 commands, and most users know five), the real role of CLAUDE.md as operating context (not documentation), and why permissions are the reason Claude Code feels slow—plus how wildcard rules stop you from babysitting every action. I also explain what the architecture suggests about multi-agent coordination, MCP/plugins/skills as the true extension layer, and feature flags that hint at internal modes we can’t access yet. The punchline: top 1% users don’t just write better prompts—they design a better operating environment, manage context like money, and build repeatable workflows that compound.