If you missed it, Anthropic accidentally published Claude Code's entire source code. 512,000 lines. 1,900 files. Everything.
Here's what I found inside (and what you should do about it):
1. CLAUDE.md is re-read on every single query
This is the most important file in your entire setup. It's not set-and-forget. Treat it like a living document — project conventions, coding style, folder structure, what NOT to touch. The more context you give it, the better it performs.
2. There are 60+ built-in tools, most undocumented
Web search, sub-agent spawning, notebook editing, computer use, memory systems — all sitting there, switched off or barely documented. Most people are treating Claude Code like a chat window in their terminal. It's not.
3. It can run 5 agents in parallel
Sub-agents fork from the same point in your conversation, meaning you pay for one context window, not five. Security audits, refactoring, tests, docs — all at once.
4. There are 5 compaction strategies
When your context fills up, don't just let it auto-compact. Sub-agent offloads, forked context, and logic-only skill loading all exist. Most people use none of them.
3 things to do today:
Rewrite your CLAUDE.md with as much project context as possible
Set up project-level settings and permissions
Try parallel agents with sub-agent spawning
I broke the whole thing down in my latest video → HERE
Speak soon
Luke
