what we *actually* learned from the claude code leak

Wait... I know that 'reg' is 'regular' but I've only ever said that out loud as rej-ex. I can't be the only one.

Any dev with more than 90 IQ knew most of what you said before the leak..

Literally all agents are just loops there are no other way to do it

Wait, what other feelings can the code detect? Can it be my therapist?

There's an entire multi-agent orchestration layer in there. Other fun things, three layers of remote context injection for "safety" and thinking depth settings that'll reduce the intelligence depth of the model without informing the user. It's interesting.

The underrated lesson from these leaks is not the exact wording, it is how much of the product is guardrails around the agent loop: when to ask, when to search, when to stop, and how to recover from bad context.

I'm more interested in how they're actually handling the tool use loops. that's where the real magic is.

It’s just echo.

YES!!!!

ELIZA: "How long have you felt this way about Claude?"

the system prompt structure in claude code is honestly a masterclass in prompt engineering. the way they layer context, use CLAUDE.md files for project-specific instructions, and handle tool orchestration is exactly how i set up my own workflows now. so much to learn from it

A very accurate summary !

It didn't mean anything to me, because it looks like a PR stunt.