Every hard rule in your CLAUDE.md is one more thing Claude Code has to reason about before it gets to the task. While older models needed that kind of guidance, the latest ones mostly don't. Running /doctor in your project flags the stale rules that are taking up space and eating tokens, and trimming them tends to make the outputs better.

Your Claude.MD stores everything about how you like to work. So why is the team that's actually building Claude saying they've removed 80% of Claude code system prompts? Well, my Claude.MD is packed with a whole bunch of dumps. Don't use this phrase, don't touch that folder. And by doing this, I thought I was training Claude, but I was actually giving it a bunch of unnecessary speed bumps. Every hard rule that you give Claude code is a speed bump that it has to slow down a reason about, before it even gets to solving the task. Okay, older models needed all of these extra rules, but the latest models, they don't. So I gave it a try. I ran the slash doctor command on my Claude.MD that I use every single day. And it flagged a bunch of old rules that I wrote years ago that were just taking up space and eating tokens. So I trimmed or deleted those stale rules and the outputs actually got better. And for my preferences that were actually worth keeping, Claude now automatically saves memories that are relevant to your work. Try running slash doctor in your project and every single rule that it flags ask yourself. Is this something that Claude really needs or is it just another speed bump?