So Anthropic just

So Anthropic just wiped out 80% of the instructions. They give to Claude Code, and their coding scores didn't move. Same test, same model, 80% less prompts and less tokens. So while your feed fills up with Opus versus Fable, and which one's dead, the people who are actually building Claude Code, publish the part that matters. If you're new here, I'm Kevin. I break down AI and how to actually build with it. So if that's your thing, hit follow. So Thark who works at Claude Code, put out what they learned rewriting their prompts for Opus 5, a list of things we all got told to do that now makes the model worse. Start with the rules. Their old prompt, hard-coded lines like default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks, one short line max. Don't create planning, decision, or analysis documents unless. That whole block and what it says is one single sentence: write code that reads like the surrounding code. And it covers rulebook got wrong. Same with examples. For the past three years said, give the model examples. Turns out examples box it into a narrow lane, you stop getting of these smarter answers. And it's already breaking real setups. Dan Shipper and folks that Every built a system called compound engineering, skills and rules stacked on top of Claude. Opus 5 shipped and it broke. You may have heard me talk about it, and it's a set of skills that gets brought up a lot in my AI Discord community I run with other creators. And so when Opus 5 shipped, it completely broke compound engineering. The agents, in this case, Opus 5, kept handing control back to the users of compound engineering mid-task, and they've had to go back to the building blocks to figure out how to build this for Opus. So they've totally wiped it off. And here's where you. Your Claude.md file, you know, that file that guides Claude every instruction prompt. That's the instruction file your AI reads before every single task. And we all treat it like a. Every rule we pick, we overload things, we keep just throwing stuff in that cabinet because we're scared that AI won't find those things otherwise. Anthropic calls that a myth now. The move is a tree of smart files that pulls only when it needs to, and they did it in their own prompt. Code review and verification moved out of it into separate skills Claude calls work shows up, when they actually need to. And so I counted mine, it's over 2,200 words. So I'm the guy on here telling you to keep it tight. And so I imagine yours is probably big too. And honestly, probably half of it is me over-explaining things that Opus 5 already knows. To do. Open Claude Code and run /doctor. They shipped this around the time they shipped Opus 5, and it tells you exactly what to cut in your stack. Then keep the gotchas, the stuff that's generally weird about your project that's sitting in your head, then move everything else either into its own file like an archive or into a reference. Your job now: deciding what it sees, and when. I also built a skill that does this and even more, and it's available on my GitHub. So hop over there if you want that, and that'll run /doctor plus some additional things to help you kind of clean up your setup.