Anthropic quietly deleted 80% of Claude Code's own built-in instructions. The result got better, not worse. Boris Cherny, the guy who created Claude Code, said the reason plainly: every instruction in your setup was written to patch a problem an older model had. The model moved. Most people's CLAUDE.md never did. So right now, a huge chunk of what's sitting in your Claude config is dead weight, instructions fighting problems that don't exist anymore, actively making the current model perform worse. The fix isn't "delete everything and hope." It's a prompt that makes Claude audit itself: → Reads Anthropic's live prompting docs for your exact model, not what it remembers from training → Goes through your CLAUDE.md, skills, and hooks line by line → Gives one verdict per line: delete, keep, or rewrite, with the reasoning in plain English → Quotes the actual source that justifies each delete, no invented justifications → Flags the sneaky stuff: "always double-check your work" rules that make the model redo work you're already paying for, "don't overthink" rules that leak internal reasoning into the output, role-padding that used to matter and now doesn't → Tells you what's missing, not just what's excess Then it hands you a table: your line, the verdict, the reason, and the exact wording to paste in for whatever's missing. You don't write the audit. You just read it and approve it. Free skill, one prompt, six months of accumulated instruction-cruft cleared in one pass. When's the last time you actually looked at what's sitting in your CLAUDE.md?