Every rule in a CLAUDE.md exists because Claude got something wrong at some point, which means the file only ever grows. Something breaks, you add a rule, and you never look at that line again. After a few months you’re carrying dozens of instructions written for models that don’t exist anymore, and with releases landing as fast as they do now, Fable 5 being the latest, some of those rules have quietly flipped from helping the model to slowing it down. A new model announcement is the natural moment to open the file back up and ask which rules still deserve to be there.

WHY RULES GO STALE WHEN MODELS UPGRADE

Most CLAUDE.md rules are workarounds. The model kept hardcoding color values, so you added a rule pointing it at your design tokens. It kept forgetting where your posts live, so you spelled out the folder structure. Each one patched a real weakness, but the weakness belonged to the model you were running at the time, and every upgrade quietly fixes a few of them. The rule stays though, and because the whole file gets read into every single session, every stale line is tokens spent on a problem you no longer have. Anthropic’s own guidance on context engineering calls out the over-specified CLAUDE.md as a real problem too, because a file stuffed with instructions makes the model worse at following the ones that still matter.

A stale rule can also do more than sit there. There’s research showing that the smarter models get, the more likely they are to push back on a rule that doesn’t fit the task in front of them, sometimes working around it instead of following it. A rule that conflicts with what the task actually needs widens the gap between what’s expected and what’s defined, which is exactly the kind of pressure that pushes Claude into its anxious, corner-cutting mode. The week Fable 5 landed, people on the Claude subreddit were finding that instructions written for the last model steered the new one wrong.

WHICH RULES TO KEEP AND WHICH TO DELETE

The rules worth keeping are the ones about how you want things done: “don’t commit without asking,” “never touch tokens.css,” “always work on the develop branch.” Those aren’t there because the model was weak, they’re there because it’s your project, and they stay true no matter how good the model gets. The rules worth deleting are the coaching ones, the ones that teach the model how to do its job, like step-by-step instructions for tasks it now handles fine or warnings about mistakes it stopped making a generation ago. A decent test is to read the rule and ask whether the model you’re on today would honestly get it wrong without it, and if the answer is probably not, it goes on the list. That’s close to the exact question Anthropic’s best practices guide recommends asking of every line (“would removing this cause Claude to make mistakes?”), and its advice when the answer is no is to just cut it.

HOW TO AUDIT ONE CLAUDE.MD, OR ALL OF THEM AT ONCE

Audit your project’s CLAUDE.md against the keep-or-delete test above with something like this:

Read my CLAUDE.md and sort every rule into three buckets: rules about how I want things done (keep), coaching for mistakes the current model wouldn't make (safe to delete), and rules you're not sure about. For each delete candidate, explain why it's no longer needed. Check with me before deleting anything.

There’s an obvious conflict in asking Claude which of its own guardrails should go, and it will sometimes argue against a rule you added for a reason it can’t see, so read the list yourself before agreeing to anything.

If you’ve built up good CLAUDE.mds across multiple projects over time, they’re all going stale on the same schedule, and a dynamic workflow can audit every one of them in a single run, one agent per project doing the same three-bucket sort and reporting back one combined list. Keep that pass read-only and review the findings yourself before anything gets deleted.

So keep an eye out, and whenever a major new model gets announced, run the audit. Or even better, ask Claude to turn the prompt into a skill so it’s a one-word command you can rerun as regular context hygiene, since new models are going to keep showing up every few months.