You’ve spent hours perfecting your Claude.md. Every rule, every instruction, every tiny tweak to make your AI assistant behave exactly the way you want. It feels like progress. It feels like mastery.
But here’s the uncomfortable truth nobody tells you: Your perfectly customized AI is now a cage.
Boris Cherny, a veteran engineer who’s seen AI evolve from clunky chatbots to something almost human, dropped a bomb on a recent podcast: delete your Claude.md every six months. Not just update it. Delete it. Start from scratch.
At first, that sounds insane. Why would you throw away months of hard-won customizations? Because the model underneath you is no longer the same model you tuned for. While you were busy optimizing for Claude 3.5, Claude 4 launched. And it has capabilities you’ve never seen—because your customizations are actively hiding them.
Your customizations are technical debt. And debt compounds.
Think about it: every rule you wrote was a reaction to the old model’s weaknesses. You told it to follow a specific format because it used to ramble. You gave it strict tone guidelines because it used to be too formal. You locked in behaviors that solved problems that no longer exist.
Meanwhile, the new model handles nuance better. It’s more creative. It can surprise you—if you let it. But your Claude.md is like a parent who still treats their 25-year-old child like a teenager.
I’ve seen this firsthand. A friend of mine refused to update his prompt library for six months. He was proud of his finely-tuned system. Then he accidentally used a fresh Claude instance and got a response that blew his mind. The new model had solved a problem he’d been wrestling with in three seconds. He’d been forcing the old model’s crutches onto a model that could run.
The most dangerous sentence in AI use is: ‘It worked before.’
This isn’t just about Claude.md. It’s about every customization you layer on top of a rapidly evolving foundation. The moment you stop questioning your setup, you start falling behind.
So here’s the challenge: delete your Claude.md today. Use the bare model for a week. Notice what it does differently. Notice where it excels. Notice where it still fails. Then rebuild your customizations—but only for the problems that actually exist right now, not the ghosts of models past.
Customization is a tool. But when the tool becomes a crutch, you’re not leveraging AI. You’re protecting yourself from change.
And change is the only constant in this space.
FAQ
Q: If I delete my Claude.md, won't I lose all my hard work and have to start over?
A: Yes, you lose the old customizations, but that's the point. The old model is gone. What worked for Claude 3.5 likely suppresses Claude 4's improvements. You're not starting over—you're starting fresh with a better baseline. The time you spent on old tweaks is sunk cost, not future value.
Q: How often should I actually delete my customizations? Is six months a hard rule?
A: Six months is a heuristic, not a law. The real trigger is when a new model version is released that claims significant improvements. If you notice the model behaving differently from your expectations, your customizations are probably outdated. The smarter approach: maintain a 'clean slate' test instance to compare against your customized one.
Q: What if I rely on my customizations for specific workflows or compliance? Deleting everything seems reckless.
A: For production-critical workflows, don't delete blindly. Instead, version-control your Claude.md and run A/B tests. But the principle still holds: periodically audit each rule. Ask: 'Does this rule still add value, or is it compensating for a weakness that no longer exists?' Most rules will fail that test.