Claude Code creator, Boris Cherny: "Every six months, delete your claude.md, delete your skills, delete your hooks, see what the model does, and it might surprise you." Anthropic deleted 80% of Claude Code's system prompt for Opus 5. It doesn't need the instructions past models did. 36 minutes of... pure insight from the creator of the fastest-growing AI coding tool in the world. Watch it, then read the guide below.show more

The delete test is really a sorting test. Half your file is compensating for the model and half is encoding your context, and only the first half rots on a release. The model will never independently learn your naming conventions or which of your three services owns billing. That part is permanent. Their 80% was almost certainly all compensation, and knowing which half of your own file is which is the actual skill.

Subscribe to my free newsletter and get one AI superpower every week:

Less prompting. More understanding of the model.

半年清一次包袱，模型反而更能发挥

Yes and this is because of reinforcement learning. Your detailed prompts became its training data.

Deleting claude.md, skills, and hooks every six months is a clean way to find instructions that became superstition. The audit matters more than the 80% number: keep only what still changes output.

Wild how much less hand-holding these models need now. Makes you wonder what other prompting conventions are just legacy cruft at this point.

Yes, i think it's right way. The real bottleneck was never missing context but constraint overload suffocating reasoning capacity

Claude is the next ftx

I have a CLAUDE.md that's grown to 800 lines. Skills, hooks, cron context, the whole thing. The idea of deleting it and watching what breaks is genuinely unsettling, which probably means I should try it.

Half my claude.md is rules for problems the model stopped having months ago. Still too scared to delete it

opus 5's system prompt cut by 80% means less manual tweaking this could shift focus from prompt engineering to model capabilities

That's a wild approach to development. It's impressive to see how much has evolved so quickly.

I definitely agree with this. When I built up my rules and harness for Claude I used Opus 4.5, but I recently cleared it and started from scratch with Fable. My projects are much more organized now and easier to manage.

Claude not needing instructions... 🤣. I am at the point where I have codex full time managing Claude code... Claude doesn't read instructions, skills, doesn't use the tools, doesn't research, lies, hallucinates...

This sounds really risky at first, but I think they're onto something.

Deletion is not the same as compression.

funny how removing context can sometimes improve results.

Less prompt engineering, more trusting the model is where things are headed

Less prompt engineering more capability is a huge win for everyone

fair, it needs way less than it used to but half of my claude.md is project facts the model cant guess, id lose those

opus 5 hit 30 percent on arc agi 3 which is wild for how far models have come

The 80% they cut was model steering. Idempotency keys, authz boundaries, terminal payment states live in code, not prompts. You can delete prompt scaffolding when the model catches up. The rails that stop side effects stay in the system layer.

The 20% worth keeping maps to real failure modes, not best practices. The rules that survive deletion in our system were all written after a production bug: stuck payment, double refund, vendor timeout that cascaded. The rest was handholding for older models.

Worth splitting two kinds of instructions here. Ones that patch the model's weaknesses - those expire, delete freely. Ones that encode your project's facts (why this module is weird, what broke last time) - no model update makes those obsolete. Most CLAUDE.md files mix both and get deleted wholesale.

This matches what I’m seeing too. As models get stronger, a lot of the old scaffolding becomes dead weight. Periodic cleanup forces you to notice what the model can already handle on its own.