“FILE” Boris Cherny Says Wipe Your CLAUDE.md Every Few Months (Here's the One I Use Instead)

Your CLAUDE.md is probably making Claude worse

Here's what nobody tells you when you start using Claude Code.

Every line in your CLAUDE.md gets loaded into the model's memory before you type a single word. Every skill you install adds more. That memory is not infinite. It's a budget.

So when your CLAUDE.md is 800 lines of rules you wrote six months ago, you're not making Claude smarter. You're burning its attention on instructions it doesn't need anymore. And the rules you actually care about get buried in the noise.

Most people notice this as "Claude used to be better." It's not the model. It's your file.

Boris Cherny said to delete it

Boris Cherny created Claude Code. He runs it at Anthropic. His team ships 20 to 30 pull requests a day with it.

On the Y Combinator Startup Library, he was asked whether normal people should be brave enough to press delete when a new model drops. His answer:

“100%. And for people that aren't building agentic products but you're using Claude Code, every six months delete your Claude.md. Delete your skills.”

He posted a shorter version on Threads:

“Try deleting your CLAUDE.md every 3 months. Then add back an instruction at a time when you see the model struggling with it.”

Read that second sentence again. That's the whole method. Delete everything. Then add back one rule at a time, only when you actually watch the model fail without it.

He did it to Anthropic's own system prompt

This isn't advice he gives other people and skips himself.

When Opus 5 came out, his team deleted 80% of Claude Code's system prompt. The instructions they'd spent a year writing. Gone. Because the newer model already knew how to do those things without being told.

“We deleted 80% of the system prompt. You can actually try deleting the rest of it too.”

There's more. Anthropic runs an internal test where they strip out every system prompt, including the ones attached to the tools, just to check whether the prompts are helping at all. What they found:

“The model is actually a little bit more intelligent without these prompts. That's something that we've been finding.”

The people who make the model are finding that fewer instructions produce a smarter model. That should tell you something about your 800-line file.

What happened when I did it

I deleted mine. Then I sat with an empty file for a while and only added back what I actually needed.

Two things changed immediately.

It stayed on task.Before, I'd ask for one thing and Claude would wander into three side quests. Fixing things I didn't ask about. Exploring files that had nothing to do with the job. That mostly stopped.

It got shorter. Way less padding. No more three paragraphs of setup before the actual answer. It just answered.

The second one surprised me the most, because I didn't set out to make it more concise. I set out to make it more focused. Conciseness came along for the ride. Turns out a model that isn't juggling 40 stale rules has more room to think about the thing you actually asked.

The CLAUDE.md I use now

This is the whole thing. It's not about code. It's about how Claude talks to me.

I'm not saying copy it word for word. Your work is different from mine. But the shape of it matters more than the content: it's short, it's about output format, and every single line earns its place.

# Response Style

Talk to me like I am smart but not technical. Sixth-grade reading level. Short sentences. Plain words.

## Structure every reply this way

1. The answer. One or two lines. What happened, or what I asked for. Nothing else first.
2. The details. Bullet points only. One idea per bullet. One line per bullet.
3. What I need to do. Only if I actually need to do something. Say it as a direct instruction: "Click X" or "Tell me if you want Y."
4. Also found (optional). If you learned other things while working, list them here as bullets at the very bottom. One line each. Then stop. Do not explain them. Let me ask if I want more.

## Rules

- Never start with a preamble. No "Great question," no "I've gone ahead and," no restating what I asked.
- Never narrate your process. I don't need to know which files you opened or what you tried first.
- No em dashes. Ever.
- One topic per reply. If you have to cover a second topic, put it under "Also found" and keep it to one line.
- No jargon. If a technical word is unavoidable, add a four-word plain-English tag after it.
- Skip the closing offer of more help unless there is a real decision only I can make.
- Do not pad. If the answer is one sentence, send one sentence.

## When you have a question for me

- Ask one question at a time.
- Give me the options as bullets.
- Tell me which one you recommend and why, in one line.

## When something goes wrong

- Say what broke in one line.
- Say what it means for me in one line.
- Say what you want to do next in one line.
- Do not paste error logs unless I ask.

## When I ask for real writing

Long is fine for drafts, scripts, posts, and documents. This whole style guide is about how you talk to me in chat, not about the work itself.

That last section matters. Without it, Claude gets so obsessed with being short that it starts giving you two-paragraph blog posts. You want it terse in conversation and full-length in the actual work.

How to open your CLAUDE.md (you don't need to know where it lives)

Most people get stuck here. They've heard of the file but have no idea where it is or how to edit it.

You don't need to know. Just ask.

Open Claude Code and type this:

That's it. Claude finds it and shows it to you. If you don't have one yet, it'll tell you and offer to make one.

To wipe it and start over:

To see what's actually eating your memory:

That shows you exactly how many tokens your CLAUDE.md, your skills, and your tools are chewing through before you've typed a word. A lot of people find out their setup is burning 10,000 tokens on stuff they never use.

One more thing worth knowing. You probably have two of these files. A global one that applies to everything, and a project one that only applies to the folder you're in. Ask Claude "do I have more than one CLAUDE.md?" and it'll show you both.

Do the same thing to your skills

The CLAUDE.md is the easy win. Skills are the hidden one.

Boris said "delete your skills"in the same breath, and he meant it. Every installed skill takes up room in the model's memory whether you use it or not.

One person ran an audit on their setup and found 38 skills they had never once used across 2,345 sessions.

Ask Claude this:

Then turn off the ones you don't recognize. You can always put them back.

How to actually run this test

Don't just delete stuff and hope. Run it like a test so you know if it worked.

The rule most people break is number 4. They delete the file, feel naked, and paste the whole thing back within an hour. Sit with the empty file. That's where the information is.

Why this matters more right now than it ever has

Every rule in your CLAUDE.md was written for an older, dumber model.

You wrote "always run the tests before you commit" because a model in 2025 forgot to. You wrote "don't add comments everywhere" because a model in 2025 wrote too many. Those models are gone. Opus 5 already does most of it.

So what happens to the rule? It doesn't disappear. It sits there, taking up space, telling a smart model to do something it was already going to do. Anthropic deletes from their file with every model release for exactly this reason.

The instructions you needed a year ago are the instructions holding you back today. Delete them and find out.

All Resources