What Is a CLAUDE.md File? The One File That Stops You Repeating Yourself to AI

CLAUDE.md is a plain text file that Anthropic's coding tool, Claude Code, reads automatically every time it starts working in a folder. Instead of re-explaining your business, your voice, or your rules in every single chat, you write it once and the AI already knows it before you say a word. That one idea, a file the AI checks on its own instead of a blank box you fill in by hand each time, is the fix behind almost every "AI forgot everything I told it" complaint.

You don't need to write code to get the idea, and you don't need Claude Code to use it. But seeing the literal, mechanical version makes the "why" click in a way that "Custom Instructions" buried in a settings menu never does.

Why your AI keeps forgetting you

An AI chat has no storage of its own. Every time you send a message, the tool gathers everything in that conversation so far and hands the whole bundle to the model fresh, as if for the first time. The model isn't consulting a file with your name on it. It doesn't "remember" last week's chat, because there's nothing to remember from. There's no record outside the conversation you're looking at right now.

Open a new chat, and that bundle is empty. Whatever you haven't typed into the box, as far as the AI is concerned, doesn't exist. So it answers with everything it has, which so far is nothing about you. That's where generic comes from.

Every serious AI tool has a workaround, because every single user runs into this. They all do the same basic thing: give you a place to put standing information so it gets bundled in automatically, every time, without you retyping it. ChatGPT calls it Custom Instructions. Claude.ai calls it Project Knowledge. Gemini calls it a Gem. Different names, same idea: a permanent note that rides along with every conversation so you stop re-explaining yourself.

The clearest version of this: CLAUDE.md

Claude Code makes the idea completely literal, which is useful because it shows the plumbing everyone else's version hides.

Claude Code reads a plain text file named CLAUDE.md, automatically, the moment it starts working in a folder. Nobody pastes anything in. There's no menu to dig through. The file just sits there, and Claude reads it before it reads you.

Two details worth knowing:

• There's a CLAUDE.md for the specific project in front of it, holding instructions and facts that only matter here. Think of this as the project's Business Brain: this business, this voice, these customers, these rules.
• There's a separate CLAUDE.md that lives outside any one project, attached to the person instead of the work, so it applies no matter what Claude is doing that day. Think of this as a Personal Brain: how you write, what you care about, how you like to work, carried into everything.

Same mechanism, two scopes: the work, and the person doing the work. That split is why we built two separate products instead of cramming both into one.

Splitting one file into several

A single CLAUDE.md can get long fast: voice notes, customer detail, pricing, examples, edge cases. Cramming all of it into one document makes it slower to update (change one price, scroll past everything else to find it) and harder for the AI to weigh properly.

So keep the main file short, and have it point to other files instead of holding everything itself. Claude Code supports this directly. Write @voice.md or @customers.md inside CLAUDE.md, and it reads that file too, every time, with no extra step from you. The main file becomes a table of contents, and the detail lives in files organized by topic.

That's a hub and spoke. One hub (short, orienting) and several spokes, one per topic, each easy to update on its own without touching the rest.

If you've ever built a Business Brain, you've already written the spokes. Look at the section headers: the business, your voice, your customers, your offers, your way of doing things, your guardrails, your examples. Same structure, because both come from the same logic: small, stable, well-labeled pieces beat one giant document every time.

Worked out for Riverbend Gallery (our running example in the Business Brain kit), the hub file says, in effect: "this is Riverbend Gallery. See the voice file for how we sound, the customers file for who buys from us, the offers file for what we sell." Each spoke is roughly a page. When prices change, you edit the offers spoke and leave the rest alone.

What Anthropic said to delete, and what to keep

In July 2026 Anthropic published new guidance on these files, and a lot of people read it as bad news. They cut more than 80% of Claude Code's own standing instructions and measured no drop: "We removed over 80% of Claude Code's system prompt for models like Claude Opus 5 and Claude Fable 5 with no measurable loss on our coding evaluations."

Headlines turned that into "instruction files are over." Then look at the sentence where they say what to keep: "Keep your CLAUDE.md lightweight and briefly describe what your repo is for, but spend most of the tokens on gotchas inside of the codebase."

Two different things live in one of these files, and only one of them expired.

Rules about how the AI should behave. "Always use bullet points." "Never write more than three sentences." "Ask before you assume." Most of those were written to stop an older, weaker model from doing something dumb. A 2026 model handles it on its own, and the rules just take up space.

Facts nobody could guess. What you sell and for how much. Who actually buys from you, and why. How you talk. What you refuse to do. No model release generates these, because they are not general knowledge. They are your particulars. A better model with none of them writes a more polished generic email.

The advice for your business is the same one Anthropic gave developers: cut the instruction manual, spend the space on the specifics. Your gotchas are things like "our customers are first-time buyers, so reassurance matters more than price," or "we never quote a delivery date we have not confirmed."

Boris Cherny, who created Claude Code, takes it further. He suggests deleting the whole file every six months or so and adding it back line by line, to see what each line is actually doing. Worth trying on yours. The lines you cannot bring yourself to leave out are the ones carrying the weight, and they are almost always facts about you.

What this looks like for your business

A Business Brain's voice section matters everywhere, but it earns its keep most in the moments where the default AI voice (stiff, hedge-everything, vaguely corporate) is the opposite of what builds trust with a customer who's on the fence.

Say the one line that matters most for your voice is "plain English, no jargon." Hand that to AI as standing context, and a routine job like nudging a customer who started something with you and never finished stops reading like a form letter:

"Quick nudge: you're about 20 minutes from done, and finishing it means you're all set with nothing left hanging. No pressure, no jargon. Reply here and I'll walk you through the last step whenever you're ready."

Nobody re-explained your tone before that email got drafted. It was already in the file. (The video below walks through this exact example.)

Where you'll actually use this (no code required)

CLAUDE.md itself belongs to a coding tool, and most people reading this will never open it directly. That's fine. The mechanism is what carries over to any tool. Here's the same move in the ones you already use:

• ChatGPT: paste your standing context into Settings → Personalization → Custom Instructions. For something longer, build a Custom GPT and add it as the GPT's knowledge.
• Claude: create a Project, and add your context to Project Knowledge. Everything in that project uses it from then on. Or just paste it at the top of any chat.
• Gemini: create a Gem with the context loaded in.
• Any tool, any time: paste the relevant section at the top of a fresh chat: "Here's my business, keep it in mind." Slower than the above, but it always works.

None of these literally read a file called CLAUDE.md off your computer. They all do the job CLAUDE.md does: hold your standing context somewhere the AI checks automatically, so you say it once instead of every time.

Try it this week

1. If you've already gathered your context (a Business Brain, a voice doc, anything like it), stop pasting it in by hand. Put the short version into Custom Instructions, Project Knowledge, or a Gem today, once, and leave it there.
2. If one section keeps changing (pricing, a seasonal offer), give it its own file or its own clearly-labeled block, so updating it doesn't mean rewriting everything else.
3. Revisit it quarterly. Add the new example you're proud of. Fix the line that's gone stale. The file is only as useful as the last time you touched it.

The shortcut

All of this assumes the context is already written down somewhere, and that's usually the real blocker. The technical step is easy. Most people don't have a clean writeup of their voice, their customers, and their offers sitting around, and gathering it from scratch is the part that eats the afternoon.

The Business Brain does exactly that ($50). It walks you through gathering what you already have, your site, your sent emails, your past proposals, into the hub-and-spoke shape this guide describes, ready to drop into whichever tool you use. There's a Personal Brain version too, scoped to you instead of the business, for the same reason Claude Code keeps two separate files.

Build it once. Give it a home. Stop re-explaining yourself.