Set up your CLAUDE.md file

Open a folder. Paste a template. Replace six placeholders. Test that it works. Done.

The thing that turns Claude from a generic AI into one that knows your business is a forty-line text file in the folder where your work lives. You don't have to be a developer to write one. You don't have to know what a markdown file is before you start. The only thing you need is thirty minutes and a clear picture of what your business actually does.

This is the how-to companion to what is a CLAUDE.md file. If you haven't read that one, skim it first; the rest of this assumes you know what a CLAUDE.md is and why one matters.

What you'll have when you're done

A working CLAUDE.md sitting at the root of your project folder, roughly under one screen on a laptop, covering six things: what your business does, how the folder is organized, the voice you write in, the commands and tools you use, who's on your team, and the things you never want done without confirmation. The next time you type claude in that folder, Claude opens with full context. The thirty-minute setup is permanent; you'll feel the leverage from the first prompt of the next session.

A folder Claude doesn't know yet is wasted Claude

For a year of running Claude inside Headphones.com, the first thing I did in every chat tab was paste the same context paragraph. Nine-figure ecommerce, headphones and audio, X thousand SKUs, our voice sounds like this, here's where the analytics queries live. Then I'd ask my actual question. Then I'd close the tab and the context died with it.

The first time I dropped that paragraph into a CLAUDE.md at the root of the Headphones folder and watched Claude pick it up on the next session without me typing a word, I felt like I'd been doing it wrong the whole time. The single highest-leverage thirty minutes you'll spend on Claude this month is writing the file you're about to write.

What you need first

• Claude Code installed. If you haven't yet, start at claude.ai/code. One-time install.
• A folder where your work lives. A project folder, a ~/notes directory, your home folder. Pick the one you'd be in if you were using Claude.
• A text editor you're comfortable with. TextEdit, VS Code, Cursor, whatever you already use. CLAUDE.md is just a text file.
• Thirty minutes of clear time. Most of it goes to thinking about your business, not to anything technical.

Step-by-step

Step 1Pick the folder

The folder is whichever directory on your laptop contains the work you'd want Claude to know about. For me that's ~/headphones-com/ for the ecommerce business and ~/desk-theory/ for Desk Theory. If you don't have a folder yet, make one. Open the terminal and type:

mkdir ~/my-business && cd ~/my-business

That creates the folder and moves you into it. From here, every Claude Code session you start in this directory will pick up the CLAUDE.md you're about to write.

Step 2Create the file

From the same terminal:

touch CLAUDE.md

That creates an empty file named CLAUDE.md. (If you'd rather use Finder: new file, name it CLAUDE.md. The capitals matter on case-sensitive systems.) Open the file in your text editor of choice.

Step 3Paste the template

Paste this into the empty file. Don't customize anything yet. Save. The point of this step is to see the shape of a working CLAUDE.md before you touch the content.

# CLAUDE.md · {Your Business Name}

## What we do
{One paragraph: revenue scale, customers, market, what you actually sell.
Specific numbers beat ranges.}

## How this folder is organized
{Where files live. Naming conventions. Canonical sources of truth.
Folders to avoid touching.}

## Voice and writing rules
{The tone you write in. What to imitate. What to never do. If you have
a voice profile document or a style guide, point to it here.}

## Commands and tools you use
{Scripts you run weekly. CLI tools like `gws`, `vercel`, whatever.
External services this folder talks to.}

## Team and ownership
{Names. Roles. Who owns what. Who not to email about what.}

## What NOT to do
{Hard boundaries. Sensitive folders. Destructive commands. Compliance
limits. Anything you'd flag to a smart contractor on day one.}

Save. Six headings, six empty sections, ready for content.

Step 4Replace the placeholders

This is the part that actually takes time, and it's the part where most CEOs get stuck. You're not writing code. You're writing the brief you'd hand to a smart contractor on day one. The trick to getting unstuck: open Claude Code in the folder and ask Claude to do the first pass for you.

claude

Then type:

Read CLAUDE.md. For each section that's still a placeholder, draft a first
version based on what you can see in this folder. Use what you find in the
files. Where you don't have enough information, leave a [QUESTION] for me
to fill in. Don't make anything up.

Claude will draft the sections it can ground in the folder (file conventions, command tools, whatever's documented in your code or notes). It'll leave [QUESTION] markers in the sections that need you (business model, voice, team, hard boundaries). Read the draft. Edit anything that's wrong. Fill in the [QUESTION] markers.

The first pass takes longer than the rest. Once the What we do and Voice and writing rules sections are solid, you can lean on Claude to ask follow-up questions ("what's missing from the team section?") and tighten things up.

A few notes on the harder sections:

• What we do. Specific numbers beat ranges. "$12M ARR, 8,000 paying customers, North America DTC" beats "growing ecommerce business." If a smart contractor would Google your name to figure out what you sell, this section is too vague.
• Voice and writing rules. Either describe the voice in a paragraph, or point to a voice profile document (See ./voice-profile.md). If you don't have a voice document yet, the description-paragraph path is fine; you can graduate to a document later.
• What NOT to do. This section earns its keep on the day Claude almost does something you'd regret. Sensitive folders, destructive commands, compliance limits, things you handle personally. Write it like a contractor warning, not a config.

Step 5Trim aggressively

CLAUDE.md gets read on every session. Every line costs context. Once you have a complete draft, go back through and cut anything that isn't load-bearing. Aim for under one screen on a laptop, roughly 60-80 lines. If a section runs long, point to a separate document and link from CLAUDE.md instead of inlining everything.

Step 6Test it

Open Claude Code in the folder fresh. Ask one question that exercises the CLAUDE.md. Mine is "What's our return policy?" in the Headphones.com folder. The version of Claude that doesn't know my business says "I'd need more context." The version with a working CLAUDE.md answers the question, references the policy document by name, and writes in the brand voice without me asking for it.

Pick whatever question is most diagnostic for your business. Operational ones work best because the answer is either right or it's wrong; there's no middle ground.

How you'll know it's working

The test query in Step 6 is the immediate signal. The longer signal lands a few days later. By your third Claude session that week, you'll notice you stopped pasting context paragraphs at the start of conversations. By the fourth, you'll start asking questions you didn't think to ask before because they would have taken too much setup to be worth it. That's the unlock. It isn't the test query passing; it's that you start asking questions you didn't think to ask before.

When it breaks

• Claude doesn't seem to know the CLAUDE.md exists. Check the file sits at the root of the folder you opened Claude Code in, named exactly CLAUDE.md (capitals matter on case-sensitive systems). Run pwd to verify your working directory; run ls CLAUDE.md to confirm it's there.
• Claude knows the file but the answers feel stale. You edited the file after Claude already loaded it. Quit the session (Ctrl+D or type /exit) and start a fresh one.
• The CLAUDE.md is too long and Claude starts missing things. You stuffed it. Trim. Aim for under 80 lines, and move long-form content (voice profiles, runbooks) into separate files that CLAUDE.md links to.
• Claude is confidently making things up about your business. A specific section is too vague. Find the section that produced the wrong answer and add concrete numbers, names, or examples.
• You're embarrassed by your CLAUDE.md and don't want to commit it to git. Either trim what's embarrassing (usually means too much detail) or move it to the user-level file at ~/.claude/CLAUDE.md, which lives in your home directory and doesn't get committed with the project.

Where this fits in your harness

CLAUDE.md is one of two foundational primitives in Claude Code. The other is skills · packaged, named capabilities Claude can invoke on demand. CLAUDE.md gives Claude context; skills give Claude moves. Pair them and you have an agent that knows your business and knows how to do your work.

Once the file is live and you're a few sessions in, the next layer is persistent memory across sessions so notes, decisions, and past work stay available beyond a single conversation. From there, routines put recurring work on autopilot.

Write the file. Run the query. When it works, send me the question you asked and the answer you got. I love seeing what people put in their first CLAUDE.md.