Claude Code for Business: Run Your Company From One Folder

If you run a business on AI chat, you know the routine. You paste the same proposal, explain the same clients and describe the same tone of voice, then open a new chat next week and do it all again. The chat window keeps nothing about how your company works.

Claude Code and Codex fix that with a plain folder on your computer. You put your business files in it, point the agent at it, and before every task the agent reads two short text files that describe your business: CLAUDE.md for Claude Code, AGENTS.md for Codex. Writing those two files is the part that stops most people, so this guide includes a free prompt that reads your folder and drafts both for you.

Below is the full setup, from building the folder to keeping the files current, plus what it costs, which files the agents can read and the limits worth knowing.

Why a chat window keeps forgetting your business

A file you upload to a chat is a copy stuck inside that one conversation. When you change the real proposal on your computer, the chat still has the old one. When you open a new chat, the explaining starts over.

An agent that works inside a folder turns this around. The files stay where they are, and the agent opens the current version each time. It can also save new files next to them: a summary, a proposal draft, a client email. And because it reads CLAUDE.md or AGENTS.md first, it already knows who your clients are, how you price and how you write before you type a word.

What Claude Code and Codex are, in plain words

Claude Code is Anthropic's AI agent that works on the files in a folder you choose. It reads them, writes new ones and edits old ones. Codex is OpenAI's version of the same idea. Both were built for software teams, which is why "code" is in the name, but a folder of proposals, client notes and price sheets works the same way as a folder of code. You talk to them in plain English.

The memory comes from two plain text files that sit in the folder:

• CLAUDE.md is for Claude Code. Each Claude Code session starts with a fresh context window, and Anthropic's docs say Claude reads CLAUDE.md at the start of every session.
• AGENTS.md is for Codex. OpenAI's docs say Codex reads AGENTS.md before doing any work.

AGENTS.md is an open format, looked after by the Agentic AI Foundation under the Linux Foundation. Its site says more than 60,000 open-source projects use it and lists tools that read it, including Cursor, GitHub Copilot's coding agent, Gemini CLI and Windsurf. So the file you write for Codex works in those tools too.

Before you start

Step 1: Build your business folder

1. On your desktop, create a new folder and give it a simple name, such as My Business.
2. Add the files you want the agent to know about: proposals you have sent, your price list, client notes, saved email threads, templates, process docs, and voice notes if you have them.
3. Sort them into a few subfolders. The agent finds things faster, and the prompt in step 3 turns this layout into the folder map in your files. A layout that works for most service businesses:

   CopyMy Business/
     Clients/        one subfolder or file per client
     Proposals/      past proposals and the current template
     Pricing/        price list, packages, discounts
     Templates/      emails, contracts, onboarding docs
     Process/        how you deliver, checklists, tools
     Notes/          meeting notes, transcripts, ideas

4. Fix the formats the agent struggles with. Claude Code's built-in reader refuses Excel .xlsx files, so save your key sheets as CSV too. Long PDFs need a free tool called poppler (more on file types below). If you have voice notes or videos, transcripts help, because the agent may not be able to transcribe them.
5. Take out anything you would not paste into a chat: passwords, bank details, card numbers, and regulated data such as health records.
6. Copy the whole folder somewhere safe as your backup.

Step 2: Open the folder in Claude Code or Codex

In the Claude desktop app:

1. Download the app from claude.com/download, install it, sign in, and click the Code tab.
2. Under the prompt box there is a row of controls. Use the environment selector, the first one, to choose Local, so Claude works on the files on your computer.
3. Click the project folder control next to it and pick your My Business folder.
4. Check the permission mode. Manual is the safe start: Claude asks before editing files or running commands, and you see each change and can accept or reject it. Accept edits lets it write files without asking. Plan has it read and propose a plan without editing anything.

If you prefer the terminal, the same thing works there. Install Claude Code from the quickstart, move into the folder with cd, and type claude.

In the ChatGPT desktop app, choose Codex instead of ChatGPT, start a new chat, and open your folder. Sign in with your ChatGPT account if it asks.

Step 3: Paste the prompt that writes both files

With the folder open, paste this whole block and press Enter:

CopyRead every file in this folder and its subfolders before you write anything. That means documents, notes, emails, spreadsheets and PDFs. If there are voice notes or videos and you can transcribe them, do that too. If you can't, list them so I know.

Then create two files in the main folder: CLAUDE.md and AGENTS.md. Claude reads CLAUDE.md at the start of every session. Codex reads AGENTS.md. Put the same content in both.

Use these sections:
1. About the business: what we do, who we sell to, and what we charge, if the files say so.
2. Clients and projects: one line each, plus the file where the details live.
3. How we work: our process, tools, deadlines, and who does what.
4. How we write: our tone and the words we use, with short examples from my own emails and docs.
5. Rules: anything the files say we always do or never do.
6. Folder map: what lives in each folder, so you know where to look.
7. Open questions: anything you couldn't work out. List it here instead of guessing.

Keep it under 150 lines. Plain words, short bullets. Only write what the files actually say.

When you're done, show me both files and ask me what to fix. Whenever we change how we work, update both files.

Prefer a one-click copy? The same prompt is on our free prompt page.

What to expect while it works:

1. The agent opens your files one by one. A bigger folder takes longer, and in Manual mode it may ask before running a command to read a file type it cannot open directly.
2. It asks permission to create CLAUDE.md and AGENTS.md (unless you chose Accept edits), then writes them in the main folder.
3. It shows you both files and asks what to fix. Any voice notes or videos it could not transcribe should be listed, and anything it could not work out should sit under Open questions.

If it stops partway on a large folder, tell it to continue from where it left off. If it cannot open a file type, ask what it needs, or convert that file to text or CSV and run the prompt again.

Why the prompt is written this way

• It tells the agent to "only write what the files actually say" and gives it an Open questions section. That keeps it from guessing your prices or inventing a client.
• The "How we write" section asks for short examples from your own emails and docs, so the agent copies how you really write instead of working from a vague word like "friendly".
• The folder map tells the agent where things live, so it opens the right client file instead of searching the whole folder each time.
• The 150-line cap sits under Anthropic's advice of 200 lines per file. Its docs say longer files take up more context and reduce how closely Claude follows them.
• Sections with short bullets match Anthropic's advice to group instructions under headings and bullets, which Claude follows more easily than dense paragraphs.

Step 4: Review and correct the files

1. Read both files from top to bottom. In the Claude desktop app you can also open the Changes view to see what was added, line by line.
2. Check the facts first: prices, client names, who does what. These are the lines a wrong guess would hurt most.
3. Answer the Open questions. Type the answers in plain words, such as "Our standard retainer is billed monthly, and we never offer refunds after kickoff", and ask it to add them to both files.
4. Make vague rules concrete. Anthropic's guidance is to write instructions you could check, such as "Use 2-space indentation" instead of "Format code properly". For a business that means "Proposals go in Proposals/ as PDF, named Client-Month-Year" instead of "keep files organized".
5. Remove anything that contradicts itself. If two instructions clash, Anthropic says Claude may pick one at random.
6. Check that it loaded. In Claude Code, start a new session and type /context: CLAUDE.md should appear under Memory files. In Codex, ask "Summarize the current instructions."

Step 5: Keep the files current

• When something changes, say so in plain words: "We raised the starter package price. Update CLAUDE.md and AGENTS.md." Asking for CLAUDE.md by name matters, because when you tell Claude Code to "remember" something, it saves it to its own auto memory instead.
• Auto memory is notes Claude writes for itself from your corrections. It loads the first 200 lines or 25KB every session. You can switch it on or off with the /memory command, which also lists your CLAUDE.md files and opens them for editing.
• Every few months, trim old rules as you add new ones. In recent versions of Claude Code (v2.1.283 and later), /doctor prompt-audit checks your instruction files for outdated or conflicting lines and proposes edits without changing anything until you say so.
• For rules about you rather than the business, such as how you like replies formatted, a personal file at ~/.claude/CLAUDE.md applies across every folder you open in Claude Code.

One file or two

The prompt puts the same content in both files, which means two copies to keep in sync. There is a tidier option. Keep one AGENTS.md, and make CLAUDE.md a single line that says @AGENTS.md. Claude Code treats that line as an import and loads the AGENTS.md content.

Recent versions of Claude Code (v2.1.277 and later) also read AGENTS.md directly when there is no CLAUDE.md at all. If both exist, Claude reads only CLAUDE.md. On Windows, use the import line rather than a symlink: Anthropic notes that symlinks there need admin rights or Developer Mode, and Git can turn them into a plain one-line file.

Codex has a size limit worth knowing: it stops adding instruction files once they reach 32 KiB combined, its default. A 150-line file is far below that.

Prompts to try once the files are in place

These work because the agent already knows your business. Change the names and folders to match yours.

CopyWhat did we agree with [client] about scope and price? Quote the file you found it in.

Which proposals in Proposals/ have no reply noted in Clients/? List them with the date sent.

Write a proposal for [new client] using our last three proposals as the template. Use our current prices from Pricing/ and save it in Proposals/ as a draft.

Draft a follow-up email to [client] in our usual tone. Keep it under 120 words and do not send anything.

Read Notes/ from this week and write a one-page summary of decisions and open tasks. Save it in Notes/.

Asking it to quote the file it used is a good habit. You can open that file and check the answer yourself.

What it costs

From the official pricing pages on October 27, 2026. Prices change often, so check both before you pay.

The cheapest way to try the idea is Codex on a free ChatGPT account, which includes it for quick tasks. For a real week of business tasks, Claude Pro or ChatGPT Plus is the usual starting point.

What files the agent can read

• Text, Markdown and CSV files are read directly.
• Images (PNG, JPEG, GIF and WebP) are seen as pictures, so a screenshot of a dashboard or a photo of a whiteboard is fine.
• PDFs are read whole when they are short. Past 10 pages, Claude reads them in chunks of up to 20 pages, which needs poppler installed. Without it, those reads fail. You can ask Claude to install it.
• Excel .xlsx files are refused by Claude Code's built-in reader, so it falls back to a skill or a small script. CSV copies of your key sheets avoid the detour.
• Voice notes and videos are hit and miss, which is why the prompt asks for a list of anything it could not transcribe. Check that list before you assume it heard anything.

Limits and gotchas

• The files are guidance, not locks. Anthropic says Claude treats CLAUDE.md as context, not enforced configuration. A line like "never delete anything in Clients/" is not a guarantee, so keep the backup and keep Manual mode on until you trust it.
• Your data leaves your computer. Claude Code runs locally, but it sends your prompts and its replies to Anthropic over an encrypted connection, and what it reads from your files becomes part of that conversation. On Free, Pro and Max, Anthropic trains on that data only if the model improvement setting is on, which you can change any time in your privacy settings. Team, Enterprise and API use are not used for training under Anthropic's commercial terms.
• Computer control is a preview. In the Claude desktop app, computer use is a research preview on macOS and Windows. It needs a Pro or Max plan, is off by default, and is not available on Team or Enterprise.
• Everything it writes is a draft. Read proposals and client emails before they go anywhere.

When you want a setup like this across a whole team, with shared folders, rules and automations, that is the kind of work our AI automation team does.

Who this is not for

• If you ask AI one quick question a week, a chat window is simpler and fine.
• If your business data is regulated, such as health records or card data, read your provider's commercial terms and talk to whoever owns compliance before you put files in the folder.
• If you are not ready to check the agent's work, start with a copy of your folder, not the original.

Frequently asked questions

Do I need to know how to code to use Claude Code?

No. You type what you want in plain English, and Claude Code reads, writes and edits the files in the folder you picked. The name comes from its software roots, but it works on proposals, notes and documents the same way.

What is the difference between CLAUDE.md and AGENTS.md?

They do the same job for different tools. Claude Code reads CLAUDE.md at the start of every session, and Codex reads AGENTS.md before it does any work. Many other agents, including Cursor and Gemini CLI, also read AGENTS.md. You can keep one AGENTS.md and make CLAUDE.md a single line, @AGENTS.md, so Claude imports it.

Is Claude Code free?

Not on the free Claude plan. It comes with Claude Pro ($20 a month, or $17 a month billed yearly), Max and Team. Codex is included in every ChatGPT plan, including Free and the $8 Go plan, which are meant for light tasks.

Can Claude Code read my PDFs and Excel files?

It reads short PDFs whole and longer ones in chunks of up to 20 pages, which needs a free tool called poppler installed. Its built-in reader refuses .xlsx files, so it falls back to a skill or a script. Saving key sheets as CSV is the easy fix.

Does Claude Code remember my past conversations?

Each session starts fresh. What carries over is your CLAUDE.md file, the files in your folder, and auto memory, which is notes Claude writes for itself from your corrections and loads every session. You can switch auto memory on or off with the /memory command.

Is my business data safe in Claude Code?

The agent runs on your computer, but your prompts and its replies, including what it reads from your files, go to Anthropic over an encrypted connection. On Free, Pro and Max, Anthropic trains on that data only if the model improvement setting is on. Team, Enterprise and API use are not used for training. Keep passwords and bank details out of the folder.

Sources

• Anthropic: How Claude remembers your project (CLAUDE.md, AGENTS.md, auto memory)
• OpenAI: Custom instructions with AGENTS.md
• AGENTS.md, the open format
• Anthropic: Claude Code in the desktop app
• OpenAI: Codex in the ChatGPT desktop app
• Anthropic: Claude Code tools reference (file types)
• Anthropic: Claude Code data usage
• Claude pricing
• ChatGPT and Codex pricing