What to put in CLAUDE.md

Hey 👋

Welcome back! Last week Anthropic filed to go public on Nasdaq. OpenAI released GPT-6.1 Terra, and Google announced Gemini 4 Carbon, which it calls its strongest model so far.

Our focus in this issue is CLAUDE.md. "What is CLAUDE.md?" is the question we get most in our bootcamps, so here is what it is and what to put in it.

In this issue

• A 3D game Claude Code built while Aleksander was at the beach
• Anthropic's wild numbers
• What CLAUDE.md is, and how to start with one folder
• AI News, with GPT-6.1 Terra and Gemini 4 Carbon

A 3D game built from the beach

Aleksander let Claude Code build a full 3D adventure game in five days, mostly while he was at the beach. Claude Code led a team of agents, Python scripts made every 3D model in Blender, and ElevenLabs made the music, sounds and voices. He made the decisions from his phone.

Even if you are not making games, it is interesting to see how Aleksander collaborates with an army of AI agents on a task as complex as an advanced game.

The video also shows what went wrong, such as a sprint that passed every test and still needed a full night to be good.

Anthropic's wild numbers 🤯

The filing is the first time Anthropic has shown its numbers.

Revenue went from $41 million in 2023 to $4.6 billion in 2025. In the first six months of 2026 it was $16.3 billion, fourteen times the same months a year earlier.

Customers paying $100,000 or more a year grew from about 200 at the end of 2024 to more than 11,900 by the end of September. The full filing is at the SEC.

What is CLAUDE.md?

CLAUDE.md is a plain text file in your folder that Claude reads at the start of every session. Each session starts fresh, without what you told it last time, so the file holds what Claude should know and cannot work out from the files. Think of it as the note you would leave a capable colleague on their first morning.

You write it in plain language, with a few headings for what the folder is for, where the sources are and which rules apply. CLAUDE.md is not only for Claude Code. It works in Claude Cowork too, when you work with an agent in a folder.

What to put in it

• Keep it under 200 lines, as Anthropic recommends. Claude follows a long file less reliably, so put additional instructions in separate markdown files that Claude opens when it needs them, or in skills.
• Add a line when you correct Claude twice, so the third time does not happen.

You do not have to write the file yourself. Type /init in Claude Code for a first draft, and end each working session with this prompt.

Add what we learned today to CLAUDE.md.

This is where it compounds. Every lesson you save is there in the next session, so Claude gets a little better each week instead of starting from zero. Read the lines it adds, because a wrong line repeats itself too.

What is CLAUDE.md? goes deeper, and it is also in Norwegian.

Start with one folder

If you are new to working this way, start with one job you do every week or every month. Choose one where you know what a good result looks like, such as a summary after a client meeting.

Put the sources and one good example in a folder, and open it in Claude Cowork or Claude Code.

I want an agent in this folder that helps me with [the job]. Look at the files, then make three folders called sources, examples and output, and move my files into the right one. Then write a CLAUDE.md for the folder, and ask me about anything you cannot see in the files.

An agent in a folder shows how one folder grows into a team of agents. It is also in Norwegian.

Our bootcamp last Tuesday

On Tuesday 27 October we were at Mesh for our Claude Code bootcamp for everyone. Thank you to everyone who spent the day with us.

We also run company bootcamps built around your team's tools, workflows and priorities. Get in touch if you want one for your team.

Coming up

NDC Porto, 17 to 20 November. Aleksander speaks at Kopke Hotel in Porto.

AI News

Here are the other updates from last week worth a look.

AI stocks fall on doubts about the payoff. The biggest AI and chip stocks have lost about $5 trillion in value since 6 October, by Yahoo Finance's count. Last week Reddit fell 36 per cent and ServiceNow 34 per cent, while the S&P 500 fell only 0.6 per cent.

Experts mostly go along with their AI agent. In an Anthropic study, Claude Code was set to lean one way on four key choices for 50 analysts. They went with it 86 per cent of the time, and 77 per cent repeated its arguments, but only 15 per cent said so. Form your own view before you ask.

Claude's Max plans get new names. Max 5x is now Max and Max 20x is now Ultra, at the same prices. Verified Ultra accounts no longer have the five-hour usage limit, while the weekly limits stay.

ChatGPT and Codex will share one usage limit. From December, Plus, Pro and Business plans get one allowance for both, measured in tokens. Instant chat does not count against it.

OpenAI releases GPT-6.1 Terra. It is a fast model for writing, working with images and using a computer for you. OpenAI says it matches GPT-6 Astra on a test of agents using software at work, at a tenth of the price per token. It is in ChatGPT Work and Codex on Plus, Pro, Enterprise and Edu.

Google announces Gemini 4 Carbon. Google calls it its strongest model so far, built for long coding and security tasks. For now only a group of cyber defenders can use it, and Google gives no date for everyone else.

Gemini makes documents, sheets and slides in the chat. You can create and edit files inside a Gemini chat, then open them in Google Workspace or download them. It works on both personal and work accounts.

Copilot puts Chat and Cowork in one place. The Copilot app now has Chat for questions and Copilot Cowork for tasks with several steps on the same Home screen, rolling out worldwide in November. The same month you can build your own skills in Copilot Chat by describing them in plain language.

If you know someone who keeps repeating the same instructions to Claude, feel free to forward this email.

Best, Aleksander and Simon