Guide · Way of working

Build a second brain with AI: a knowledge system that belongs to you

A second brain is your own knowledge system outside your head, and with AI it becomes a workplace. Tiago Forte popularized the term with his book “Building a Second Brain”. The most common mistake today: many build their second brain in an AI provider's account. There it belongs to you only as long as you stay a customer.

The key points in brief
  • A second brain with AI is a folder with text files on your computer that an AI program reads and edits.
  • The program and the AI model are interchangeable, the folder stays. That way you own your knowledge and only rent the tool. According to agents.md, more than 60,000 open-source projects use the instruction file for this, AGENTS.md.
  • You need 3 files to start: one about you, an instruction file called AGENTS.md and a status file for your first project.
  • According to the manufacturers' documentation, Codex, OpenCode and Hermes Agent read the instruction file AGENTS.md, and according to agents.md, so does Cursor. Claude Code reads it as long as there is no CLAUDE.md of its own in the folder.
  • Getting started takes about 1 hour, plus installing the AI program. Obsidian for reading the files is free, and the AI program uses your account with the model provider. According to OpenAI, Codex is also included in the free ChatGPT plan.
  • What the AI program reads goes to the model provider for processing. Confidential customer data therefore needs its own rules.

A second brain with AI decides whether you explain your company to an AI anew every morning or whether it continues where you left off. I run my whole company from such a folder: projects, customer knowledge, rules and templates. In this guide you build the smallest usable version of it, with 3 files in about 1 hour, plus installing the AI program. If you first want to know what an AI program actually does with files, read AI agents explained simply first.

A second brain is a folder, not a program

A second brain with AI consists of a folder with text files on your computer and an AI program that reads and edits this folder. The program is the tool, the folder is the knowledge. If you swap the tool, the knowledge stays where it is.

Behind it is a principle I took from Nick Milo: own your ideas and only rent the AI tool. AI tools change quickly. Prices, models, features and terms and conditions will look different in half a year than they do today. Your thoughts, decisions and processes should therefore not be tied to any single tool.

Text files in Markdown format meet that requirement. Markdown is plain text with a few characters for headings and lists. Every computer opens such files, even in 20 years, and every AI program can read them.

The folder in the middle belongs to you. The programs around it are interchangeable.

Knowledge in a provider's account can only be taken with you in part

Knowledge you store in an AI provider's projects, memories and custom instructions lives in that provider's account and only partly comes back out. According to Anthropic, the data export from Claude contains the conversations and the account data, but not the original files you uploaded. The help page does not describe an export of project knowledge (as of November 2026).

For a long time you don't notice this in everyday use. You create a project, upload documents, write instructions, and the AI knows you better from week to week. The price shows up when you switch: another program knows nothing of all this. You start there with an empty chat.

With your own folder, this turns around. What the AI needs to know about your work, it writes into your files. The next program reads the same files.

You build links in the provider's account, on the right in your own folder.

Build a second brain in 5 steps

To get started you need a folder, 3 text files and an AI program. The 5 steps take about 1 hour together, plus installing the AI program. You don't need to program for this.

Step 1: Create a folder (5 minutes)

Create a folder on your computer, for example "My Company", with 3 subfolders: "Projects", "Knowledge" and "Templates". You don't need more structure at the start. New folders emerge later from the work, not on the drawing board.

For reading and writing, I recommend Obsidian. You open the folder there as a “vault” and see your files on the left and the text on the right. According to its own help, Obsidian stores all notes as plain text files in the local folder, and you can edit them with any other text editor. According to its price list, the program is also free for work use.

This is what the start looks like: 3 folders and the first files.

Step 2: Write the file about yourself (20 minutes)

The file ueber-mich.md tells every AI who you are and how you work. It replaces the explanation you would otherwise type in every new chat. Write in it what a new employee would need to know on their first day.

File ueber-mich.md
# About me

## Who I am
- Name, company, location
- What my company offers and for whom

## How I work
- Answer briefly and concretely, without a preamble.
- When in doubt, make a sensible assumption and state it.
- Ask before you delete or send anything.

## What annoys me
- Clichés and marketing language
- Follow-up questions you can answer yourself

## Current status
- What I'm working on right now (with date)

Write the file in your own words, ideally dictated. It may be unfinished. The last section with a date is important, because your focus changes and the AI would otherwise work with an outdated state.

Step 3: Create the instruction file AGENTS.md (10 minutes)

The file AGENTS.md sits at the very top of the folder and tells the AI program how to work in this folder. It is the reason why switching programs later costs little: according to agents.md, the standard is maintained by the Agentic AI Foundation under the Linux Foundation, and many programs read the same file.

File AGENTS.md
# Rules for this folder

First read the file ueber-mich.md.

## Filing
- Projects are located under Projekte/<Name>/, each with a file Status.md.
- Lasting knowledge goes in Wissen/, templates in Vorlagen/.
- Don't create any new folders at the top level without asking me.

## Way of working
- If I name a project, first read its Status.md and tell me in 2 sentences where we stand.
- After every completed step, enter in Status.md: status, next step, date.
- Delete nothing and send nothing without my yes.

Keep the file short. Every line is read at every start, and a long rules file costs the AI attention. Mine was around 35,000 characters long in September 2026. Today it is less than half that, and everything that only one area needs is in that area's rules file.

If you use Claude Code and the folder already contains a file CLAUDE.md, or if Claude Code doesn't read your AGENTS.md, write the line @AGENTS.md at the very top of the CLAUDE.md and create the file if it doesn't exist yet. Then Claude Code loads your rules from the same file as the other programs, and what applies only to Claude Code goes below it.

Step 4: Complete the first task with AI (15 minutes)

Install the AI program following the manufacturer's instructions and sign in with your account. Then, in the program, choose the folder “Mein Unternehmen” as the working folder. If you work in the terminal, first switch to this folder and start the program there.

Then give the program a real task from your week. Beforehand, create a folder for this project under “Projects” with a file Status.md and write 3 lines in it: what it is about, what is already done and what comes next.

Then say, for example: "Continue with Website Tischlerei." The program reads the instruction file, finds the status file and replies with the current state before it gets started. At this point you notice the difference from an empty chat.

The program first reads your files and then replies with the status.

Step 5: Record the status after every piece of work (2 minutes per session)

A second brain only grows if something ends up in the files at the end of every piece of work. The rule for that is already in your AGENTS.md: the AI enters the status, next step and date in the status file itself. You read it over briefly.

The same goes for mistakes. If you correct the AI a second time in the same place, have it write the correction into a file as a rule. A correction in the chat is forgotten tomorrow; a rule in the file applies to every future session.

The tools: Obsidian for reading, an AI program for working

You need 2 tools: one to read the files, and one for the AI to work in them. For reading, Obsidian or any text editor is enough. For the AI program, what matters most is which subscription you already have and how important independence from the provider is to you.

ProgramManufacturerAI modelsReads as instruction fileFits if
Claude CodeAnthropiconly ClaudeCLAUDE.md, without this file AGENTS.mdyou have a Claude subscription
CodexOpenAIOpenAI modelsAGENTS.mdyou have a ChatGPT account
OpenCodeAnomaly, open sourcemany providers, including local modelsAGENTS.md, otherwise CLAUDE.mdyou don't want to be tied to a single provider
Hermes AgentNous Research, open sourcemany providers, including local modelsAGENTS.md and othersyou want an agent to run for you continuously

The details come from the manufacturers' documentation, as of November 2026. I work with Claude Code every day myself, and since November 2026 also with Hermes Agent. I did not test OpenCode and Codex myself for this guide.

2 programs that are often mentioned are not a first step. According to its maker, T3 Code is an interface that controls existing programs such as Claude Code or Codex, and it needs one of them to be signed in. It is built for programming work. Claude Cowork works without a terminal and, on the desktop, also works with connected folders, but according to Anthropic it runs on the provider's servers and stores sessions and files in the Claude account.

My second brain: one folder, 2 AI programs

My second brain is a folder with around 20,000 text files that Claude Code and Hermes Agent use at the same time. It holds project statuses, customer knowledge, brand profiles, templates and 57 fully written workflows, from quote to newsletter. I open the folder in Obsidian, and the AI programs read it directly.

I call my about-me file christoph.md. Claude Code loads it via its own rules file, and all other programs read it first. For me, the file AGENTS.md is a pointer to the same rules file, so that I only maintain rules in one place.

The hurdle with the second program: When I connected Hermes Agent to the same folder in November 2026, the knowledge and workflows came along immediately. The locks did not. In Claude Code, a small script of mine checks before every write whether the AI is creating a text for a customer without having read the brand profile, or inventing a new top-level folder. These locks belong to the program, not to the folder. So I keep writing customer texts in Claude Code until the locks also run in the second program.

The lesson from this: everything that is written as text in the folder moves with you. Everything you set up in the program, you rebuild when you switch. The more is written as text in the folder, the freer you are.

What is stored as text in the folder moves with you. What is set in the program stays there.

For individual questions, a ready-made product is enough

A folder of your own is not worth it for everyone. If you use AI only for individual questions and texts, a project in ChatGPT or Claude is enough, and the effort for folders and rules does not pay off.

3 signs point to your own folder: you explain the same thing to the AI again and again. You work on projects that run for weeks. Or you want your knowledge to survive a change of provider.

2 limits you should know about. A folder of text files is not a backup, it needs one, especially when an AI writes in it. How that works is explained in the guide on backup strategy in the age of AI. And a second brain for one person is not yet a knowledge system for a company: as soon as several people work in the same folder, you need permissions, an approval step for changes, and rules for confidential data.

From your own folder to the company's knowledge

The step from a personal second brain to company knowledge is an implementation in the company, not another folder. We clarify which knowledge goes in, who may see and change what, how customer data is protected, and how your team works with it.

I set this up for companies and train the employees: knowledge management with AI. If you'd rather keep reading on your own first, introducing AI in your company shows the order of the first steps.

Frequently asked questions about the second brain with AI

What is a second brain?

A second brain is a system outside your head in which you store and find knowledge, decisions and the status of your projects. Tiago Forte popularized the term with the book “Building a Second Brain”. With AI, another task is added: the system is no longer just storage, an AI program reads it and works with it.

Do I need Obsidian for a second brain with AI?

No. Obsidian is a window onto a folder of text files and makes reading and linking convenient. According to Obsidian, the files are plain text in Markdown format and can be opened and edited with any other text editor. The AI program doesn't need Obsidian; it reads the folder directly.

Which AI program is right for getting started?

Pick the program that goes with the subscription you already have: Claude Code with a Claude subscription, Codex with a ChatGPT account. If you don't want to tie yourself to one provider, OpenCode is the choice with the least lock-in, because according to its own documentation it works with many providers and with local models. The files in the folder are the same in all cases.

Does my data stay on my computer?

The folder stays on your computer. But what the AI program reads from it goes to the model provider for processing. For Claude Code, Anthropic writes that inputs and responses go to the model over the network. For personal accounts, Anthropic only uses the data for training if the setting for it is turned on.

Can I export my knowledge from ChatGPT or Claude?

Only in part. According to Anthropic, the Claude data export contains the conversations and account data, but not the original uploaded files. An export of project knowledge is not described on the help page. If your knowledge is stored as text files in your own folder from the start, you don't need an export.

What is the AGENTS.md file?

AGENTS.md is a text file in the folder that tells an AI program how to work there. According to agents.md, the standard is maintained by the Agentic AI Foundation under the Linux Foundation and is used in more than 60,000 open-source projects. Because many programs read the same file, you don't have to rewrite your rules when you switch programs.

Sources

  1. Forte Labs (Tiago Forte): “Building a Second Brain: The Definitive Introductory Guide”, fortelabs.com, accessed on 19 November 2026
  2. Obsidian Help: "How Obsidian stores data", obsidian.md, accessed on 19 November 2026
  3. Obsidian: "Obsidian Pricing", obsidian.md, accessed on 19 November 2026
  4. Agentic AI Foundation, Linux Foundation: “AGENTS.md”, agents.md, accessed on 19 November 2026
  5. Anthropic, Claude Code Docs: “How Claude remembers your project (section AGENTS.md)”, code.claude.com, accessed on 19 November 2026
  6. Anthropic, Claude Code Docs: “Data usage”, code.claude.com, accessed on 19 November 2026
  7. Anthropic, Claude Help Center: “How can I export my Claude data?”, support.claude.com, accessed on 19 November 2026
  8. Anthropic, Claude Help Center: “Get started with Cowork”, support.claude.com, accessed on 19 November 2026
  9. OpenCode Docs: "Rules", opencode.ai, accessed on 19 November 2026
  10. OpenCode Docs: "Providers", opencode.ai, accessed on 19 November 2026
  11. OpenAI, Codex Docs: "Custom instructions with AGENTS.md", learn.chatgpt.com, accessed on 19 November 2026
  12. Nous Research, Hermes Agent Docs: "Context Files", hermes-agent.nousresearch.com, accessed on 19 November 2026
  13. OpenAI: "Pricing (Codex in den ChatGPT-Plänen)", learn.chatgpt.com, accessed on 19 November 2026
  14. Nous Research, Hermes Agent Docs: "Hermes Agent FAQ", hermes-agent.nousresearch.com, accessed on 19 November 2026
  15. Nous Research: "Hermes Agent (Repository, Lizenz MIT)", github.com, accessed on 19 November 2026
  16. Anomaly: “OpenCode (Repository, MIT license)”, github.com, accessed on 19 November 2026
  17. Ping: "T3 Code (README)", github.com, accessed on 19 November 2026

Read more

  1. AI Brain Fry: the study and working in rhythmWhat the study on AI Brain Fry shows, what its figures do not prove and how you run several AI runs in rhythm.Way of working · Updated 10 November 2026
  2. Create an anti-visionDecide more clearly with a list of the things you don't want, and how AI helps with that.Way of working · Updated 22 October 2026
  3. Constant availability: 15 studiesPhone, notifications and switching off: what the research shows and what follows from it.Way of working · Updated 21 October 2026

All posts in the guide

Contact

Let's talk.

About AI in your business, your network or your visibility. A short conversation is usually enough to see where things get stuck.

Christoph Harrisonavailable for new projects