Build an AI Brain

Your AI work is everywhere. All over the shop.

ChatGPT has one bit. Claude has another. Codex has half a project. Your phone has the voice note. GitHub has the code.

Sure you are doing a lot. Very productive. Lovely.

But the more you create the more mess you make.

The problem is that none of these things automatically know what the other ones are doing. So you become the USB stick. You are the one copying context from one model to another, re-explaining your business, re-explaining the client, re-explaining the project, re-explaining the weird decision you made last Tuesday because Future You apparently enjoys admin.

That’s precisely the opposite of what we want to use AI for. We are just making MORE work for ourselves.

This is why people are suddenly talking about AI brains, Obsidian vaults and personal operating systems. It looks new. It is sorta new. But (as always) the genuinely useful bit is not the flashy viral stuff.

It is mostly folders and markdown files.

Ho hum.

Obsidian is not the brain

You’ve probably seen posts about Obsidian. And think I’m going to focus on Obsidian.

Obsidian is cool. I like Obsidian.

But Obsidian is the HUMAN interface. It is the pretty layer. It gives you the graph view, backlinks, search, daily notes, plugins and all the nerdy little goodies that make a folder of text files feel like a system.

For you, that is useful.

For the AI? Less so. They don’t NEED this.

Claude Code, Codex and the rest do not need a lovely purple graph of your knowledge. They need the actual files. They need the README. They need the decisions log. They need the project notes. They need the bit where you wrote "my god do not touch this payment flow because it breaks checkout!!1!" three months ago and then forgot about it.

That is the brain.

And it’s basically a folder, some sub folders and some text files. That’s it.

Boring bit that gets things done

The base layer here is a markdown file. .md.

Nothing to do with doctors. No stethoscope required.

Markdown is basically a plain text file with a bit of simple formatting. Humans can read it. AIs can read it. GitHub can track it. Obsidian can display it. Claude Code and Codex can scan it before doing work.

Markdown is not sexy. It is aggressively unsexy. It is a Word document with the formatting fluff removed and, frankly, thank you very much.

But boring is useful here. You do not want the memory of your business trapped inside one app's proprietary format. Or heavy files like PDFs and Word docs.

You want files that can be copied, searched, versioned, backed up and read by whatever AI tool you use next month when this week's favourite app inevitably changes its pricing or does one.

The useful files are simple:

• README.md for what the brain is

• AGENTS.md for how AI tools should behave

• CLAUDE.md if you want Claude-specific instructions (hint, point it at AGENTS.md instead so you have one set of instructions)

• tasks.md for open loops and todos

• decisions.md for things that should not be re-debated

• project READMEs for active work

• agent logs for what AI tools actually did

You can of course add more. But generally something simple like this is all you need.

Also don’t worry about how to make all this - AI will do it for us shortly.

Access anywhere

Where should this folder and set of documents live?

We will connect ALL our AI tools to the same set of folders.

So: we need it accessible anywhere. In the Cloud.

Someone asked on the stream whether this could live in Google Drive.

Technically, yes.

I would not.

If you are using AI tools properly, you will soon have multiple agents working on different bits of your life or business. Codex doing one thing. Claude Code doing another. Maybe one thread sorting your newsletter, another checking your website, another helping with a client project.

If those tools all bash away inside the same Google Doc, it gets messy fast. Conflicts. Random edits. "Who changed this?" energy. Lots of faff.

GitHub is built for this kind of thing.

I’ve done another full guide on Github 101 previously so go check that if have no idea what Github is.

Simply put though it’s for online file storage in the cloud. It tracks changes. It lets tools work in branches or worktrees. It gives you a history of decisions. It means you can see that on June 6th you changed the website, added a rule, moved a file, broke something, fixed it, then told Future You never to do that again.

So: use Github here to make your life much easier. Bonus - it’s free.

Do not copy someone else's brain

This is the bit people will ignore - I know some of you will still go and find some guru’s “perfect AI brain” template. Maybe even buy it! Don’t!

You should not start with someone else's template.

I know. Annoying. Templates are comfy. They make the work feel done before it is done. You download a folder structure, put it in Obsidian, stare at the cool looking knowledge graph and think: "Ah yes, productivity."

Nope!

Your vault needs to map to YOUR actual life. Your roles. Your projects. Your clients. Your products. Your weird repeated context. Your privacy boundaries. Your working style.

Your mess. Your very own chaos!

That is why the starting point should not be a template. It has to start with you.

The starting point therefore should be an interview.

Let the AI ask you questions. One at a time. Then answer with your microphone. Yap for 20 minutes. Tell it about the projects, the priorities, the stuff you keep forgetting, the workflows you repeat, the things you definitely do not want it to touch, and the boring context you are sick of retyping.

I personally did this over DAYS when setting my systems up. I kept remembering important things and yapping some more!

The AI can then propose the folder structure for you. It will make up the files, the project folders, the whole shebang for you.

Much better.

Steal this prompt

Use Codex or Claude Code for this. ChatGPT can help design the structure, but it cannot reliably create the folder tree and files on your machine unless you connect it to tools. Codex is free if need to just getting started

Here is the basic prompt:

I want to create a vendor-agnostic AI vault that acts as a shared brain across my tools.

Use markdown files and folders.

Interview me one question at a time about:
- my daily life and roles
- my work and active projects
- repeated context I keep explaining to AI
- workflows I want AI to help with
- privacy boundaries
- how I like AI tools to behave

After the interview, propose:
- a vault folder structure
- README.md
- AGENTS.md
- project README templates
- tasks.md
- decisions.md
- an agent-log format
- a migration plan

Ask me the first question now.

This will live in Github. 
If I do not have this set up help me.

Then talk.

Seriously: actual talk. Use the microphone. Do not sit there trying to write the perfect answer like a LinkedIn thought leader. Ick.

Just talk. Yap. Unload everything.

The AI will sort the structure out afterwards. That is the whole point. We are not doing artisanal folder design here. Let the AI do the work.

The rule: explain it twice, save it

One issue right now - it’s hard to capture “normal” chats into your Github folders and files. Everyday chatbot Claude and ChatGPT don’t do this by default.

Instead use Codex or Clade Code to make sure everything is captured.

Does this mean you need to always use Claude Code or Codex? Nope!

The daily habit is simple: Anything you explain to AI twice belongs in the vault.

For any one off usage of AI the normal chatbot is entirely fine. Those chats are disposable. For anything you want to capture use your new system.

And the magic thing here is that over time your system gets better. The more context and information it has about you the better your AI will be able to help you.

Not on day one. Day one is mostly a folder and a mild headache as you get it all set up. Sorry…

But after a few weeks, your AI tools stop starting from zero. They stop asking the same questions. They stop making the same bad assumptions. They can read the vault first, understand the context, then do the work.

If your coding tool knows your business it’ll stop making errors that seem “dumb” to you. “What the hell are you doing Codex you KNOW I don’t sell that those services?” sort of annoyances disappear because your AI tools will start to collate info from across different domains of your life.

This is also why chat history is not enough. It doesn’t pass between tools. We need a centralised location for that!

Obsidian is still useful

Full circle!

Should you use Obsidian?

Yeah, probably!

It’ll give you a great human interface to “see” your system.

It is free for the basic local version. It is good. It makes markdown folders easier to browse. It gives you backlinks and search and graph views and plugins and all the nice human bits.

Do you need to pay? Probably not! The paid plan is for synching. But we are using GitHub for that (for free) so we’re good!

Set the GitHub repo up locally, open that folder as your Obsidian vault, and let GitHub Desktop or your AI tool handle the pushing and pulling. Tada! Saved some money.

Again, keep the layers clear:

• folder = the brain

• markdown = the memory

• GitHub = history and sync

• Obsidian = human interface

• AI tools = workers

Grossly simplified. Also useful.

Start small

You do not need a perfect personal operating system by Friday.

Please do not spend three days building a 97-folder cathedral to productivity and then never use it. Productivity porn is appealing. But valueless.

Start with one repo. One README with a basic description. One AGENTS file with simple rules for your AI tools. One decisions file. One task list. One project folder.

Then USE it.

Tell every AI tool to read the vault first. By default they’ll always check AGENTS.md first (that’s its purpose) so put important instructions in there. Again the AI will help you set this up so don’t worry about the how.

Have your AI update the vault when a decision is made. Have it write logs when it changes something. Have it add repeated context instead of letting it disappear into a chat thread you will never find again.

And over time it’ll grow and improve. For now though: keep it simple.

To the Task,

Kyle