May 26, 2026
12 mins
What you're actually doing when you "use AI"
You've had hundreds of conversations with ChatGPT and Claude. Research for a pitch last March. A content strategy workshopped for a client in July. A framework you built at 11pm that you thought was brilliant and never opened again.
All of it is sitting in a database. Inert. Unsearchable. Tagged with dates and nothing else.
Every time you open a new chat, the AI starts from zero. It doesn't know you worked on that pitch. It doesn't know your content frameworks. It doesn't know you've already done the research. You brief it from scratch, every single time, and then the conversation dies and takes everything with it.
You're not building anything. You're spinning the wheel and wondering why the output always feels generic.
The AI isn't the problem. The setup is.
The shift most people never make
Search engines are retrieval tools. You ask, they fetch. The knowledge stays in the index, not with you.
What AI actually makes possible is a thinking partner with a memory. But only if you build the memory. It doesn't come default. Nobody ships you a second brain out of the box.
According to Andrej Karpathy's widely-shared knowledge base proposal, the core idea is simple: give your LLM a folder of markdown and let it organise your thinking for you. His post hit 16 million views in 48 hours. Not because it was technically complex. Because it named something everyone already felt.
The people getting the most out of AI right now aren't the ones with the best prompts. They're the ones who built infrastructure. A place where knowledge lives, connects, and compounds. Where their past work informs the next thing instead of disappearing.
That's the gap. Not AI capability. Not prompt quality. Infrastructure.
What the AI second brain infrastructure actually looks like
Three components. One folder on your desktop. Thirty minutes to set up.
Component 1: Your AI Brain
Export your entire ChatGPT and Claude history. Dump it into Obsidian. Point Claude Code at the folder and tell it to organise everything into tagged, linked markdown files by topic, project, person, and theme.
What you get: two years of your own thinking, made searchable. Cross-referenced. Wikilinked so you can see how ideas from six months ago connect to work you're doing today.
I did this and stared at the graph view for fifteen minutes. Client research from last year, linked to a project I'm actively building now. A topic I'd circled back to six different times without realising.
Component 2: The Living Wiki
The brain is your past. The wiki is your future.
Two folders: raw and wiki. Drop source material into raw: articles, PDFs, meeting notes, transcripts. Claude Code reads the raw folder and writes structured pages into wiki. Summary pages. Topic pages. Cross-references. An index.
You don't write a single wiki page yourself. You add sources and the wiki grows.
The setup takes one prompt. The habit takes one minute. Drop a source, tell Claude Code to update, done. Next time you research that topic, you're building on everything you've already read.
As Alex Freedman documents in his workflow, this approach scales to hundreds of interconnected articles without you writing a single one yourself.
Component 3: The Heartbeat
A static wiki is useful. A wiki that updates automatically is a different thing entirely.
MCP connectors let Claude Code pull directly from your apps: Gmail, your meeting notes tool, Google Drive. Every email thread, every call transcript, every forwarded article becomes a potential source without you dragging files around.
And then there's iMessage Channels. Connect iMessage to your Claude Code session and you can text yourself from anywhere. Claude Code reads the message, checks it against your entire vault, and replies in the same thread.
On the train. In a meeting. At 11pm with an idea you don't want to lose. Your second brain, in your pocket.
The compounding problem with how most people work
Here's what happens when you use AI like a search engine for two years.
You have a lot of useful conversations. Zero accumulated knowledge. Every project brief starts from scratch. Every research session rediscovers things you already knew. Every piece of content draws on the same surface-level prompting instead of the two years of context you've actually built.
Here's what happens when you build the infrastructure instead.
Month one, the wiki is small. Month three, you start noticing it already knows things. Month six, you're not briefing Claude from scratch anymore. You're pointing it at your own knowledge base and asking it to build on what's there.
The output quality isn't different because the AI got better. It's different because your context got richer.
That's the compounding effect most people are missing.
The practical version
Start with one component, not three.
If you want to understand your past: export your conversations and run the AI Brain. One afternoon. Done.
If you want to start accumulating knowledge from now: set up the wiki. Ten minutes. Drop in five sources. Run the first update.
If you want the full system: add the connectors. Connect Gmail. Connect your meeting notes tool. Set up a slash command or two.
As Greg Isenberg breaks down on YouTube, twelve slash commands built on an Obsidian vault can automate most of your daily knowledge workflow. Start with three.
The system compounds every day you run it. The best time to start was when you first opened ChatGPT. The second best time is now.
One more thing
Most people read something like this, think "that's smart," and do nothing.
It's not laziness. It's that "build a second brain" sounds like a project. Something to plan for when you have time.
It isn't a project. It's a folder and a prompt.
The hardest part is the thirty minutes you spend on it this afternoon. Everything after that runs itself.
Frequently Asked Questions
What is an AI second brain?
An AI second brain is a personal knowledge system where your AI conversations, research, and notes are stored, tagged, linked, and queryable instead of disappearing after each session. Tools like Obsidian, Claude Code, and MCP connectors let you build one that grows automatically every time you add new information.
Why does AI output feel generic even with good prompts?
Because the AI has no context about you, your work, or your thinking. Every session starts blank. Generic input produces generic output. The fix isn't better prompts. It's a richer knowledge base that gives the AI real context to work with.
How long does it take to build an AI second brain?
The initial setup takes about thirty minutes: export your AI conversation history, load it into Obsidian, and run Claude Code to organise and tag everything. The living wiki takes another ten minutes to configure. After that, maintaining it is a one-minute habit per new source you add.
What tools do you need to build an AI second brain?
The core stack is Claude Code (CLI), Obsidian (free), and your exported conversation history from ChatGPT and Claude. For the living wiki you need a raw folder and a CLAUDE.md instruction file based on Karpathy's gist. For live updates, MCP connectors link Claude Code to Gmail, Granola, Google Drive, and other apps you already use.
Charlie Hills
