Obsidian plus AI, three levels deep

Most people treat Obsidian as a fancy note-taking app. Turns out the real value shows up when you stop writing notes yourself and let AI do the writing. I just watched a 24-minute breakdown from Tina Huang, the creator behind the AI education company Lonely Octopus, and it flipped how I think about “second brains.”

Here’s the twist that hooked me. The author admits she rarely takes notes at all. She rates herself a three out of ten on the note-taking scale, and when she does learn something, she dumps it into one giant Google Doc. And yet her whole company documentation system and her personal productivity setup run on Obsidian. The reason is simple: Obsidian happens to be almost perfectly shaped for AI.

The key idea: Obsidian is just a folder

At its core, an Obsidian “vault” is one local folder full of plain markdown (.md) files. That’s it. Any software can open them, you can sync them through Dropbox or Google Drive, and the only real paid feature is Obsidian Sync, which is end-to-end encrypted. The app is so private that its own CEO can only estimate the user base at 5 to 10 million.

The expert speedruns the main features, and three of them matter most for AI:

• Plain markdown files. An AI agent can read a folder of text files with zero setup. No export, no API, no database.
• Links between notes. Type two brackets around a phrase and you’ve created a connection. The graph view shows how everything relates, including “stubs” for notes you haven’t written yet.
• Flexible metadata and plugins. You can add all kinds of metadata, turn notes into tables or Kanban boards, and plug in community tools like Smart Connections or Copilot.

Those three things turn a note app into something an AI can actually navigate and maintain.

Old way vs new way: the three levels

The traditional second brain, popularized by Tiago Forte, works like this: your brain is for thinking, not storing. So you write everything down, then go search for it later. Bobby’s favorite color is yellow, you save that, you find it when gift season comes. Simple, but you do all the work on both ends.

The creator lays out three levels that change who does the work.

🧠 Level 1: AI second brain

You still write the notes. AI retrieves, interprets, and analyzes them. The author points Claude Cowork at her vault of YouTube production notes and asks, “How do I come up with a video concept?” It reads the whole vault and answers with references to her own files. When she wants to build something, she switches to Codex, which turned her notes into an “Is this worth filming?” checklist. For full privacy, she suggests an open source agent harness like Hermes or OpenCode running a local model.

📥 Level 2: AI database

This is the reverse. AI writes into the vault, you read what you want. This is where the note-averse author lives. Her “Life Bot” logs a photo of green tea as zero calories. Her “Taco Bot” writes process documentation for the company. Desktop widgets drop Pomodoro logs and to-do lists into the same folder. Agents run on her laptop, her Mac Studio, and a VPS, all synced through Obsidian Sync. Then she asks another agent, “How do I improve my deep work sessions based on my personal data?” and gets specific advice: when to work, which tasks to sequence first, when to break. Her go-to models for the writing side are Gemini Flash, DeepSeek V4 Pro, and Qwen 3.6 36B running locally.

📚 Level 3: LLM Wiki

This is Andrej Karpathy’s pattern, and it’s fully AI-managed. AI writes the notes, organizes them, and answers questions. You never touch the vault. Karpathy’s framing: Obsidian is the IDE, the LLM is the programmer, the vault is the codebase. The creator demos her Hermes wiki, where she pastes a resource about Kanban boards into a Telegram bot, the agent ingests it, and later she asks “How do I use a Hermes Kanban board?” and gets a step-by-step answer built from her own wiki.

Under the hood there are three layers:

1. Raw sources. Articles, papers, transcripts, your own notes. Immutable. The AI reads these but never edits them.
2. The wiki. LLM-generated summaries, entity pages, comparisons. The AI owns this layer completely.
3. The schema. A file like CLAUDE.md or AGENTS.md that tells the model how the wiki is structured.

And three operations: ingest (drop in a source, the AI files it), query (ask questions, the AI synthesizes), and lint (a health check that finds contradictions, stale claims, and orphan pages). Two special files hold it together: an index.md that catalogs every page, and an append-only log.md that records everything the AI did.

Practical steps to get started

The author’s advice is to not jump straight to Level 3. Here’s the path I’d follow based on her breakdown:

1. Make a vault. Create a folder, open it in Obsidian, write a few notes with bracket links. Ten minutes, tops.
2. Point one AI at it. Claude Cowork, Claude Code, Codex, or a local Hermes agent. Ask it a question only your notes can answer.
3. Let something write to it. Start with one bot logging one thing: food, workouts, meeting summaries. This is the level the expert recommends first if you want a rich dataset fast.
4. Graduate to a wiki. Karpathy’s gist is deliberately vague, so you can paste it into almost any coding agent and it’ll implement it. Hermes has a ready-made “LLM wiki” skill, and there are many Claude Code repos you can clone.

Things to watch out for

• Privacy. Cloud models see everything in your vault. If that bothers you, the local model route exists for a reason.
• Drift. A wiki that ingests for months will accumulate contradictions. Run the lint step regularly or the AI starts answering from stale pages.
• Index bloat. The index.md is what lets the AI find the right page. As the vault grows, that file matters more, not less.
• Starting too big. The creator’s own team is still at Level 2 and only now exploring Level 3. That’s a healthy pace.

I think the real insight here is that “second brain” stops being a discipline you have to maintain and becomes infrastructure that maintains itself. That’s the part I can’t stop thinking about!

If you want to see the actual demos, the summary slides, and the little quiz at the end, go watch the full video. It’s worth every one of the 24 minutes.