AI in Obsidian: how I actually use it

TL;DR: I don’t use any AI plugins in Obsidian. My setup is plainer than that: Obsidian holds my notes as ordinary markdown files, and Claude Code reads and writes those same files from outside the app. Obsidian is the window I look through when I want to view a file. The files are the system, the AI works on the files, and I talk to the AI to pull together and distill what’s in them.

I’m Julie Kaiser, a science writer. This vault runs a huge portion of my freelancing work and a lot of my hobby projects, and I work in it extensively, daily. Here’s what it actually looks like – capture, retrieval, writing – plus straight answers to the questions people ask when they’re new to it. The vault is the knowledge layer of a bigger system; the whole setup is in How I built an AI that actually knows my work.

A note on names before we start. People call a setup like this a second brain, a personal knowledge management system (PKM), a Zettelkasten, a digital garden, a personal wiki. I call mine an LLM wiki – written for me, written by an AI, and managed by an AI – and I named it the Julieverse. Underneath the name it’s a vault in Obsidian, and it’s just files on my computer. If you’re building one of these things, whatever you end up calling it, this is what mine looks like.

What I actually use Obsidian for

I’m so new to Obsidian, and my use of it is almost embarrassingly plain. Three things, in order of how much they matter to me:

It’s a markdown editor with the whole vault in view. The file tree sits in the side pane, I click into a note, read it, change it, add a new one. That’s the most important thing it does for me: it lets me see, touch, and edit the files my system is made of – without going through something like Word.

It’s where Claude leaves work for me. Claude writes me to-do lists and checklists in the vault – a file of things I need to review, questions waiting on my answer – with a link straight to it. I click the file, the questions are right there, and I give my feedback on the spot.

The graph view is fun. The map of notes and links makes the vault feel like a place. I won’t claim it runs my workflow – honestly, it’s not useful to me at all – but it’s what drew me to the idea of building an LLM wiki in the first place: in the example I was watching, I saw the pieces of information, and how Claude connected them and put them into files it could later query – files you could interact with. That was the most amazing idea to me.

And one thing to be clear about up front, because it’s the heart of this setup: Claude doesn’t need Obsidian. If I uninstalled the app tomorrow, Claude Code would keep reading and writing the same files – I’d have to install some other markdown viewer to see them. Obsidian is for my eyes; the files are for both of us.

What Claude does with my files

This is where the arrangement pays off.

Retrieval. This is the one I’d keep if I had to drop the rest. For years I had a recurring pain: I know I did something great, it was really useful, and I can’t ever find it again. Now I describe what I’m looking for – roughly, badly, “that thing about audience fears from a few months ago” – and the AI finds it in the vault. I don’t have to remember what I named it or where I filed it. My brain can’t hold everything I’ve captured. The vault can.

Capture. Ideas are fleeting – if you don’t catch one at the moment it arrives, it’s gone, and for me it’s hard to get it back. So I dictate voice memos on my phone, they land in a folder, and Claude files each one where it belongs: this one’s a content idea for this project, that one’s an insight about a science writing project. Even before anything gets used, the capture alone takes the stress off. Sometimes I never come back to a memo. This system doesn’t punish that.

Writing (well, drafting actually). When I draft, the AI pulls what the vault already knows – the research I saved, the decisions I made in March, the position I locked last month – and puts it in front of me. Then we draft, and I edit hard. It handles the blank page and the polish; I bring the truth in the middle. And Obsidian is where I do that hard editing: I click into the file, read it, change it, and I don’t have to click save – my changes are already in.

Can Claude use Obsidian?

Claude doesn’t use Obsidian, and it doesn’t need to – because an Obsidian vault is a folder of markdown files on your computer. No database, no proprietary format.

Here’s what I do: I open Claude Code in the Claude desktop app, pointed at my Obsidian vault. It can find notes, create them, reorganize them – and it manages that folder according to the “rules” we established when it built the wiki, based on the Karpathy idea (that build story is its own post). When Claude writes a file, the note appears in Obsidian, because Obsidian is a viewer on my Finder folder: the file appears in my Finder, and I can also see it in Obsidian. Claude works on the files. I work with Claude. Obsidian is how I look at what we did.

Remember, it’s just files on my computer. If I switch AI tools next week, the vault doesn’t care. Nothing is trapped or tied to a particular tool.

That’s why I have, so far, skipped all the plugins people mean when they say “Obsidian AI” – the chat sidebars and related-notes panes inside the app. The plugins are community add-ons, mostly free to install, but none of that is useful to me, because Claude accesses my files directly. Obsidian is free for personal use, and the vault you create is yours forever.

What a day in it actually looks like

Morning: I wake up early with some grand idea in my head, so I pick up my phone and dictate it before I get out of bed. Then I lie there and listen to the birds while the rest of me wakes up, my stress relieved – my brilliant idea is already being transcribed and sent to a folder in my vault.

Working: I open Claude Code in the vault and start with what it already knows. No re-explaining who I am or what the project is. I tell Claude there’s a new voice memo in the vault, what it’s about, and what I want to do with it. I could automate the processing of those files, but I don’t – I usually want to do something with a memo right when I’m working with Claude, and I don’t want to show up to already-processed information. But that’s me. You can set that up; I don’t.

End of a session: nothing to rescue, because the way mine is set up, the files are the wiki – Claude saves and documents as it goes. When I tell it I’m done, it goes back through the session and harvests the learnings – lessons, what we could have done better – and writes those to the wiki too.

One admission to keep this honest: none of this was instant. Early on, I was so paranoid that the system wasn’t saving things that I was constantly asking, wasting tokens: did you save that? Did you save that? Later I adapted the system so it’s constantly saving and documenting on its own. Now the payoff is huge: a wiki full of useful information – ideas, topic research, my expertise, my experiences – and when I want to draw a little bit from everywhere to make something, I do it by asking Claude.