The Best AI Integration Is a Folder

I have about 11,000 notes in Obsidian. Project outlines, article drafts, meeting notes, random ideas I captured at 2am. Claude Code can read all of them.

No plugin. No API key. No integration marketplace. Claude is a terminal tool that reads files. Obsidian is a UI for a folder of Markdown files. Point one at the other and everything just works.

I didn’t plan this. When I moved to Obsidian, I was solving a different problem: getting my notes out of Evernote before they became unreadable. The AI part was an accident. A good accident, but an accident.

What the workflow actually looks like

I write notes the way I always have. Fragmented, messy, full of half-finished thoughts. An idea in one file, a counterargument in another, a quote I saved six months ago that I half remember.

When I’m ready to write something, I point Claude at the vault and say, “find everything I’ve written about X.” It greps through the files, pulls out the relevant notes, and shows me connections I’d forgotten about. Sometimes it finds threads I didn’t know were there. Three notes from different months that were circling the same idea.

From there, I can ask it to draft an outline from those notes, or reorganize a messy document, or find holes in an argument I’m building. It’s working with my own thinking, stored in my own files, on my own machine.

The whole thing runs in a terminal. No cloud sync involved in the AI part. No data leaving my laptop unless I choose to send it. The notes stay where they are.

What it’s good at

Finding threads. I have notes scattered across months of writing. Claude can search the entire vault and show me things I’d forgotten. Not because it understands my thinking, but because it can read 11,000 files faster than I can.

Restructuring. I write messy first drafts. Long paragraphs that cover three ideas. Sections in the wrong order. Claude is good at reorganizing existing text without losing the content. The structure gets better. The words stay mine.

Outline flow from fragments. I have scattered bullet points, half-formed ideas across a dozen files, notes that don’t connect yet. Claude can pull those together into an outline that actually flows. Not a finished piece. A shape I can work with.

Rubber ducking. Describing what I want to write and having something respond with questions and pushback is more useful than it sounds. It’s not a colleague, but it’s better than talking to myself.

What it’s bad at

Voice. This is the big one. Claude writes clean, competent prose that sounds like nobody in particular. If I let it draft without heavy editing, the result reads like a well-written Wikipedia article. That’s not what I want. I want it to sound like me, which means every sentence Claude produces gets rewritten or thrown out. The value is in the structure and the research, not the final words.

Grammar. This one surprised me. You’d think a language model would at least get the language part right. But I find grammar mistakes all the time. Comma splices, dangling modifiers, subject-verb disagreements buried in long sentences. It writes confidently enough that you don’t notice until you’re reading it aloud and something trips you up. You still have to proofread everything, even the basic stuff.

Over-connecting. Point it at a bunch of notes and ask for themes, and it will find themes whether or not they’re actually there. It’s a pattern-matching tool, and it would rather find a weak connection than admit there isn’t one. You have to push back.

Knowing when to stop. Ask Claude to improve a paragraph, and it will keep improving it until all the edges are sanded off. The rough parts of your writing are sometimes the point. You have to know when “that’s enough” means closing the terminal.

The plugin trap

Every note-taking app is rushing to add AI features. Notion has Notion AI. Evernote has an AI assistant. Apple is adding it to Notes. The pitch is always the same: we’ll make your notes smarter.

But these companies are building AI features that only work with their data, in their format, through their interface. Notion AI works on Notion documents. Evernote AI works on Evernote notes. You can’t swap in a better model when one comes out. You can’t run it locally for privacy. You can’t use it with a different tool.

It’s vendor lock-in wearing a new hat. The same pattern I ran into with Evernote, repeated with an AI label on it. Your notes become more dependent on the platform, not less.

Obsidian has AI plugins too. The difference is that they’re optional, and the underlying data is still just Markdown files. If every Obsidian AI plugin disappeared tomorrow, I could use any other AI tool that reads text. The lock-in lives in the format, not the feature, and Markdown is nobody’s proprietary format.

The accidental advantage

The people who chose plain text years ago, before AI tools existed, are now in the best position to use them. Not because they predicted this. Because open formats are compatible with everything, including tools that didn’t exist when the choice was made.

This is the text files argument extended one step further. Text files were already durable, scriptable, and greppable. Now they’re also the easiest thing for AI to work with. Every advantage compounds.

If your notes are in a proprietary database, your AI options are whatever the vendor decides to offer you. If your notes are in a folder, your options are everything.

The broader pattern

This isn’t only about notes. It’s about data formats in general.

Code in git repos works well with AI tools because it’s text files in folders. So do dotfiles, Markdown docs, and plain-text accounting ledgers. Anything that lives in a filesystem and can be read by cat is ready for AI without any integration work.

The things that don’t work well are databases behind apps, proprietary document formats, data trapped in SaaS platforms. The same things that were already hard to search, hard to script, and hard to migrate.

AI didn’t create a new divide. It widened an existing one. The people who valued portability and openness got an unexpected bonus. The people who traded control for convenience got another bill.

Credit is useful. But the value of what you build with it depends on what you own.

What’s your note-taking setup? I’m curious whether AI tools have changed how anyone thinks about file formats.