Don't use Obsidian for AI knowledge management: hand over all three jobs

Give it credit first, then talk about the problem

Whenever knowledge management comes up, it seems to mean installing Obsidian—and if you haven’t, you look like you’re not serious. I call that tool anxiety.

I’m not going to call Obsidian worthless. Quite the opposite: I’ll start with what it got right. And once I do, you’ll notice something—the things it got right don’t need Obsidian to do them.

A note app, when you get down to it, sells three jobs: store, maintain, view.

Store: it bet right, but the upside goes to the AI

Plain text, file over app. Your data is just your files, not locked in any platform’s database, and you can take it with you when you switch tools. It bet beautifully here, and in the AI era that bet hasn’t aged—it appreciated, because plain text is the easiest thing for an AI to read and write.

But look closely. That “friendliness” is the AI finding it friendly, not you finding it comfortable.

So my verdict is one line: the format won, the software lost.

Maintain: people simply cannot keep this up

Backlinks, tags, MOC, the Zettelkasten card-box method. All of it amounts to making you your own librarian. You build the links by hand, you attach the tags by hand, and you keep feeding an index that only gets bigger.

People cannot do this. I’ll use myself as the sample: I’m an NTJ personality, J to the extreme. Every document since I started working was renamed by hand and uploaded to the cloud one at a time, for over a decade, without missing a day. But I could only keep up because I produced a few documents a day. Even someone as J as me would quit at dozens a day.

So this isn’t a discipline problem. It’s a human-nature problem, and maintenance has to be handed off.

Now the people who go deep will jump in: I already handed it off, I run Claudesidian and let the AI manage my vault. I agree with that—an AI manages Obsidian better than a person does. But go read its own documentation. It says right there that the Obsidian layer is optional, and that folder structure runs fine in an ordinary plain folder.

Which means: for the AI to do the managing, you don’t need the Obsidian layer at all.

View: the one reason it’s still here, and the job it does worst

By now the first two jobs are settled. Storage—the upside goes to the AI. Maintenance—the AI does it, without Obsidian. So what is that software still doing on your machine?

Only the third job: view. It gives you an interface so you can pull things up, use them, and look things over. That’s the one reason it’s still there. And here’s the problem—the one job it was kept for is the job it does worst. Two reasons.

First, md is deeply unfriendly to humans. I burn a few hundred million tokens a day now and produce over a hundred documents a day. At that volume I can’t be bothered to read a single line of md. Don’t jump on that—it doesn’t mean I read less. I read far more than before. I just don’t read md. It was born as a markup language for machines to read and write, so using it as a human reading surface never made sense.

Second, and this one is worse: what you actually work with isn’t only md. Every meeting note, client proposal and research report I have right now is a web page or a PDF. Not one of them is md. Obsidian either can’t handle those or handles them badly. The moment you most need to pull something up and reference it is precisely when it’s least helpful.

The conclusion: entirely redundant

Put those two together and the conclusion is sitting right there. The two jobs it’s good at don’t need it, and the one job left to it, it does badly. On that basis alone, Obsidian is entirely redundant.

What I actually need instead

Maintenance already went to the AI, so it generates the presentation layer along the way. I don’t want a fixed interface. I want something that manages itself: a multi-format viewer and editor, plus a filter-and-retrieval tool.

I’m not trying to open one document. I want to slice the whole pile by contacts, companies, projects, and life goals. How the relations get built and which documents to pull—the AI does all of that, not me.

I’m not the only one saying this. Karpathy’s LLM Wiki is even more direct: the knowledge base is handed to ClaudeCode to maintain and the human only reads. Maintenance is no longer something a person does.

How I actually manage my knowledge base now

One line: I manage it directly with ClaudeCode.

The page opens and 200-odd nodes lay themselves out into a Morandi-coloured starfield. Colour maps straight to file format, so at a glance you can tell which is a web page, a note, a spreadsheet, a slide deck. The bigger the dot, the more information it holds. Click any one of them and a floating window pops up with the file content right there—you can read it, and you can edit it.

The panel on the right is the slicing: contacts, organisations, projects, goals. Click one and the whole graph keeps only the related documents and re-clusters itself. That’s the retrieval tool I wanted—I never have to remember what a document was called.

The most interesting part is the edges. No human drew them. The AI inferred them, searching four levels down: explicit declarations in the document first, then topic membership, then in-body citations, and only then title similarity. So the backlinks of the AI era are written by the AI.

One more thing I love: I create a new file in the terminal, do no tidying at all, refresh the page, and the node grows in by itself. Filing is indexing.

Where did this starfield come from? An earlier episode covered letting the AI tidy local files and rename them automatically, and I open-sourced that. This starfield is the second half of the same job: renaming handles where a file sits, the starfield handles how files relate to each other.

How to build it: not complicated at all

• Directory convention: finished pieces in one directory, notes by topic in another, all plain text
• One iron rule written into CLAUDE.md: declare tags and relations the moment a document is created, not afterwards
• Old documents need no page-by-page rework: one central registry table, one line governs a whole topic

A skill isn’t enough either

At this point you might think, fine, so what’s needed is a skill. But a skill isn’t enough—a skill is still passive. You have to call it.

I want something that moves on its own: an autonomous loop plus a scheduled job. It wakes up on time, sweeps the library, files new documents, fills in the relations, and flags what should be retired. By the time I want to look, it’s already there. The human isn’t in the process.

One step further: online documents are transitional too

Last episode I covered plugging AI into online documents. Now think about why online documents were valuable in the first place. Because an individual couldn’t stand up a platform—you had no way to build something that stored documents well, sent them to other people, and let everyone edit together. So we all had to get on somebody else’s platform.

With AI, that premise is gone. The AI manages the documents, the interface is generated on demand, and sharing means sending out a web page. So what’s left of the platform layer? What I think the future actually needs is some kind of A-to-A protocol, letting my AI talk directly to yours.

Which makes online documents a transitional form too.

What has to change isn’t the software. It’s who does the managing

Back to that tool anxiety. What you’re anxious about is not having installed the right software, but that’s not the real problem. The real problem is that those files of yours are still managed by you.

No software swap fixes that. What has to change is who does the managing. I’ve handed all of mine to ClaudeCode. So the question I asked at the very start no longer applies to me: I don’t go back and dig through my old notes any more. Whatever I ought to see gets delivered to me on its own.

Source: EP0071 digital-avatar voiceover (pinpinvoice0717b voice) · text = final voiceover script (ground truth) · timecodes aligned sentence-by-sentence by gemini-2.5-pro (chunked parallel ASR, 99.0% coverage)

[00:00] Nowadays, the moment you mention knowledge management, it seems like it equals installing Obsidian. If you don’t install it, you look unprofessional. Let me ask you a real question first. When was the last time you opened your note vault and seriously read an old note you wrote yourself? So think about that—we’ll come back to it later. I’m not here to trash Obsidian completely today—quite the opposite, I’ll start by pointing out the things it got right. And once I do, you’ll notice one thing: the things it got right—none of them actually need Obsidian to pull off. Let’s take a note-taking app apart. At the end of the day, it’s really just selling three things: storing, managing, viewing. And that storing, managing, viewing thing—none of it is something only Obsidian can do.

[00:27] First thing—saving. Obsidian got this one right, and got it beautifully right—plain text, file over app. Your data is your files, not locked inside some platform’s database. Switch tools and you can take it with you. Now in the AI era, this hasn’t become outdated—it’s actually gained value. Because plain text is the most AI-friendly—it reads smoothly, writes smoothly. But here’s the thing—this “friendly” is friendly for AI, not friendly for you to use.

[00:54] So my verdict comes down to one line: the format won, the software lost. Second thing: management. Backlinks, tags, MOCs, the Zettelkasten card-box method—let’s be real, it’s just you playing librarian for yourself. You have to build links by hand, tag things by hand, and then hand-nurture an index that keeps growing. Humans just can’t do that. I’m an NTJ personality, J dialled all the way up. Every document I’ve ever produced at work—I’ve manually renamed every single one and uploaded them all to the cloud, one by one, for over a decade without missing a day. But I can pull that off because I only produce a handful of files a day.

[01:21] Even someone like me—as J as it gets—would still give up after a few dozen files a day. So this isn’t a discipline problem. It’s a human-nature problem. Maintenance has to be handed off. Now the folks who’ve gone deeper will jump in and say, “I already handed it off—I’m using Claudesidian.” Fair point. AI manages Obsidian way better than a human does. But go read its documentation. The authors say it themselves: the Obsidian layer is optional. That folder structure runs just fine inside a plain old directory. Which means, when AI takes over, you don’t actually need Obsidian at all.

[01:48] Okay, so up to here, we’ve sorted out the first two things. Storing—its benefit is for the AI. Managing—the AI just handles that on its own, no need for it. So why is Obsidian, this piece of software, still sitting on your machine? It’s for the third thing: viewing. It gives you an interface, a way to go fetch, use, browse your own stuff. That’s the only reason it’s kept around. But here’s the problem: this one thing it’s kept around for—

[02:15] And they’re the worst at it. First, let’s go back to the question I asked you at the start. Right now I’m burning through hundreds of millions of tokens a day, and I generate over a hundred documents daily. At that scale, I’ll be honest—I can’t be bothered to read a single line of md. Don’t jump in to argue just yet. I’m not saying I don’t read. I read way more than I used to. I just don’t read md anymore. The md format is way too unfriendly for humans. It was born as a markup language for machines to read and write. Using it as a human reading interface never made sense in the first place. Second, and I think this one’s even more fatal. The stuff you actually use day to day—it’s not just md.

[02:42] All my meeting notes, client proposals, research reports right now—they’re all web pages or PDFs. Not a single one is an md file. Obsidian either can’t touch them, or handles them really poorly. The exact moment you most need to pull something up and reference it—that’s precisely when it’s least helpful. Put those two things together, and the conclusion just sits there staring at you. The two things it’s good at, you don’t need it for. The one thing left for it, it can’t do well. On that point alone, Obsidian is honestly completely redundant. So what do I actually need? That brings us right back to what I was just talking about.

[03:09] Since maintenance is already handed off to AI, the presentation layer gets generated along the way too. I don’t want a fixed interface. I want something that manages itself—a multi-format viewer, an editor, plus a filter-and-retrieval tool. I’m not trying to open one specific document. I want to slice this whole pile of stuff by contacts, by company, by project, by life goals. How the connections get built, which pieces to pull—that’s all done by AI, not by me. And I’m not the only one saying this.

[03:36] Karpathy’s LLM Wiki is even more direct. He just throws the knowledge base at ClaudeCode to maintain, and people only read. Maintenance is no longer a human job. Alright, talking about it isn’t enough. So how do I actually manage my files and my knowledge base right now? To put it bluntly—I just use ClaudeCode to manage it. Let me show you what that looks like. The moment the page opens, over two hundred nodes spread themselves out into a Morandi-coloured starfield. The colours map directly to file formats, so at a glance you can tell which is a webpage, which is a note, which is a spreadsheet, which is a slide deck.

[04:03] The bigger the dot, the more information it holds. Tap any one of them, and a floating window pops right up—the file content is right there. You can read it, you can edit it. The panel on the right—that’s the slice I was just talking about. Contacts, organisations, projects, goals—you click on one, and the whole graph instantly filters down to only the documents related to it. They pull themselves back together into a new cluster. And that’s exactly the kind of retrieval tool I wanted. I don’t have to remember what any of those files are called. Not at all. But the most interesting part? Those edges. Those edges aren’t drawn by a human. The AI inferred them on its own.

[04:30] It searches in four levels. First, it checks if there’s an explicit declaration in the document. Then topic membership. Then in-body citation. And only at the very end does it match title similarity. So I say, backlinks in the AI era are written by AI itself. And there’s another point that really makes me comfortable. I create a new file in the terminal, do absolutely no organizing, just refresh the page, and that node just grows out by itself—filing is indexing. So how does this starfield come about? I did a video on this before, specifically about how to let AI organize local files on its own, auto-rename them.

[04:57] That one’s open-source too. Today’s starfield is really the second half of that same project. Rename handles where files go. Starfield handles how files relate to each other. So is this whole thing complicated? No, not at all. Finished pieces go in one folder. Notes sorted by topic go in another. All plain text. There’s one hard rule baked into CLAUDE.md: every time you create a document, tag it and mark its connections right then and there. For all those old legacy documents, you don’t need to go back and fix them one by one. Just drop in a central registry table—one row covers an entire topic. Now at this point you might be thinking, okay—

[05:24] So isn’t that just needing a skill? But I think even a skill is not enough. A skill, at the end of the day, is still passive. You have to call it for it to do anything. What I want is something that moves on its own—an autonomous loop, plus a scheduled job. It wakes up on schedule, scans the whole library on its own, puts new files where they belong, fills in missing connections, and flags the ones that should be retired. And when I want to check, the stuff is already there. I’m not in that process. Now take that logic one step further. Last episode, I just talked about how to connect AI into online documents.

[05:51] So, think about it—why were online documents valuable in the first place? Because an individual simply couldn’t build a platform. You couldn’t just roll your own thing that managed documents, sent them out, and let people edit together. So we had to go onto someone else’s platform. But now with AI, that premise is gone. AI handles the documents, the UI is generated on the fly. If you want to share, just fire off a web page—done. So what does the platform layer even do anymore? I suspect what we’ll actually need in the future might just be some kind of A-to-A protocol, so my AI can talk directly to your AI.

[06:18] So online documents—they’re a transitional form too. So let’s go back to that tool anxiety at the start. You’re anxious about not having the right software installed, but that’s not the real issue. The real problem is, to this day, you’re still the one managing all those files. And that’s something no software can fix. You need to change who manages them. I’ve already handed everything over to ClaudeCode. So that question I asked you at the beginning—it just doesn’t apply to me anymore. I don’t need to go back and dig through old notes. Whatever I should be seeing just shows up in front of me on its own.

[06:45] That’s it for this one. I’ll see you in the next video. Bye.