I Used AI to Fix My Broken Obsidian Vault

On failing to replicate the most prolific academic in history

Photo by Christa Dodoo on Unsplash

Almost as soon as I went digital with my notes, I started changing my mind about how to organise them, creating a mess it would have taken days to untangle… so I didn’t bother.

I was inspired by reading How to Take Smart Notes and discovering how the German sociologist Niklas Luhmann obsessively maintained 90,000 index cards, enabling him to publish 70 books and nearly 400 articles. I wanted that system, but with modern tools. Instead, I got distracted by Tiago Forte’s PARA structure, then the inevitable YouTube rabbit hole.

The result was predictable.

Whilst polishing my process, I ended up with over 80 notes stuck in the inbox, duplicates I’d forgotten about, links pointing to notes I’d never created, all tagged inconsistently. By the time I knew what I wanted (Maps of Content, a proper tagging taxonomy, notes at different maturity stages), fixing 200 notes felt like an enormous chore.

So last week I got Claude to do it.

The Audit

A friend came up with the term “thought refactoring” to describe how I was using AI to rework documents. Co-work, Claude’s desktop tool, helped me expand that idea from a single document to an entire folder of notes. What made it possible was Obsidian storing everything as plain Markdown files rather than locking it in some proprietary cloud format.

I made a backup, pointed Co-work at my Obsidian root folder, and used this prompt:

I’d like you to review my Obsidian vault.

I’ve tried to mix a Zettelkasten-style approach with atomic notes. 
That is one idea per note, connected through links, not nested folders. 
Maps of Content as index notes instead of a folder hierarchy.

I want a tagging system with two dimensions: 
- a type tag (concept, framework, principle, quote, book-note, or person) 
- a maturity tag (seedling, developing, or evergreen)

Can you:
- audit the vault structure
- identify what’s working and what isn’t
- put together an action plan

Don’t move anything until I’ve approved the plan.

It grasped the intent quickly. Not just Zettelkasten, but the added complexity of MOCs and a two-dimensional tagging system on top.

Minutes later, Claude described the extent of my personal knowledge management shame. On top of the notes stuck in the inbox, it explained that 114 were unfinished and around 30 were essentially empty. Oh, and I’d misspelt filenames, had orphan notes with no connections, and created duplicates months apart without realising. And yes, my Resources folder conflicted with Zettelkasten principles.

But it had a six-phase plan to fix it, which I checked before it started moving anything. Because I still have some use, right?

Then it executed. Every note moved, renamed, tagged, and wired into the right MOC. It extracted atomic concepts from my notes, so now autopsia, disequilibria as the motive for learning, and shamanic rituals all have their own home.

Claude had flagged my plugins, too. I got recommendations for Dataview and Homepage. All without a single YouTuber telling me not to focus on the process, whilst spending twenty minutes describing their process.

The anxiety of adding to broken notes and creating more problems for future Mike to fix lifted, and I started to think about what else was possible.

The Skill

The cleanup fixed my past mistakes. I also needed help maintaining my notes so they didn’t drift back. So I built a custom knowledge-linker skill, a set of instructions that tells Claude exactly how my vault works, what conventions to follow, and how to create notes that fit.

You can see the skill here → knowledge-linker.skill

Now, when I come across a new idea, I can say “add Jevons paradox to my vault”. The result is a properly formatted, tagged note linked to related notes and filed into the right Map of Content.

I tried it after finishing a book last week. Instead of writing half-finished notes and telling myself I’d process them later, I talked through what I’d learned with Claude. In under 20 minutes, we extracted 4 new ideas into atomic notes and filed them based on what interested me.

Talking it through also helped with recall, though it did make me think about what I’m losing by not making the connections myself. Curating notes manually had forced me to think deliberately about the connections. So I’ve kept some agency over that, and I review everything after.

I also built a second capability into the skill: it reads a draft and searches the vault for any connections that might be useful. My hope is that using them in writing is what makes them stick, rather than just curating them.

The skill is there to handle the admin, so I can focus on applying the ideas.

In Defence of Not Doing It Yourself

The emergent connections I’d imagined, the Luhmann dream of ideas linking across years of reading, never appeared. I couldn’t replicate his obsession even with a tool like Obsidian. My result was a pile of unfinished notes and a growing dread about a weekend I’d eventually spend sorting it out, or quietly give up on.

But that weekend never came. Claude sorted it in an afternoon, knew exactly what I was trying to build, and gave me the tools to keep it that way.

Luhmann’s obsession was the connections, each card placed deliberately, the system maintained by hand for forty years. I’ve automated that part. What’s left is the bit he’d have actually envied: the time to read more, think more, and find ways to apply it.