I've been "ingesting" sources by reorganizing into a note with information types. I actually don't find Karparthy's idea as useful, because I want to be the one to shape the knowledge ... just use AI to do the grunt work.
Agree. Shaping the knowledge is the part worth keeping. In my setup, the AI suggests each change and waits for my yes before it edits a note. It does the sorting and linking. I decide what the note says. What types do you sort your notes into?
I've been using information types. The basics are: reference, process, task, concept. Then I've added a few like Claim, counter-claim, etc. How about you?
Mine split by topic into 5 main types: people, projects, ideas, decisions and tasks, plus a daily note. Raw sources stay untouched in their own folder.
Your claim and counter-claim are closest to how facts work in mine. Every fact carries a date and a source, and when 2 sources disagree, it opens a Conflict note for me to settle.
Great article and very interesting skill. I implemented the baseline version of Karpathy’s L zlM Wiki. I’ll give your skills a try.
I’m still experimenting with kb topics. I’d like to use the wiki to cover the internal kb of my business (mktg, sales, support, operations) and use it as a single source of truth to run the business (my context).
Is it a viable approach in your opinion? Today I use a bunch of “satellite md files” within the project. I suppose I may need a custom template for this.
I like the reliability of the daily process you defined.
Yeah, viable. That's exactly the use case. A lot of people running the skill use it this way now. Founders and CEOs run it as their company brain (marketing, sales, support, ops in one graph), and even a few researchers use it for lit notes. The feedback is consistent: the win isn't the folders, it's that everything cross-links through shared entities, and the vault maintains itself instead of rotting.
So the shift from satellite MD files is less about a template and more about entities. Give every client, tool, and project its own page, then let your sales/support/ops notes link to those. One client lives on one page but shows up across every function. That's what makes it a real context for an agent, not just organized files.
And you don't have to build the _CLAUDE.md by hand. It's in the repo, and /obsidian-init generates one for your vault. It tells Claude how you write, where things go, and how to reconcile a note that contradicts an older one. Start with ~5 notes, skip the big folder taxonomy, let it grow from real work.
One more thing, since you mentioned running a whole business off it: I'm building a team version right now. Same idea as the personal vault, but the graph is shared across the team instead of just you. v0 is tested, v1 is close. I'll share the output soon. It sounds like exactly your use case and many other cases - I'm solving this pain now!
I've been "ingesting" sources by reorganizing into a note with information types. I actually don't find Karparthy's idea as useful, because I want to be the one to shape the knowledge ... just use AI to do the grunt work.
Agree. Shaping the knowledge is the part worth keeping. In my setup, the AI suggests each change and waits for my yes before it edits a note. It does the sorting and linking. I decide what the note says. What types do you sort your notes into?
I've been using information types. The basics are: reference, process, task, concept. Then I've added a few like Claim, counter-claim, etc. How about you?
Mine split by topic into 5 main types: people, projects, ideas, decisions and tasks, plus a daily note. Raw sources stay untouched in their own folder.
Your claim and counter-claim are closest to how facts work in mine. Every fact carries a date and a source, and when 2 sources disagree, it opens a Conflict note for me to settle.
Great article and very interesting skill. I implemented the baseline version of Karpathy’s L zlM Wiki. I’ll give your skills a try.
I’m still experimenting with kb topics. I’d like to use the wiki to cover the internal kb of my business (mktg, sales, support, operations) and use it as a single source of truth to run the business (my context).
Is it a viable approach in your opinion? Today I use a bunch of “satellite md files” within the project. I suppose I may need a custom template for this.
I like the reliability of the daily process you defined.
Appreciate your input 🙏
Yeah, viable. That's exactly the use case. A lot of people running the skill use it this way now. Founders and CEOs run it as their company brain (marketing, sales, support, ops in one graph), and even a few researchers use it for lit notes. The feedback is consistent: the win isn't the folders, it's that everything cross-links through shared entities, and the vault maintains itself instead of rotting.
So the shift from satellite MD files is less about a template and more about entities. Give every client, tool, and project its own page, then let your sales/support/ops notes link to those. One client lives on one page but shows up across every function. That's what makes it a real context for an agent, not just organized files.
And you don't have to build the _CLAUDE.md by hand. It's in the repo, and /obsidian-init generates one for your vault. It tells Claude how you write, where things go, and how to reconcile a note that contradicts an older one. Start with ~5 notes, skip the big folder taxonomy, let it grow from real work.
One more thing, since you mentioned running a whole business off it: I'm building a team version right now. Same idea as the personal vault, but the graph is shared across the team instead of just you. v0 is tested, v1 is close. I'll share the output soon. It sounds like exactly your use case and many other cases - I'm solving this pain now!
Thanks
Really valuable insight here, originality comes from depth of experience and perspective.
Thanks, Aniket! Yes, I have received many positive feedback as well on this.