I’ve been testing an agent-only @obsdmd vault in parallel with my...

The @karpathy LLM-wiki does what you'd expect — digests sources, chunks them, synthesizes answers.
But the structure mapped almost one-to-one to the Zettelkasten I've been teaching for years. It just does the grunt work automatically.
Maybe more people can get the benefits without the years of system-noodling it usually takes?
There's just one thing missing though...
The system captures everything (synthesis, connections, patterns). But my thinking ends up mixed in with the agent's output.
There's no place where I can see just my ideas, review them, and build on them.
It's all in there somewhere.
Remember that time when you argued with a dead author and won?
Good luck finding it.
Thankfully, I didn't have to create anything new. I added what I call the CONVOs layer — a place where only my thinking lives.
I reply to authors. They reply to me. Not a wiki — an async conversation applied to problems I want to solve.
The agent isn't allowed to write here without checking with me first.
Which authors do you want to have a group chat with?
All the ammo’s there. When an idea hits, examples, stories, and arguments are already in the conversation.
It’s not just a wiki. It’s a writing tool.
The CONVO is the playground. But some ideas stop being experiments.
When an idea becomes something I live by — my niche, my mission — it graduates to Ops.
Ops is the rules the LLM lives by. Whenever it needs context, it knows where to look. I never re-explain myself.
And we continuously evolve together.
Now...
it's not all obsidian butterflies and claud rainbows. There's still a lot of evolving that can happen :)
- Not loving how it uses tags
- It actually doesn't use its own index
- How its search doesn't really scale (echoed by Karpathy himself)
If you're on the fence, just create a new vault and put Karpathy's idea file in. You always learn more when you just give it a try.
Let me know if you have any questions about starting.




