I’ve been playing with a Codex-assisted LLM Wiki in Obsidian for a week or so.
Today it got its first serious exercise helping me plan for tomorrow’s VerseNotes office hours about Jesus as the firstborn, and it smashed it.
It started from a few source notes I provided that I knew I had on the topic, surfaced a bunch more (including a cluster I wrote on “second sons in Scripture”), extracted Scripture references from all those notes, and compiled it all into a speaking outline with links to the source notes.
Process Notes
There are a thousand “do it for me” blog posts out there, and even an Obsidian plugin.
But as usual, I want to build it myself so I know how it works. So I’ve written the prompts by hand and iterated on them as I’ve used them.
Also, I don’t like Karpathy’s concept of an LLM-only wiki on top of/below my notes. If a note is useful, it should be useful to me and the machine. So I’ve baked that assumption into my prompts, and I’ve worked very slowly, one source at a time, inspecting the results and injecting directly into the vault.
The biggest problem I’ve encountered is that I have had to introduce a lot of machine-readable frontmatter, and that’s annoying visually. Eventually I suspect I’ll have to make an architectural change so the mechanics don’t get in my way as I’m working.
