LLM wiki ops
Karpathy's llm-wiki concept. A shift in what's become available to me as an individual; a valuable source of compounding knowledge, with an agent doing the bookkeeping that I abandoned.
My personal knowledgebase of thousands of notes was dying. Not because the thinking was hard, but because the housekeeping was. The cross-linking, the deduplication. Karpathy's LLM-Wiki gist (subject of my last post) inspired me to close that gap and get rid of the tedium. Two weeks into my opinionated implementation, all I now care about is what to read and write; everything downstream like where it goes, what it links to, and how it slots into existing knowledge is handled for me.
What I kept faithful to Karpathy: the constraint. One page per concept. Every page densely cross-linked. The wiki gets richer because the schema is narrow.
Where I diverged: Karpathy's design is wiki-centric and lossy, everything compounds in. So only for content where compounding is wrong, I created a References/ folder: long-form technical playbooks, certification prep etc. References/ keeps it verbatim; concepts can still be mined into the wiki on demand with my /vault skill.
The other extension: going beyond user-driven ingest. A nightly GitHub Action scans frontier researchers for novel content. Just this week, Simon Willison coined "vegan model" for LLMs trained exclusively on licensed/out-of-copyright data, and Dwarkesh Patel articulated the "RL Verification Gap", about the structural mismatch between RL's rapid feedback loops and science's decade-long validation cycles. Scans don't just add pages; they thicken the connective tissue between existing ones in my wiki. And I hooked up Hermes-Agent to deliver a daily luminary summary to telegram for coffee reading every morning.
It's been pretty cool to use daily. Every Claude session shows up already aware of the vault, because the global CLAUDE.md tells it where to look. Always-on for context, on-demand for workflow. A /vault skill handles explicit operations. Those wonderful snippets of knowledge you arrive at through conversation with LLMs, that quietly die as the session ends and you forget? /vault capture drops them into the right inbox. /vault extract mines long-form references for wiki-worthy concepts. And /vault reference places verbatim docs into the right References/ subfolder.
The full replication kit: workflows, the /vault skill, the schema, verify-wiki.py, the day-zero prompts, the claude.md, a README walking through every step, is linked in the comments. It's no plugin, and not for the faint-hearted, designed to tackle a big, brownfield archive, but there you go.
Curious to know if others are using llm-wiki plugins or crafted their own flavour.