The LLM-wiki pattern: the idea this notebook runs on

Andrej Karpathy's pattern for an LLM-maintained personal wiki — kept verbatim at the repo root as LLM_Wiki.md — is the founding idea mantua.io itself instantiates. What the pattern claims, and where this repo's implementation matches or departs from it.

Karpathy's argument starts from a complaint about RAG: upload a collection of documents, retrieve relevant chunks at query time, generate an answer — and the LLM re-derives the same synthesis from scratch on every question. Nothing accumulates. His fix is to have the LLM incrementally build and maintain a persistent wiki that sits between the owner and the raw sources: read a new source once, extract what matters, fold it into the existing pages — updating summaries, flagging contradictions, adding cross-references — so the wiki compounds instead of being re-computed every time.

The pattern has three layers: raw sources (immutable, the source of truth), the wiki itself (markdown, entirely agent-owned), and a schema document — a CLAUDE.md or AGENTS.md — that tells the LLM the conventions well enough to act as a disciplined wiki maintainer rather than a generic chatbot. Three operations run on top: ingest (file a new source, however many pages it touches), query (answer a question, and file good answers back as pages so explorations compound too), and lint (a slower-cadence health check for contradictions, stale claims, and orphan pages).

This isn't an adjacent idea for mantua.io — it's the blueprint. The essay is kept verbatim at the repo root (LLM_Wiki.md) rather than paraphrased away, and the mapping onto this repo is close to literal:

A few places this repo's implementation goes further than the essay describes, or departs from it:

• Schema enforcement is mechanical, not just documented. The essay's schema layer is prose the LLM reads and follows; here it's also a Zod schema (velite.config.ts) that fails the build on a violation. Convention became a compiler check.
• Provenance is a first-class field. The essay doesn't dwell on who wrote what — it assumes the LLM writes the wiki and the owner reads it. mantua.io adds origin: human | agent | mixed to every page, precisely so the owner's own writing (like Agentic engineering: finding my role in the loop) stays legible as his rather than folding into agent prose. See How this notebook works for why that rule exists.
• No search tool yet. The essay's optional layer — a local search engine such as qmd once an index file stops being enough — is on this repo's roadmap, not yet built.
• "Lint" means two different things in this notebook. The essay's lint operation — contradictions, stale claims, orphan pages — maps directly onto this repo's /lint. Linting as an agent guardrail, from the harness-design cluster, is a different concept that happens to share the name: code linters as guardrails on what an agent writes, not wiki health. Worth keeping straight — they don't contradict each other, they're just namesakes.

Where the essay is most directly relevant to the rest of this notebook: its framing of the owner's job — curate sources, direct the analysis, ask good questions, while the LLM does the grunt work of summarizing, cross-referencing and filing — is the same shift the owner describes wrestling with in the harness-design cluster, just one level up: there it's managing a team of coding agents instead of a wiki-maintaining one. Same bottleneck, same fix.

The same bet, arrived at from teaching

Karpathy makes a second version of the argument that isn't in the essay, and it is a stronger statement of what the pattern is for. Talking about documentation (NoPriors, March 2026): stop writing HTML docs for humans and write markdown for agents, because an agent will re-explain any part of it to any reader at their level with infinite patience. He abandoned a walkthrough video for microGPT on exactly that reasoning — the 200 lines are simple enough that an agent can already explain them better than a fixed explanation can, in as many different ways as the reader needs.

Which reframes what a wiki like this one is. Not a document you write for future-you to read, but the corpus an agent reads in order to answer future-you — which is precisely why /oracle files good answers back as pages, and why the schema lives in markdown the agent loads rather than in a style guide a person is supposed to remember. His formulation of what a person still contributes fits the provenance rule above: the few bits the agent cannot come up with on its own. An agent understands microGPT completely and cannot invent it. Everything downstream of that — the explaining, the cross-referencing, the filing — is not the human's job any more.

He makes the same argument earlier in the interview about smart-home apps, which is what makes it a position rather than an aside. Both halves, and the places the position thins out, are in The customer is not the human anymore — including the dependency this version doesn't name: "write for the agent, it will explain it to people" holds only while the agent's explanation is faithful, and a wrong explanation reads exactly like a right one to whoever came to the docs because they didn't know.