LLM Wiki Skill (proposal)

Purpose: Keep a compounding markdown wiki of external knowledge (articles, papers, transcripts, clipped web content) maintained by an LLM so the bookkeeping is free and the knowledge accumulates. Based on Andrej Karpathy's "LLM Wiki" pattern published April 2026. See references/architecture.md for the full design rationale and how it interoperates with existing AgentOps skills.

Status: PROPOSAL — not yet merged. Opened 2026-04-11 for council review. See .agents/research/2026-04-11-karpathy-llm-wiki-integration.md for the research that motivated this skill.

The core idea in one paragraph

RAG has a critical flaw: there is no accumulation. Each query re-discovers fragments from the raw source corpus. The LLM Wiki pattern replaces per-query RAG with a persistent LLM-maintained markdown wiki that sits between raw sources and the user. The LLM reads each raw doc once, extracts concepts, writes summaries, cross-links related ideas, and keeps the whole thing current. On query time, the LLM reads the already-compiled wiki — much faster, much richer, and it compounds.

Karpathy's metaphor: "Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase." Same as AgentOps's internal-flywheel pattern, applied to external knowledge.

What this skill IS

A specification for four operations against a wiki-structured vault:

1. ingest — process new raw source → write wiki pages + update index + log
2. query — answer a question from wiki pages → file synthesis results back as new pages
3. lint — periodic health-check for contradictions, stale claims, orphans, missing concepts
4. promote — move a mature wiki page from draft to reviewed, or from wiki to authored content