llm-wikilisted

# LLM Wiki Skill Pattern by Andrej Karpathy. Build a personal knowledge base where the LLM maintains an interlinked markdown wiki between the user and the raw sources. Knowledge is **compiled once and kept current**, not re-derived per query. > **Compile, don't retrieve.** Raw sources are the source code. The wiki is the binary. The LLM is the compiler. Obsidian is the IDE. ## Core idea Unlike RAG (rediscovers per query, no accumulation), LLM Wiki: - Reads new source → integrates into existing wiki - Updates entity/concept pages, revises summaries, flags contradictions - Cross-references already materialised in the file system - Synthesis reflects everything read so far - User curates sources + asks good questions. LLM does the bookkeeping ## Three layers 1. **`raw/`** — immutable source documents (PDF, MD, articles, images, data). **Read-only for the LLM. Never modify. Never rewrite. Never replace originals with summaries.** 2. **`wiki/`** — LLM-owned markdown. Entity, concept, source, and synthesis pages. LLM creates/updates. 3. **`CLAUDE.md`** (or `AGENTS.md`) — schema. Conventions, workflows, ingest/query/lint rules. Co-evolves with user. ## Critical failure mode — read this before ingesting Batch ingesting many sources at once saturates the context window. Under pressure, the LLM will silently compress raw files into summaries and store **those** in `raw/` instead of the originals. The wiki then becomes summaries-of-summaries: orphan pages, broken backlinks, hall