# LLM Wiki > Build a persistent compounding wiki from raw sources using the LLM Wiki pattern - Skill: `jellydn/llm-wiki` (Agent Skill, multi-file: 4 files) - Install (CLI): `npx skillmds@latest add jellydn/llm-wiki` - Raw SKILL.md: https://api.skillmd.com/api/skills/jellydn/llm-wiki/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: AI & ML - License: MIT - Author: jellydn (https://skillmd.com/u/jellydn) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/jellydn/llm-wiki --- # LLM Wiki Build and maintain a persistent, compounding knowledge wiki from raw sources. Based on [Karpathy's LLM Wiki pattern](https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f): the LLM incrementally integrates each new source into a structured wiki of markdown files — updating entity pages, flagging contradictions, strengthening cross-references. Synthesis is compiled once and kept current, not re-derived on every query. ## Usage `/llm-wiki [ARGUMENTS]` ## Actions - **init [DIRECTORY]** - Initialize a new wiki (default: `wiki/`) - **ingest ** - Process a source file and integrate it into the wiki - **query ** - Answer a question using the wiki as the knowledge base - **lint** - Health-check the wiki for contradictions, orphan pages, and stale content - **status** - Show a summary of the wiki's current state --- ## Architecture ``` your-wiki/ ├── raw/ # Immutable source documents — you add, LLM never modifies │ └── assets/ # Downloaded images referenced by sources ├── wiki/ # LLM-generated and maintained pages │ ├── index.md # Catalog of all pages with one-line summaries │ ├── log.md # Append-only record of all operations │ ├── overview.md │ ├── entities/ # People, places, organizations │ ├── concepts/ # Ideas, topics, themes │ └── sources/ # One summary page per source └── CLAUDE.md # Schema: wiki conventions and per-operation workflows ``` Templates for all three layers are in `$SKILL_PATH/templates/`. --- ## For "init [DIRECTORY]": 1. Create the directory structure: `raw/`, `wiki/`, `wiki/entities/`, `wiki/concepts/`, `wiki/sources/` 2. Copy `$SKILL_PATH/templates/schema.md` to `CLAUDE.md` (or `AGENTS.md`) 3. Create `wiki/index.md` from `$SKILL_PATH/templates/index.md` 4. Create `wiki/log.md` from `$SKILL_PATH/templates/log.md` 5. Create a starter `wiki/overview.md` stating the wiki's domain and purpose 6. Tell the user to add sources to `raw/` and run `/llm-wiki ingest ` next **Done when:** all five files and directories exist and the user has been told the next step. --- ## For "ingest ": 1. Read the source file; share key takeaways and ask the user what to emphasize 2. Create `wiki/sources/.md` with a structured summary 3. Update `wiki/overview.md` if the source shifts the big picture 4. Create or update entity pages in `wiki/entities/` for key people, places, organizations 5. Create or update concept pages in `wiki/concepts/` for key ideas and themes 6. Add a contradiction note on any existing page where this source conflicts with current content 7. Update `wiki/index.md` — every new or modified page must appear with a one-line summary 8. Append to `wiki/log.md`: `## [YYYY-MM-DD] ingest | ` **Done when:** every affected page is updated, every new or changed page appears in `index.md`, and the log entry is appended. A single ingest typically touches 5–15 pages. --- ## For "query ": 1. Read `wiki/index.md` to identify the most relevant pages 2. Read those pages in full 3. Synthesize an answer with citations to specific wiki pages 4. If the answer is a valuable synthesis (comparison, analysis, discovered connection), offer to file it as a new wiki page; if filed, update `wiki/index.md` and append to `wiki/log.md` **Done when:** the question is answered with citations, and any filing decision (page created or skipped) is resolved. --- ## For "lint": Scan the entire wiki and report on each of the following: - **Contradictions** — conflicting claims across pages - **Stale content** — pages superseded by newer ingested material - **Orphan pages** — pages with no inbound links - **Missing pages** — concepts mentioned on multiple pages but lacking their own page - **Broken links** — `[[Page Name]]` links pointing to non-existent pages - **Data gaps** — areas a targeted source or web search could fill Suggest new questions and sources. Ask before making any fixes. Append to `wiki/log.md`: `## [YYYY-MM-DD] lint | health check` **Done when:** all six issue types are checked and reported, suggestions are given, and the log entry is appended. --- ## For "status": Report: source count in `raw/`, wiki page count, last 10 entries in `wiki/log.md`, and any obvious health issues (orphan pages, missing index entries). **Done when:** all four items are reported. --- ## Index and log **wiki/index.md** — content catalog, updated on every ingest. Every page with a link and one-line summary, organized by category. The LLM reads this first when answering queries. **wiki/log.md** — append-only operation log. Entry format: `## [YYYY-MM-DD] | `. Grep-parseable: ```bash grep "^## \[" wiki/log.md | tail -5 # last 5 operations grep "^## \[.*\] ingest" wiki/log.md # all ingests grep "^## \[.*\] query" wiki/log.md # all queries ``` --- ## Optional: qmd search For larger wikis, add [qmd](https://github.com/tobi/qmd) — local hybrid BM25/vector search with an MCP server: ```bash bun install -g @tobilu/qmd qmd collection add ./wiki --name my-wiki qmd embed qmd query "how does X relate to Y" -c my-wiki ``` --- ## Tips - **Obsidian Web Clipper** converts web articles to markdown — fast source collection for `raw/`. - **File good answers**: valuable query results (comparisons, analyses, new connections) should become wiki pages — explorations compound the knowledge base just like ingested sources do. - **Ingest one source at a time** — read the summaries, check the updates, guide emphasis. You get more value staying in the loop than batch-ingesting unattended. - The wiki is a git repo of markdown files — version history and collaboration come free.