Build a persistent compounding wiki from raw sources using the LLM Wiki pattern
Installation
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Build and maintain a persistent, compounding knowledge wiki from raw sources.
Based on Karpathy's LLM Wiki pattern: 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 <ACTION> [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]":
Create the directory structure: raw/, wiki/, wiki/entities/, wiki/concepts/, wiki/sources/
Copy $SKILL_PATH/templates/schema.md to CLAUDE.md (or AGENTS.md)
Create wiki/index.md from $SKILL_PATH/templates/index.md
Create wiki/log.md from $SKILL_PATH/templates/log.md
Create a starter wiki/overview.md stating the wiki's domain and purpose
Tell the user to add sources to raw/ and run /llm-wiki ingest <file> next
Done when: all five files and directories exist and the user has been told the next step.
For "ingest ":
Read the source file; share key takeaways and ask the user what to emphasize
Create wiki/sources/<slug>.md with a structured summary
Update wiki/overview.md if the source shifts the big picture
Create or update entity pages in wiki/entities/ for key people, places, organizations
Create or update concept pages in wiki/concepts/ for key ideas and themes
Add a contradiction note on any existing page where this source conflicts with current content
Update wiki/index.md — every new or modified page must appear with a one-line summary
Append to wiki/log.md: ## [YYYY-MM-DD] ingest | <Source Title>
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 ":
Read wiki/index.md to identify the most relevant pages
Read those pages in full
Synthesize an answer with citations to specific wiki pages
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.
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 — local hybrid BM25/vector search with an MCP server:
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.