Obsidian Starter Kit - System - LLM Wiki System

Canonical version: Obsidian Starter Kit - System - LLM Wiki System.

🏠 User Guide / Systems / LLM Wiki System

The LLM Wiki System is the OSK's framework for building persistent, AI-maintained knowledge bases. Inspired by Andrej Karpathy's LLM Knowledge Base pattern (April 2026), it lets you collect raw sources on any topic and have an LLM compile, cross-reference, and maintain a structured wiki. The human curates sources and asks questions; the LLM handles the bookkeeping.

Part of the AI Assistant System.

TL;DR

• Three-layer architecture: Raw Sources (immutable) -> Wiki (LLM-maintained) -> Schema (shared conventions)
• Core operations: create, ingest, explore, query, lint, list, deepen, graduate, absorb; plus activity reports, canvases, HTML visualizations and static-site export
• Note type: type/ai_wiki with wiki-specific tags ai_wiki/<wiki-slug>
• Naming convention: AI Wiki - <Wiki Name> - <Article>.md prevents clashes
• Each article tracks: wiki_name, wiki_role, explored status, confidence level, maturity, sources, graduated_notes
• Index.md catalogs all articles; Log.md records all operations chronologically
• Speculative links (red links) allowed; they signal future exploration targets
• Dedicated agent (OSK Wiki Curator) and panels (osk-panel-wiki-review, osk-panel-wiki-graduate, osk-panel-wiki-cross)

🎯 Philosophy

The bottleneck in traditional knowledge management is human attention: cross-references rot, indexes go stale, contradictions go unnoticed.

The LLM Wiki approach moves the mechanical bookkeeping (writing summaries, maintaining cross-references, updating indexes, flagging contradictions) to the LLM. The human focuses on source curation and strategic questioning.

Key insight from Karpathy: "The tedious part of maintaining a knowledge base is not the reading or the thinking. It's the bookkeeping." LLMs excel at exactly this.

This creates:

• Compounding knowledge. Every ingestion and query enriches the wiki
• Source provenance. Every claim traces back to a source with confidence levels
• Living structure. Cross-references stay current automatically
• Low maintenance. The LLM does the tedious work; you do the thinking

🏗️ Architecture

Three Layers

Directory Structure

10 Meta/99 AI Assistant/Wikis/
  <Wiki Name>/
    AI Wiki - <Wiki Name> - Index.md       # Catalog of all articles
    AI Wiki - <Wiki Name> - Log.md         # Chronological operation record
    AI Wiki - <Wiki Name> - <Article>.md   # Wiki articles
    raw/                                    # Source documents
  AGENTS.md                                 # Wiki instructions for any AI assistant
  CLAUDE.md                                 # Points Claude Code at AGENTS.md

Wiki instructions

The Wikis/ folder carries its own AGENTS.md (and a CLAUDE.md that points to it). Any assistant that works inside a wiki reads it first: load osk-wiki-shared before any operation, the wiki skills available, the rules (articles start with explored: false and only you mark them explored; maturity is computed, never set by hand; raw/ is never modified; the Log is append-only), the OSK Wiki Curator agent and the three wiki panels. So the conventions hold whichever assistant opens the wiki, without loading the whole AI Assistant System first.

Note Type: AI Wiki Articles

Tag: type/ai_wiki Wiki-specific tag: ai_wiki/<wiki-slug> (e.g., ai_wiki/machine-learning) Template: TPL AI Wiki Article.md Folder: 10 Meta/99 AI Assistant/Wikis/<Wiki Name>/

That folder is excluded from Templater's folder template (agents write there, and Templater would run template code in their new notes). A note you create there by hand gets no template: create articles through the osk-wiki-* skills, or apply TPL AI Wiki Article with Alt + E.

Properties:

Special Files

Index.md: The entry point for both humans and LLMs. Contains:

• Wiki description and scope
• Statistics (article count, explored ratio)
• Categorized article links with one-line summaries
• Unexplored topics list

The Index lists every article of the wiki, published or not: its lists don't filter on public_note.

Log.md: Append-only chronological record of all operations:

## [YYYY-MM-DD HH:MM] ingest | Article Title
Ingested: URL/file/note reference
- Files created: list
- Files updated: list

📝 Core Operations

Create (osk-wiki-create)

Scaffold a new wiki: directory, Index, Log, optional seed articles. Can seed from a topic description or existing vault notes.

Ingest (osk-wiki-ingest)

Process a source (URL, file, vault note) into the wiki:

1. Fetch/read the source
2. Create source summary article
3. Update existing concept articles with new info
4. Create stub articles for new concepts
5. Update Index and Log

A single source usually touches many wiki articles: its own summary, the concepts it updates, and the stubs it creates.

Explore (osk-wiki-explore)

Deeper exploration of topics. Three depth levels:

• Shallow: Stubs with key points
• Medium: Substantial articles (default)
• Deep: Full coverage with web research

When several independent topics need exploring, it runs them in parallel agents for speed.

Query (osk-wiki-query)

Ask complex questions against the wiki:

1. Read Index to find relevant articles
2. Search and gather from multiple articles
3. Synthesize answer with wikilink citations
4. Optionally file the answer back as a new article (compounding!)

Lint (osk-wiki-lint)

Health checks:

• Frontmatter completeness
• Index consistency
• Orphan detection
• Red link analysis
• Contradiction detection
• Confidence assessment
• Exploration gaps
• Cross-link density
• Naming convention compliance
• Maturity assessment (the maturity value matches the article's word count, sources, cross-links and confidence)

List (osk-wiki-list)

Show all wikis with stats: article count, word count, explored ratio, confidence distribution, last activity.

Deepen (osk-wiki-deepen)

Depth-first enhancement of existing articles: moves them up the maturity ladder (stub, draft, substantial, mature).

Graduate (osk-wiki-graduate)

Decomposes a mature article into atomic permanent note candidates. Only articles with maturity: mature qualify.

Absorb (osk-wiki-absorb)

Bridges wiki articles and permanent notes in both directions through "Vault Perspective" sections instead of copying text across.

Activity, canvas, visualization and export

• osk-wiki-activity: what changed in a wiki over a time window (articles created and updated, maturity changes, exploration progress)
• osk-wiki-canvas: generates or updates an Obsidian Canvas for a wiki
• osk-wiki-visualize: interactive HTML graph of the articles, color-coded by maturity
• osk-wiki-to-html: turns a wiki into a static website

🤖 Agent and Panels

OSK Wiki Curator agent: Light agent that owns wiki quality, maturity assessment, graduation, and vault bridging. Accumulates knowledge about what works. Always loads osk-wiki-shared.

osk-panel-wiki-review panel: reviews wiki articles from multiple angles:

• Wiki Curator (structure) + Researcher (accuracy) + Skeptic (claims) + Beginner (accessibility) + Editor (prose)

osk-panel-wiki-graduate panel: quality-gates graduation candidates:

• Wiki Curator (decomposition quality) + Skeptic (claim strength) + Editor (standalone clarity) + Power User (depth/value)

osk-panel-wiki-cross panel: analyzes two or more wikis together to find connections, overlaps and contradictions:

• Wiki Curator + Researcher + Skeptic + Strategist

🔄 Workflow Examples

Knowledge Maturation Pipeline

1. Deepen: osk-wiki-deepen to take draft articles to substantial/mature
2. Lint: osk-wiki-lint checks that each article's maturity matches its content
3. Graduate: osk-wiki-graduate proposes atomic notes from mature articles
4. Panel: osk-panel-wiki-graduate reviews candidates before creation
5. Absorb: osk-wiki-absorb bridges vault notes back into wiki articles
6. Repeat: Each cycle moves knowledge from AI-maintained wiki to human-owned Zettelkasten

Building a Research Wiki

1. Create: osk-wiki-create with name "Quantum Computing" and seed topics
2. Ingest: Drop papers and articles into raw/, run osk-wiki-ingest for each
3. Explore: osk-wiki-explore to fill gaps and expand stubs
4. Query: Ask questions, file valuable answers back as articles
5. Lint: Periodic health checks to maintain quality
6. Repeat: Each cycle compounds knowledge

Turning a Wiki into Content

Ask for the wiki-to-content team (osk-agent-team), e.g. "Turn my wiki on X into an article":

1. The OSK Wiki Curator answers from the wiki with osk-wiki-query, citing articles and their confidence
2. The writing step drafts the piece with osk-writing-note-writer (in your voice when your voice profile is filled)
3. The OSK Agent Editor reviews the draft

Personal Knowledge Wiki

1. Create wiki from existing vault notes on a topic
2. The LLM extracts concepts, creates cross-referenced articles
3. New learning gets ingested; wiki grows organically
4. Becomes a queryable personal knowledge base

🔧 Integration with Other Systems

• AI Assistant System: Wikis are a subsystem; the Wiki Curator agent lives in the agent framework
• Learning System: Wiki articles complement literature notes; ingest from Readwise highlights
• Zettelkasten System: Wiki articles are NOT permanent notes, but mature articles can be graduated into atomic permanent notes via osk-wiki-graduate. The graduated_notes property tracks provenance
• Maintenance System: Wiki lint integrates with vault health checks

🎨 Best Practices

1. One wiki per topic domain. Don't mix unrelated subjects
2. Ingest high-quality sources. Garbage in, garbage out; curate your raw/ folder
3. Review confidence levels. Articles with confidence: low need more sources
4. Run lint monthly. Catch contradictions and orphans before they accumulate
5. File query answers back. This is what makes the wiki compound
6. Use the Index as your entry point. Don't navigate the file system randomly
7. Let red links guide exploration. They show you where the wiki wants to grow
8. Don't manually edit wiki articles. Let the LLM maintain them; add your insights as sources
9. Graduate mature articles. When an article reaches the top of the maturity ladder, run osk-wiki-graduate; atomic permanent notes are the payoff, the wiki is the workshop
10. Bridge, don't duplicate. Use osk-wiki-absorb to connect permanent notes and wiki articles through Vault Perspective sections instead of copying text across

🚧 Common Challenges

• Wiki too broad: Keep scope focused; one wiki per research domain
• Stale content: Run lint to detect; re-ingest updated sources
• Low confidence everywhere: Need more diverse sources; prioritize high-quality ingestion
• Too many red links: Normal for young wikis; use explore to fill the most-referenced ones first
• Context window limits: Large wikis (400K+ words) may need vector search; the Index helps navigate

🚀 Next Steps

• Start small: Create a wiki on a topic you're actively researching
• Seed from vault: Use existing permanent and literature notes as starting material
• Ingest a handful of sources: Get enough material for meaningful cross-references
• Ask a few questions: Test the query workflow; file valuable answers back
• Run lint once: See how the health check works and what it catches
• Expand gradually: Let the wiki grow organically through research and ingestion

🔗 Related Documentation

OSK Systems:

• Obsidian Starter Kit - System - AI Assistant System
• Obsidian Starter Kit - System - Learning System
• Obsidian Starter Kit - System - Zettelkasten System

OSK Reference:

• AI Wiki Articles note type
• AI Wiki Article template

Concepts and Architecture:

• AI Assistant Architecture
• LLM Wiki
• Andrej Karpathy
• Farzapedia

Tutorials:

• Obsidian Starter Kit - Tutorial - Creating Your First LLM Wiki
• Obsidian Starter Kit - Tutorial - Using LLM Wikis

About Sébastien

Ready to get to the next level?