wiki-knowledge-compiler
A local-first LLM Wiki / Knowledge Compiler plugin for Claude Code.
Inspired by Karpathy's "LLM Wiki" concept: raw sources are immutable, the wiki is derived knowledge, and useful outputs are always written back.
What it does
wiki-knowledge-compiler turns Claude Code into an AI-powered knowledge management system that:
- Ingests raw sources (PDFs, markdown, notes, URL stubs) into an immutable archive
- Extracts atomic concepts, creates summaries, and builds topic pages — all with provenance
- Queries the knowledge base with citations and writes back reusable answers
- Maintains structural health: detects orphans, duplicates, stale pages, and broken links
- Synthesizes cross-cutting insights with explicit confidence ratings and uncertainty markers
Everything is plain markdown with YAML frontmatter. No databases, no cloud services, no proprietary formats.
Install
Option 1: Claude Code Marketplace (recommended)
The easiest way to install. Open Claude Code and run:
/plugin marketplace add bradduy/wiki-knowledge-compiler
Or use the Add Marketplace dialog in Claude Code settings:
- Open Claude Code
- Go to Settings → Plugins → Add Marketplace
- Enter:
bradduy/wiki-knowledge-compiler
- Browse the marketplace and install wiki-knowledge-compiler
Then install the plugin:
/plugin install wiki-knowledge-compiler@wiki-knowledge-compiler
Once installed, run /setup-wiki to configure your project size and search backend.
Option 2: Clone as a standalone knowledge base
git clone https://github.com/bradduy/wiki-knowledge-compiler.git
cd wiki-knowledge-compiler
Open the folder in Claude Code — it will automatically detect CLAUDE.md and .claude/ configuration. All commands are immediately available.
Option 3: Add to an existing project
Copy the plugin files into your repo:
git clone https://github.com/bradduy/wiki-knowledge-compiler.git /tmp/wkc
# Copy plugin structure
cp -r /tmp/wkc/.claude/ your-project/.claude/
cp -r /tmp/wkc/.claude-plugin/ your-project/.claude-plugin/
cp -r /tmp/wkc/{agents,skills,templates} your-project/
cp /tmp/wkc/CLAUDE.md your-project/CLAUDE.md
cp /tmp/wkc/wiki.config.md your-project/wiki.config.md
# Initialize the knowledge base
bash /tmp/wkc/scripts/init-kb.sh your-project/knowledge-base
# Clean up
rm -rf /tmp/wkc
Verify installation
After any install method, open Claude Code in the project directory and run:
/setup-wiki
This walks you through project size configuration and confirms everything is working.
Commands
| Command | Description |
|---|
/setup-wiki | Start here. Interactive setup — choose project size, configure search backend |
/ingest-source <path-or-url> | Add a raw source and extract knowledge from it |
/query-wiki <question> | Search the wiki, answer with citations, write back if useful |
/update-index | Rebuild all knowledge base indexes from the file system |
/health-check | Audit for orphans, duplicates, stale pages, weak links |
/generate-insights [focus] | Synthesize new insights from existing knowledge |
Knowledge Base Structure
knowledge-base/
raw/ # Immutable source documents
concepts/ # Atomic concept pages (one per file)
topics/ # Broader topic overviews
summaries/ # Per-source summaries with provenance
insights/ # Cross-cutting observations
index/ # Auto-generated indexes
drafts/ # Work-in-progress pages
references/ # External URL/resource stubs
log.md # Append-only activity log
Usage Examples
Ingest a paper
/ingest-source ~/papers/vaswani-2017-attention-is-all-you-need.pdf
Creates: raw copy, summary, concept pages (attention, self-attention, multi-head attention, etc.), topic updates, index updates, log entry.
Ask a question
/query-wiki How does self-attention differ from cross-attention?
Searches the wiki, answers with citations to specific pages, identifies knowledge gaps, and writes back the synthesis as an insight if it's novel.
Check wiki health
/health-check
Reports: 2 orphan pages, 1 suspected duplicate (attention.md vs attention-mechanism.md), 3 summaries with broken source links. Offers to auto-fix safe issues.
Generate insights
/generate-insights scaling laws
Reads across all sources related to scaling, identifies cross-cutting patterns, generates insight pages with confidence ratings.
Architecture
Agents
bradduy