llm-wiki

Git-based markdown wiki for software teams. No server, no database. Pure git + markdown + Python scripts. Bring your own AI tool.

Inspired by Andrej Karpathy's vision of LLMs as a new kind of operating system — where the file system IS the knowledge base and AI tools are first-class citizens.

Install

npx skills add junbjnnn/llm-wiki

Works with Claude Code, Cursor, Codex, Gemini CLI, and 20+ other agents.

Why

• Tool-agnostic: Works with Claude, Cursor, Copilot, Gemini, Codex — any tool that reads markdown
• AGENTS.md standard: Universal schema readable by any AI tool
• Compounding knowledge: Every query can feed insights back into the wiki
• Zero infrastructure: git repo = your wiki. No hosting, no server, no API keys

Architecture

Data Flow

                    ┌─────────────────────────────────────────┐
                    │            your project repo            │
                    └─────────────────────────────────────────┘
                                       │
  PDF, HTML, TXT, MD ──► [ingest.py] ──┤
                                       ▼
                                   sources/          (raw parsed documents)
                                       │
                              [AI compile] ──► wiki/  (generated pages with wikilinks)
                                                 │
                              [AI query] ◄───────┘
                                   │
                                   ▼
                                Answer
                                   │
                                   └──► new insight? ──► wiki/  (feedback loop)

Directory Structure

project-repo/
├── .wiki/                     # Wiki root (default: subfolder)
│   ├── AGENTS.md              # Universal schema (any AI tool)
│   ├── CLAUDE.md              # Claude-specific additions
│   ├── .llm-wiki.toml         # Config
│   ├── index.md               # Auto-maintained catalog
│   ├── log.md                 # Append-only activity log
│   ├── sources/               # Raw parsed documents (8 categories)
│   │   ├── product/ design/ architecture/ development/
│   │   └── operations/ meetings/ references/ data/
│   ├── wiki/                  # AI-generated pages (9 types)
│   │   ├── summaries/ entities/ concepts/ comparisons/
│   │   ├── syntheses/ chronicles/ decisions/ runbooks/
│   │   ├── postmortems/ glossary.md
│   │   └── _template.md       # In each subdir
│   └── scripts/               # Python utilities
└── src/, docs/, ...           # Your project code

Quick Start

# 1. Clone
git clone https://github.com/your-org/llm-wiki.git
cd llm-wiki

# 2. Install dependencies
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# 3. Initialize wiki in your project
python scripts/init-wiki.py --target /path/to/your-project --name "My Project"

# 4. Ingest your first document
python .wiki/scripts/ingest.py meeting-notes.pdf --category meetings --output .wiki/sources/meetings/

# 5. Ask your AI tool to compile and query
# (Open your AI tool, it reads .wiki/AGENTS.md automatically)

Scripts

Script	Description
init-wiki.py	Initialize wiki structure
ingest.py	Parse documents → markdown sources
update-index.py	Rebuild index.md catalog
lint.py	Check wiki health (orphans, broken links)
stats.py	Wiki statistics
graph.py	Generate Mermaid knowledge graph
setup-qmd.sh	Setup qmd semantic search (optional)

Workflows

1. Ingest — Parse docs into sources/ (script, no AI)
2. Compile — AI reads sources → creates wiki pages
3. Query — AI searches wiki → answers → feeds insights back
4. Digest — AI deep-analyzes a topic across all sources
5. Lint — Check wiki consistency and health

Real-World Workflows

"How does our auth system work?"

New developer needs to understand a feature.

/wiki query "authentication flow"

Wiki searches all ingested docs, synthesizes answer with citations. If the answer reveals new insight, a wiki page is created automatically.

Feature spec changed after meeting

PM updates requirements after stakeholder meeting.

/wiki ingest meeting-notes-apr7.pdf --category meetings
/wiki compile

AI detects conflicts with existing spec and annotates them. Related wiki pages are cascade-updated.

Adding a new feature

Team starts building a new module.

/wiki ingest prd-payment-v2.pdf --category product
/wiki ingest api-spec.yaml --category architecture
/wiki compile

Creates summaries, extracts entities (services, APIs), links concepts. Query the wiki anytime during development for up-to-date context.

Post-incident analysis

Production incident happened, need to document.

/wiki ingest postmortem-2026-04-07.md --category operations
/wiki compile

Wiki links to related runbooks and past incidents. Lint ensures nothing falls through cracks.

Sprint planning prep

PM needs a comprehensive overview of a topic.

/wiki digest "payment module"

Deep cross-source report: progress, gaps, contradictions, open questions. Saved permanently, compounds over sprints.

Onboarding new team member

New hire reads the wiki instead of asking 10 people.

/wiki query "project overview"
/wiki query "how to deploy to staging"
/wiki status

Claude Code Integration

bash install.sh    # Manual install for Claude Code
# Or: npx skills add junbjnnn/llm-wiki  (works with all agents)
# Then use: /wiki init, /wiki ingest, /wiki compile, /wiki query, etc.

Requirements

• Python 3.11+
• git
• Optional: qmd for semantic search (recommended for 100+ pages)

Docs

• Getting Started
• For PM/PO
• For Developers
• Tool-Specific Setup

License

MIT