A gitagent-compatible AI agent that helps you build, maintain, and query personal knowledge systems using Obsidian + AI. Covers 4 distinct approaches — each solving a different problem for a different user.
| # | Approach | Best For | Setup Time | Complexity |
|---|---|---|---|---|
| 1 | Second Brain Building | Book/YouTube consumers who want AI-cited frameworks | Days | 6/10 |
| 2 | Personal Search Engine | Practitioners with 5-15 trusted blogs/podcasts | 1 week | 4/10 |
| 3 | AI Coding Vault | Developers spending too much on tokens re-explaining architecture | 6-12 weeks | 6/10 |
| 4 | Karpathy's LLM Wiki | Anyone with raw data piling up who wants zero-friction organization | 5 minutes | 2/10 |
Start with Approach 4 (Karpathy's LLM Wiki) unless you have specific needs. It's the fastest to value and simplest to set up.
# Install gitagent
npm i @open-gitagent/gitagent
# Run this agent directly from GitHub
gitagent run https://github.com/mingrath/obsidian-ai-knowledge-agent.git --adapter claude-code# Clone this repo
git clone https://github.com/mingrath/obsidian-ai-knowledge-agent.git
cd obsidian-ai-knowledge-agent
# Export to Claude Code format
gitagent export --format claude-code
# Or just start Claude Code in this directory
claudegitagent export --format system-prompt # Works with any LLM
gitagent export --format openai # OpenAI Agents SDK
gitagent export --format cursor # Cursor rules
gitagent export --format gemini # Google Gemini CLI| Skill | Description | Trigger |
|---|---|---|
vault-setup |
Set up a new Obsidian vault for any of the 4 approaches | "set up a vault", "help me get started" |
vault-compare |
Compare approaches, recommend best fit | "which approach?", "compare them" |
cross-vault-wire |
Connect vaults via CLAUDE.md references | "wire my vaults", "connect to my wiki" |
wiki-ingest |
Ingest raw sources into organized wiki pages | "ingest this", drops file in raw/ |
wiki-query |
Query wiki with citations, file answers back | Any question about the knowledge base |
wiki-lint |
Health-check for contradictions, gaps, orphans | "lint the wiki", "check wiki health" |
1. Create vault: raw/ + wiki/ + CLAUDE.md
2. Dump source: Drop article/transcript/PDF into raw/
3. Ingest: "Hey, ingest that article"
4. AI organizes: One article → 5-25 wiki pages with cross-references
5. Query: "What do I know about X?" → cited answer
6. Lint: "Check wiki health" → find gaps and contradictions
7. Compound: Each new source enriches the entire wiki
Wire multiple vaults together so any AI agent can access any vault's knowledge:
# In your executive assistant's CLAUDE.md:
## Wiki Path
Path: /path/to/my-brain-vault/
Reading order: hot.md → index.md → specific pages- One article → 23 wiki pages (10 min processing)
- 36 YouTube transcripts → full knowledge system (14 min batch)
- 383 files + 100 meeting transcripts → compact wiki → 95% token reduction
The knowledge/ folder contains the complete technical breakdown:
obsidian-ai-knowledge-systems.md— Full analysis of all 4 approaches with agent debate, verdict, layering strategy, and Learning in Public sectionnate-herk-transcript.md— YouTube video transcript showing practical Karpathy LLM Wiki implementationkarpathy-llm-wiki-gist.md— (Add Karpathy's original gist here)
obsidian-ai-knowledge-agent/
├── agent.yaml # gitagent manifest
├── SOUL.md # Agent identity & philosophy
├── RULES.md # Hard constraints & approach-specific rules
├── AGENTS.md # Framework-agnostic fallback instructions
├── skills/
│ ├── wiki-ingest/SKILL.md # Ingest raw sources into wiki
│ ├── wiki-query/SKILL.md # Query wiki with citations
│ ├── wiki-lint/SKILL.md # Health-check the wiki
│ ├── vault-setup/SKILL.md # Set up new vaults
│ ├── vault-compare/SKILL.md # Compare & recommend approaches
│ └── cross-vault-wire/SKILL.md # Connect vaults together
├── knowledge/ # Raw source documents (immutable)
│ ├── index.yaml # Source catalog
│ ├── obsidian-ai-knowledge-systems.md
│ └── nate-herk-transcript.md
├── memory/
│ ├── wiki/ # Organized wiki pages
│ │ ├── index.md # Master catalog
│ │ ├── hot.md # Recent context cache
│ │ └── sources/ # Source summary pages
│ ├── log.md # Operation history
│ └── runtime/ # Live agent state
│ └── context.md
└── README.md
- Andrej Karpathy's LLM Wiki — The original idea
- Nate Herk's implementation video — Practical walkthrough
- gitagent — Git-native agent standard
- swyx's Learn in Public — The missing output layer
- Sönke Ahrens' How to Take Smart Notes — Zettelkasten method
- Tiago Forte's Building a Second Brain — PARA method
MIT