AI Second Brain – A personal knowledge‑base powered by an LLM agent

What it is – A complete, open‑source guide and starter kit for turning an Obsidian vault into a second brain that an LLM (Claude Code) continuously ingests, links and curates for you. The repository contains:

• a vault template (wiki + project layers) you clone locally,
• 18 “skills” (LLM‑driven workflows) and 72 slash commands that automate ingestion, linking, linting, stats, graph export, etc.,
• 6 sub‑agents that run on a schedule or on‑demand, and
• Python scripts for health checks and graph statistics.

All of the heavy lifting is done by an LLM agent that reads raw files (articles, PDFs, transcripts, chat exports), extracts concepts, creates densely‑linked markdown pages, and keeps the vault up‑to‑date. The user only needs Obsidian (as a markdown viewer) and the Claude Code runtime; no external vector database is required until the vault grows large enough for RAG.

---

Why it matters

• Self‑hosting & ownership – Your knowledge stays in plain markdown files you control; the only “cloud” component is the LLM you call.
• Automation of the boring work – Linking, contradiction tracking, and incremental updates are performed automatically, solving the classic “I never file what I save” problem.
• Two‑layer architecture – A wiki for what you know and a project folder for what you are doing, each with its own CLAUDE.md pipeline, keeping reference material separate from active work.
• Gradual scaling – Starts with pure markdown; only when you hit the “vector‑DB threshold” does the guide suggest adding GraphRAG.

---

Core components (as described in the README)

Component	Role
vault-template/	Starter Obsidian vault with the wiki‑structure, project pipeline, page templates and a sample CLAUDE.md file.
skills/	18 LLM‑driven skills, each implementing a workflow from the guide (e.g., ingest → extract → link → lint).
commands/	72 slash commands that expose the skills inside the Claude Code UI (e.g., /ingest, /graph‑export).
agents/	6 sub‑agents (four read‑only) that run scheduled maintenance, health checks, or on‑demand queries.
scripts/	Dependency‑free Python utilities for link checking, vault statistics, and graph export.
plugins/	Claude Code plugins that the Agents Course installs in two commands; they provide a context auditor, goal‑test generator, gate finder, harness auditor, eval bootstrapper, and loop critic.
docs/	The 65‑page “Second‑Brain Guide” (10 sections, 77 pages) plus the seven‑module Agents Course and several handbooks on graphs, harnesses, loops, etc.
resources/	Curated list of tools, papers, and other repos the author recommends.

---

How it works (high‑level flow)

Clip / save raw material → raw/ folder (articles, PDFs, chat logs)
        │
        ▼
   Claude agent runs the appropriate skill
        │   – extract concepts
        │   – create/ update wiki pages under wiki/
        │   – link to existing pages, record contradictions
        ▼
   Structured wiki (one idea per markdown page, dense wikilinks)

A parallel project layer mirrors the same pattern but is scoped to an active effort, each with its own CLAUDE.md pipeline (Inputs → Process → Outputs → Feedback).

---

Getting started (quick‑start steps from the README)

1. git clone the repo and copy the vault template to ~/brain.
2. Copy the skills, commands, agents and scripts into the appropriate hidden .claude folders.
3. Install Obsidian and open the copied folder as a vault.
4. Set up Claude Code (the LLM runtime) and optionally enable MCP for remote access.
5. Run claude inside the vault to start the agent.
6. Use the built‑in /ingest command (or the Web Clipper) to feed new sources.
7. Schedule maintenance via the provided agents (/metrics, /health, /graph‑export).

The guide walks through each step in detail, with separate pages for setup, ingestion, structuring, graph handling, agents, retrieval, outputs, maintenance, and troubleshooting.

---

Who should use it

• Researchers, engineers, writers, students – anyone who consumes a lot of content and wants it organized without manual filing.
• Solo users (the repo explicitly says it is not for team wikis or Notion replacements).
• People comfortable with a terminal and Obsidian, and who have access to Claude Code (or any LLM that can read files).

---

License & community

• MIT License – free to fork, modify, and use commercially.
• Contributions are welcomed; the CONTRIBUTING.md asks for source‑backed factual updates and justified tool additions.

---

TL;DR

second‑brain‑os is a fully‑documented, open‑source implementation of the “LLM‑wiki” pattern (Karpathy’s LLM‑wiki). It gives you a ready‑made Obsidian vault, a suite of Claude‑Code skills/commands, and a step‑by‑step guide to let an LLM automatically turn your raw reading material into a living, linked knowledge base you own.