Comment je me suis construit un « deuxième cerveau » IA (Obsidian + Claude Code)

Summary

This video demonstrates how to build an “AI second brain” that behaves more like a persistent personal wiki than a chat model with short-term memory. The core problem addressed is that LLMs often respond by re-selecting small chunks from your documents each time—meaning knowledge doesn’t truly accumulate between sessions (a kind of LLM forgetting).

Main idea: a persistent wiki that “self-compiles”

Inspired by a concept discussed by Andrej Karpathy, the approach is to transform Claude into an agent that maintains a growing wiki. Instead of having the LLM reread sources for every question, it builds a 3-layer persistent knowledge system:

1. Raw sources layer (read-only)

   • Contains PDFs, articles, transcripts, and similar materials.
   • The LLM reads these but does not modify them.

2. Wiki layer (generated/maintained)

   • Uses Markdown pages (e.g., *.md) as concept pages.
   • Pages represent concepts/entities/sources/summaries, connected with Wikipedia-like links.
   • As new knowledge arrives, existing pages are enriched/updated.

3. “Diagram / conventions” layer

   • A cloud.md file acts like a contract/schema.
   • It instructs the LLM on how to structure and follow naming/content conventions.

The system is positioned as:

• Obsidian = the IDE
• Claude = the “programmer” agent
• Wiki pages = the growing codebase

Goal: a system that self-compiles, self-maintains, and grows as you ingest new sources.

---

Why Obsidian (features highlighted)

• Local + Markdown: stored as local files (not locked to a proprietary database or cloud service).
• Graph view: visualizes connections between notes/concepts.
• Markdown-friendly: Claude can write/summarize Markdown easily.

---

Implementation tutorial (Obsidian + “Claude Code” workflow)

1. Install Obsidian (Windows/Mac tutorial noted as working).
2. Create a vault in Obsidian (example name shown like “AI monitoring system”).
3. Open the vault folder in VS Code.
4. Launch Claude Code (triggered at the folder level).
5. Copy prompt/structure instructions from a GitHub link (referenced in the description) to create the required files, including:
   • cloud.md (schema/rules)
   • index.md
   • log.md
   • plus custom command files under a command system

---

Custom commands (agent operations)

The generated structure includes agent commands such as:

• ingest.md: processes new raw documents and generates/updates wiki pages
• lint.md: checks for errors/contradictions/inconsistencies
• query.md: answers questions using the wiki
• save.md: stores/classifies outputs back into the wiki (as summaries, etc.)

---

Ingestion flow (browser extension + ingest command)

1. Install Obsidian Web Clipper (Chrome/Firefox).
2. Clip web articles into the vault. Initially, they’re placed in a raw folder structure (e.g., a raw/row-style layout).
3. Example workflow:
   • Clip multiple articles.
   • Rename ingested files to avoid special characters/spaces (e.g., article 1.md, article 2.md).
   • Run ingest so Claude parses content and generates/updates:
     • concept pages
     • entity/project pages
     • summaries
4. The example shows that added articles result in interconnected wiki pages, visible via Obsidian’s graph links.

---

Query + continuous learning behavior

After ingestion, the system can answer questions like:

• “What happened with the Axios supply chain attack?”

The agent can then propose to classify the answer as a wiki summary, which is written back into Obsidian. This creates an iterative loop:

asking → producing synthesis → saving/updating wiki

---

Maintenance / quality control

• lint detects contradictions or errors as the knowledge base evolves.
• A “health/inspection” concept is mentioned: the system can report whether the second brain remains consistent as it grows.
• The tutorial emphasizes that after adding new inputs to the Obsidian vault, you can re-run ingestion and maintenance to keep everything synchronized and properly connected.

---

Main speakers / sources (as mentioned)

• Andrej Karpathy — referred to as “co-founder of OpenAI,” source of the tweet/concept
• Claude — the Anthropic model being turned into the wiki agent
• The tutorial narrator — the practical guide demonstrating the Obsidian/Claude Code setup