Build guide

I Built a Second Brain with Claude Code + Obsidian

An LLM knowledge base: dump raw sources into a folder and Claude Code builds and maintains the wiki.

Claude CodeObsidian
May 18, 2026
Optimized for

What you'll build

An LLM Knowledge Base - Karpathy's term for it. You stop using LLMs to just answer questions and start using them to build and maintain a wiki of everything you're learning. You dump raw sources into a folder, Claude Code compiles them into a clean markdown wiki with backlinks and summaries, and you view the whole thing as a connected graph inside Obsidian.

Mine holds my entire personal brand: around 100 articles, every long-form transcript I've published, a voice fingerprint, a hook library - all connected as one brain. I never wrote any of it by hand. Claude did. I just drop sources in, run an ingest skill, and the wiki grows itself.

The conceptual shift is the whole video: stop using LLMs to manipulate code, start using them to manipulate knowledge.

What you'll need

Step 1: Set up the two-layer vault

Create a folder with two layers inside:

my-vault/
  raw/    ← immutable source documents
  wiki/   ← the LLM-compiled layer

raw/ is where everything you collect gets dumped exactly as-is - articles, papers, transcripts, screenshots, Notion exports. You never edit these. wiki/ is the layer Claude writes and maintains: clean markdown pages with backlinks, summaries, and categorized concepts.

Open the whole folder as a vault in Obsidian. Obsidian isn't doing any of the thinking here - it's the IDE and the graph view for what Claude builds.

Step 2: Build an ingest skill in Claude Code

The operational core is a skill - mine is /vault-ingest. Its job: take a new source, file it into raw/, then compile it into the wiki - create or update the relevant pages, write summaries, categorize the concepts, and link everything with [[backlinks]] so the graph stays connected.

That's the division of labor the whole pattern rests on: you barely touch the wiki. The LLM writes it. You just view it.

Step 3: Drop sources in and let the wiki grow itself

From here the loop is one move: something worth keeping shows up - an article, a transcript, a screenshot - you drop it in and run the ingest skill. Every ingest makes the graph denser, because new pages link back into old ones.

I also run maintenance skills on top: /vault-lint health-checks the vault, and /vault-emerge scans it for patterns I haven't noticed myself. Same idea - Claude maintains the knowledge base, not you.

Step 4: Ask it real questions

The payoff shows up once the vault is big enough. I can ask something like "how has my thinking on voice agents shifted over six months" - and Claude goes and reads 22 transcripts and actually answers it. That's not a question a chat window with no memory can touch.

Keep raw/ immutable. The sources stay exactly as captured - all the compilation, summarizing, and linking happens in the wiki layer on top. That separation is what makes the whole thing trustworthy.

That's it. Two folders, one ingest skill, and Obsidian as the window into it. Good luck and happy building.

Resources

← Back to the library Good luck and happy building