Stop Hand-Rolling Claude's Second Brain—Obsidian CEO Has Already Open-Sourced the Shortcut

Every Claude session starts cold. Project background, decisions, last week's attempts all have to be re-explained. By evening, most of the context is already gone. The second brain is for exactly this: let Claude read a plain-text folder, write and maintain it itself, and the next session continues from where the last one stopped.

Andrej Karpathy proposed this pattern in April 2026. His gist has 5000+ stars, and related reports say views passed 20 million. But he only shared an idea, not specific instructions. Brucky turned it into a complete guide and also tested it.

Obsidian CEO kepano has already open-sourced obsidian skills, 49.8K stars, specifically teaching Claude to understand how Obsidian works. Together with the other three repos, someone has already done the setup for you.

Karpathy's division of labor is like this: Obsidian is the IDE, the LLM is the programmer, and the wiki is the codebase. Three-layer structure: raw holds raw material, which the model only reads; wiki holds pages written by the model; CLAUDE.md holds the rules. Three operations: ingest integrates new sources into 10-15 pages; query answers questions and saves good answers; lint finds contradictions, outdated claims, and orphan pages.

Four repos, all ready to use:

obsidian skills (kepano, 49.8K stars): teaches Claude Obsidian's native format.
claude obsidian (AgriciDaniel, 15K stars): wiki, ingest, query, lint as ready-made commands.
obsidian second brain (eugeniughelbur, 4.8K stars, 48 commands): runs on Claude Code, Codex, Gemini.
second brain starter (coleam00, 795 stars): interviews you and generates a nine-stage setup plan.

Note that many online tutorials still cite old star counts and installation scripts that no longer exist. These numbers are outdated.

In Brucky's guide, the most valuable points:

First, set it up in 10 minutes. Just create a folder, git init, start Claude, then have Claude interview you and write CLAUDE.md. You don't need to install any Local REST API plugin or MCP server first, unless you need to communicate with a running Obsidian.

Second, one vault for all projects. In the project CLAUDE.md, point to the same vault, first read wiki/hot.md (recent context cache), then read index.md. The last rule: don't read the wiki for general coding questions. This is cost control.

Third, add tests to the brain. Brucky planted a false statement in the wiki: Bush invented hypertext in 1941. That sentence is not in the source files. Let the model check itself, and it may let it pass because it sounds plausible. But a 95-line Python script with no AI reports at zero tokens: QUOTE NOT IN SOURCE. This script also checks frontmatter, whether links are valid, whether there are orphan pages, TODO, and whether source files have been modified.

Fourth, permissions rely on configuration, not prompts. CLAUDE.md is only context, not enforcement. Use permissions deny Edit(/raw/**) to make raw read-only. A hook automatically runs lint after every wiki edit, and error messages go directly back into the Claude conversation.

Fifth, automation. claude -p can run ingest non-interactively, with cron or the Schedule tab in Claude Desktop. Every night automatically process new sources.

This method is not without costs. Brucky wrote several failure points in the article:

Dead zone. With fewer than 50-100 sources, the wiki is like a search engine: you pay the setup cost and get search back, and many people give up here.
Bad sources spread. One wrong note is just one wrong note, but one wrong source entering the wiki pollutes 10-15 pages.
Scale. Critics say it only fits a scale indexable within one context window, roughly 50k-100k tokens. R&D World measured 760 pages in a one-month real-world test, and maintenance time roughly broke even with time saved.
The model checks itself. That's why deterministic scripts like brain.py are needed.

But Brucky's test proved that deterministic checks can catch problems that model self-review misses. By the sixth month, every session starts from where it stopped, and compared with people who open a blank session every day, the gap is structural.

The complete guide also includes the full text of brain.py, Python 3 standard library, 95 lines, and can do status, done, lint, and index. If you don't want to hand-roll it, you can also directly use one of the four repos above.

Save it if you want to save it. But since Obsidian's CEO has already open-sourced the shortcut, don't spend hours hand-writing YAML frontmatter anymore.