Obsidian CEO Has Open-Sourced His Second Brain: 7 Free Repos and a Counterintuitive Note-Taking Approach
Obsidian CEO Steph Ango has open-sourced his own note vault and seven free repositories on GitHub. But what's more worth paying attention to than the files is the underlying idea of treating a second brain as a compiler. This article breaks down how the system works—and whether it's worth copying.
Obsidian CEO Steph Ango has open-sourced his own note vault. All seven repositories on GitHub are free, someone strung the links together, and the comments section was full of “this is gold.” He even gave away the vault template he’s currently using: download it, unzip it, point Obsidian at that folder, and you have a second brain that someone else has spent years refining.
The seven repositories, divided by official and CEO-personal:
kepano/kepano-obsidian: the CEO's own vault template; its directory structure is how he organizes his notes.obsidianmd/jsoncanvas: an open canvas format. Whiteboard data is stored as.canvasfiles, and the files always belong to you.obsidianmd/obsidian-sample-plugin: the official plugin template, written in TypeScript. If you want to build a plugin, start with this code.kepano/obsidian-minimal: a minimal theme that makes your vault look clean.obsidianmd/obsidian-clipper: the official web clipper; save any page to your vault with one click.obsidianmd/obsidian-importer: import old notes from other note-taking apps, so you don't have to start from scratch.obsidianmd/obsidian-help: a plain-text version of the official docs; you can read the files directly to understand the features.
Here is the video attached to the tweet, showing what the vault actually looks like.
But if the only thing you want to copy is the folder structure, you're missing the point.
A long article by rvaniaaa mentions the LLM Wiki that Andrej Karpathy (a founding member of OpenAI) proposed in April 2026. The gist of his point was: RAG re-derives knowledge on every query; a compiled wiki derives it once and then keeps updating it. That distinction changes how you build things.
Storage and compilation are two different operations. Storage is building a library: you put notes in and search them when needed. It does not connect knowledge on its own, and it does not discover when two notes contradict each other. Compilation has a process: when new material comes in, the system reads it, extracts key points, links it to existing notes, updates related pages, flags contradictions, and then files away the synthesized understanding. A library only gets bigger; a compilation system gets smarter.
The architecture comes down to a directory structure. Three folders, one file, one loop.
raw/ is the input buffer: anything new gets thrown here first. It's not the answer; it's just raw material. wiki/ is where compilation happens; a single piece of material can touch ten to fifteen pages, automatically create links, and flag contradictions. output/ is generated from the compiled knowledge. At the core is CLAUDE.md, a living document about who you are, how you think, and what you've tried, which the model reads before every session. The whole loop doesn't require manual organizing, and after the system finishes processing, it gives you a summary of changes.
The article offers a timeline: after one month, context no longer disappears between sessions. After three months, the wiki starts surfacing connections you never deliberately made—an idea from January linked to a note from last week. After six months, the gap between you and someone starting from scratch is structural. After a year, you open graph view and see hundreds of nodes all connected.
Some commenters offered more practical perspectives. One said that compilation quality depends entirely on source quality. Bad material is easy to delete in a library, but in a compilation system it may already have touched fifteen pages. Another said token costs are a real concern: doing deep compilation for every random document is likely not worth it. High-frequency, high-value content is worth it; one-off documents, not so much. Another person warned that after a few hundred documents, automatic categorization and linking by the model becomes error-prone unless the process is defined strictly enough.
Also, this workflow depends on Claude Desktop and a paid plan; it won't run on the free tier.
The value of the Obsidian CEO open-sourcing all this isn't that you should copy the folders verbatim. It gives you a zero-cost starting point: store your second brain in plain-text files, then let AI compile it. The official Clipper handles collecting external information, and Importer handles migrating old notes. Community developers have also built tools specifically to let agents directly call on knowledge in this kind of wiki, such as AKM.
Ironically, most people still use Claude as a polite search engine. You ask, it answers, you close the tab, and tomorrow it remembers nothing. You're still the one single-handedly preserving context. The compiler approach shifts that burden to the system. That may be what's genuinely exciting about it.
发布时间: 2026-11-22 03:59