Karpathy Obsidian vault method

An Obsidian-focused X post on May 21 recirculated Andrej Karpathy’s AI knowledge-base workflow as a practical note-taking method: drop raw files into a folder, then use Anthropic’s Claude to turn them into readable notes with links between concepts. The post described a system that starts with source material rather than hand-written notes and uses AI to generate structured markdown inside an Obsidian vault. The method matches the “LLM Wiki” pattern that Karpathy described publicly in April 2026, when he said he had shifted from using large language models mainly for code generation to using them to build personal knowledge bases. The appeal of the workflow is operational rather than theoretical. Instead of tagging, filing and linking notes by hand, the user feeds the system raw material — articles, papers, transcripts or other files — and the model creates summaries, concept pages and cross-references. Community write-ups published after Karpathy’s April posts describe the same setup as a markdown-based knowledge layer, often viewed in Obsidian and maintained by Claude or Claude Code. (blog.starmorph.com) ### What is the workflow the Obsidian post is describing? Karpathy’s April 2026 pattern centers on a persistent wiki built from source documents, not a chatbot answering from scratch each time. A widely cited community guide summarizing his approach describes a three-layer structure: a `raw/` folder for immutable source files, a `wiki/` folder for AI-generated pages, and a `CLAUDE.md` file that defines the schema and rules for the system. (blog.starmorph.com) Fabian G. Williams, a Microsoft product manager who said he implemented the approach in Obsidian with Claude Code, described the flow more plainly: “raw sources in, compiled wiki out.” Williams wrote that the model generated summaries, backlinks, concept articles and cross-references in markdown, all viewable in Obsidian. ### Why are backlinks and concept notes getting so much attention? The cited Obsidian post emphasized automatic backlinks because linking is one of the chores that usually makes personal knowledge systems hard to maintain. (blog.starmorph.com) Karpathy’s public pattern, as described in April community documentation, shifts that work to the model: the LLM reads documents, identifies recurring ideas, creates structured summaries and updates related pages as the archive grows. (fabswill.com) Karpathy’s own idea, as quoted in a GitHub blueprint based on his April gist, was that standard retrieval-augmented generation forces the model to “rediscover” knowledge on every query, while a maintained wiki compiles that knowledge once and keeps it current. That is the distinction behind the Obsidian discussion: the note system is meant to become a standing artifact, not a temporary answer engine. ### How is this different from a normal Obsidian vault? (blog.starmorph.com) Obsidian is the interface in most of these examples, but the organizing logic changes. A conventional Obsidian vault often depends on the user to write notes, choose tags, create links and keep pages updated. In the Karpathy-style setup, the user curates inputs and the model handles much of the summarizing, linking, indexing and maintenance. Williams wrote that his vault had become “powerful but messy” before he added the pipeline, and said the missing piece was automation that could turn raw sources into structured knowledge. (github.com) His test run turned four YouTube transcripts into 21 cross-linked wiki articles, according to his April 5 post. ### Did Karpathy actually frame it this way? April 2026 community guides and repositories consistently trace the pattern back to Karpathy’s posts and gist about an “LLM Wiki.” One guide said his X post drew more than 16 million views and that the follow-up gist quickly drew thousands of stars, though Reuters could not independently verify those platform metrics from the original posts through publicly accessible pages. (blog.starmorph.com) (fabswill.com) The clearest public through-line is the architecture itself: source files go in, markdown pages are generated, and the wiki is maintained over time. A GitHub reference repo published on April 12 describes the result as a “persistent, compounding artifact,” with cross-references built in and updated as new material arrives. ### What would someone need to watch next? The next concrete sources are Karpathy’s original April 2026 X posts and gist, plus the community implementation guides that specify folder structure, schema files and Claude prompts. (blog.starmorph.com) The May 21 Obsidian post functions mainly as a rediscovery layer; the implementation details remain in the April documentation and the open GitHub blueprints built around it. (github.com)