Scuti AI builds LLM Wiki with Claude Code

Scuti AI’s May 24 post slots into a fast-growing developer pattern: using an LLM agent to maintain a personal knowledge base instead of querying raw documents each time. The company said it built an “LLM Wiki” with Anthropic’s Claude Code and the note-taking app Obsidian, framing the project as a “second brain” for research and accumulated notes. (gist.github.com) Andrej Karpathy’s April 4 GitHub gist is the clearest upstream source for the idea Scuti cited. (gist.github.com) Karpathy described “a pattern for building personal knowledge bases using LLMs” and said the goal was to copy the file into an LLM agent such as Claude Code, which would then help build the system in collaboration with the user. ### What did Scuti AI actually build? Scuti AI said on May 24 that it had assembled an “LLM Wiki” using Claude Code and Obsidian, rather than a conventional retrieval setup. The project follows the now-familiar “second brain” framing: a store of notes and source material that is meant to persist and improve as new documents are added. (gist.github.com) Karpathy’s gist lays out the same core premise in more concrete terms. Instead of relying on retrieval-augmented generation, or RAG, at question time, the LLM reads new sources, extracts information, and updates a structured markdown wiki that sits between the user and the raw files. (gist.github.com) ### How is this different from uploading files to a chatbot? Karpathy’s April 4 note contrasts the wiki approach with standard document-chat workflows. In his description, tools such as NotebookLM, ChatGPT file uploads and “most RAG systems” retrieve relevant chunks each time a question is asked, while the wiki model “incrementally builds and maintains a persistent wiki.” (gist.github.com) The gist says that difference matters because the knowledge is “compiled once and then kept current, not re-derived on every query.” Karpathy wrote that the result is a “persistent, compounding artifact” in which cross-references, contradictions and synthesized summaries are already in place before the next question arrives. (gist.github.com) ### Where do Claude Code and Obsidian fit in? Karpathy’s own description names both tools in the workflow. He wrote that, in practice, he keeps “the LLM agent open on one side and Obsidian open on the other,” with the model making edits while he browses updated pages, links and graph view in real time. (gist.github.com) A GitHub guide built around Karpathy’s method shows the operational split more explicitly. In that setup, Obsidian serves as the vault and browsing interface, a `raw` folder holds source material, and Claude Code processes new files, creates wiki pages, updates indexes and builds links between entities. (gist.github.com) ### What does the workflow look like in practice? The GitHub implementation describes a simple loop. A user clips or drops a new article into a `raw` folder, opens the vault in Claude Code, asks the agent to ingest the file, and then reviews the updated graph and pages in Obsidian. (gist.github.com) That guide says Claude Code can create files such as `CLAUDE.md`, `index.md` and `log.md`, then ingest new documents and update relationships across the wiki. It also lists maintenance prompts including health checks, linting for inconsistencies and plain-language queries against the knowledge base. (github.com) ### Where can readers see the underlying idea? Karpathy’s GitHub gist remains the primary public reference for the pattern Scuti cited. The gist was created on April 4, 2026, and presents the architecture as an idea file meant to be pasted into tools including Claude Code. (github.com) Scuti AI’s implementation notes were published on May 24 on the company’s website. Readers looking for the original schema and workflow can start with Karpathy’s gist, then compare it with Scuti’s write-up and other public Claude Code-plus-Obsidian builds that have appeared since April. (gist.github.com) (github.com)