Workflow
Obsidian LLM Wiki Workflow
A practical workflow for running an LLM wiki in Obsidian: raw sources, maintained wiki pages, schema rules, ingest passes, query loops, and lint cycles.
Why this page matters
The operating model has three layers and three recurring operations: ingest, query, and lint.
The three layers
Karpathy describes a simple stack. Raw sources hold the original material and stay unchanged. The wiki is a set of markdown pages generated and updated by the model. The schema is the instruction file that teaches the agent how to ingest, answer, and maintain the wiki.
That schema layer is easy to underestimate. Without it, the model behaves like a generic assistant. With it, the model can behave like a disciplined maintainer that follows naming, linking, logging, and filing conventions over many sessions.
- Raw sources: articles, notes, papers, images, transcripts, data.
- Wiki: summaries, concepts, entities, comparisons, synthesis pages.
- Schema: the conventions, directory structure, and workflows the agent follows.
The recurring operations
Ingest means reading one new source, discussing what matters, then updating the wiki and its logs. Query means answering questions against the wiki and filing valuable answers back into the system. Lint means auditing the wiki for stale claims, missing links, orphan pages, and gaps that deserve more research.
This loop is what keeps the knowledge base from decaying. The model is not just generating pages; it is keeping them healthy.
Why Obsidian fits
Obsidian gives the human a fast view into the markdown graph while the model does the writing. Backlinks, graph view, local files, plugins, and git-friendly storage all line up with the pattern.
Karpathy's description of keeping the agent open on one side and Obsidian on the other captures the feel of the workflow: the wiki behaves like a codebase, the model behaves like the programmer, and the human acts like the reviewer and research lead.