Adapting Karpathy’s “LLM Wiki” workflow (Claude + Obsidian) to Claude + Capacities
Adapting Karpathy's "LLM Wiki" workflow (Claude + Obsidian) to Claude + Capacities
Andrej Karpathy described a workflow where an LLM ingests raw sources (articles, books, videos) and automatically builds/maintains a cross-linked knowledge base — "Obsidian is the IDE, the LLM is the programmer, the wiki is the codebase." It's built with Claude Code + Obsidian Skills (wikilinks, callouts, a CLI) and runs on three commands: ingest a source, process an inbox of quick notes, and lint the wiki for broken links/gaps.
I'd love to run this same workflow with Claude + Capacities instead of Obsidian, using Capacities' native objects, properties, and tags in place of markdown wikilinks. What's the best way to do this today?
Specifically:
- Can Claude (via MCP/API) read full text of long objects like PDFs, or only metadata/search?
- Can it create new typed objects with properties programmatically, not just search/read/append to daily note?
- Any recommended pattern for "raw source" vs. "LLM-synthesized" objects so this stays organized at scale?
Would really value official guidance (or a template) for building this — feels like a natural fit for Capacities' object model, possibly even better than Obsidian's flat file structure.
2 Comments
Sign in to comment
Thanks for the detailed idea and context. Based on the current API docs, Capacities’ API and MCP connectors are still limited: there is no documented way yet to (1) have Claude read full PDF text via API/MCP, or (2) create arbitrary typed objects with properties programmatically in the same way Obsidian’s CLI workflows allow. There is also no official pattern or template documented for separating “raw source” vs “LLM-synthesized” objects at scale. This workflow is not supported as described today; if you haven’t already, you can reference the API docs here for current capabilities: https://docs.capacities.io/developer/api
This reply was written by the Capacities AI Assistant. Please note AI may make mistakes. If this response doesn't resolve your question or issue and the community can't help either, feel free to reach out to us at
Thanks for the clarification — makes sense that this isn't fully supported yet.
Given those limitations (no full-PDF read via API/MCP, no programmatic creation of typed objects, no raw/wiki separation pattern), what would be the best way to approximate Karpathy's system within Capacities today?
For context on what I'm trying to replicate, here's how Karpathy describes the three pieces:
"Raw is your junk drawer. Articles, transcripts, docs, notes. Dump it in without organizing. That's the AI's job. Wiki is what the AI builds from raw. Summaries, concept pages, entity pages, an index it uses to navigate everything. You never write the wiki. The AI writes and maintains it entirely, every session. Schema is a single CLAUDE.md file that tells the AI what your knowledge base is about and how to organize it. You write this once. The AI follows it every session."
The core value is that the AI does the bookkeeping automatically, every session, without the human maintaining it manually.
Given Capacities' current constraints, is there a recommended interim approach? For example:
Appreciate any guidance on the best current workaround, and it'd be great to know if native support for this kind of workflow is on the roadmap.