Skip to content

How it works

 you save things                 the worker (sb run)                        you read
 ───────────────                 ──────────────────                        ────────
 Web Clipper → inbox/  ┐
 forwarded email        ├─►  queue ─► fetch & extract ─► LLM writes a  ─►  wiki/ notes + index.md + log.md
 an older notes folder ┘    (SQLite)  (web, PDF, note)   JSON edit plan     wiki/daily/YYYY-MM-DD.md  ◄── start here

Three layers, as in Karpathy's design:

  • raw/: the original of every source, never edited.
  • wiki/: notes written and maintained by the worker: one source note per source, plus concept, entity and synthesis pages that grow as more sources arrive. You read it in Obsidian; the worker writes it.
  • SCHEMA.md: the conventions the model is given on every run.

Design rules worth knowing:

  • The model never touches your files. It only returns a JSON "edit plan". The worker validates the plan (drops invented links, rejects wrong-language or malformed output, retries once) and only then writes.
  • Everything is a git commit in the vault repo (when git is installed; it is optional), one per ingested source, so any change can be reviewed or reverted.
  • Risky things go to review, not to the wiki: contradictions and lint findings become notes under wiki/review/. Nothing is merged or deleted silently.
  • It works with small local models (Ollama) and with cloud models through the same interface.