Build a personal LLM knowledge base — a checklist · Resource

Use this as a bench card. Run the Foundation block once, then keep the page open and tick the Feeding and Review boxes each time you add material. A box is done when it's true, not when you've read it.

Foundation (once, about 5 minutes)

[ ] Install a notes app that stores plain text (a "vault" is just a folder of markdown notes on your disk — nothing proprietary you can get locked out of).
[ ] Create a new vault: one fresh, empty folder with nothing in it yet.
[ ] Open that folder in Claude Code, an assistant that reads and writes the files on your machine directly.
[ ] Paste in a two-line brief: the idea ("this folder is my second brain") plus one setup instruction telling the agent to scaffold the structure and show you the first ingest.
[ ] Let it build the skeleton. You describe the system once; you don't hand-build any of it.

Check you have the four parts, plus the router

[ ] raw/ — the inbox. Where you drop sources and forget about them.
[ ] wiki/ — the processed knowledge. One source becomes many atomic notes (a note about exactly one thing: one idea, one tool, one company), densely linked to each other.
[ ] index — a table of contents the agent keeps current, every note linked from it. This is the map of the territory.
[ ] log — a dated history of what you added and when, so the system grows a bit at a time instead of rebuilding from scratch.
[ ] Rules file (CLAUDE.md) — the router that sits on top of the four. It tells the agent where each thing lives so it walks straight to the right note.

The rules file — what to put in it

[ ] Folder conventions: what raw/, wiki/, the index and the log are each for.
[ ] Routing by source type: where a dropped PDF goes, where a pasted link goes, where a page of notes goes.
[ ] "Keep it flat when in doubt." Extra folder layers cost you more in search friction than they buy you in tidiness.
[ ] The rule the others answer to: it must stay searchable — for you and for the agent alike. If a choice makes either of you hunt, it's the wrong choice.

Feeding rhythm

[ ] Add a source one of two ways: drop a file into raw/, or paste a URL and ask the agent to read and ingest it (ingest = read a source and file it into notes).
[ ] Don't tell it how many notes to make. One source becomes as many atomic notes as it needs — a short article might be one, a dense document a whole cluster.
[ ] After each ingest, glance at the index and the log. Fresh entries in both mean the source landed where it should.

Review (every so often)

[ ] Open the index and click through a handful of links. Does the map still make sense?
[ ] When it stops making sense, fix the rules file and re-run — don't rebuild the whole vault.
[ ] Watch for near-duplicate notes covering one thing under two names, and fold them together.

Signals to watch

✅ The agent finds the right note fast and answers straight from it. Leave the setup alone.
⚠️ It stalls, or burns through tokens hunting for something it should know where to find → the routing is off; tighten the rules file.
⚠️ The structure stopped fitting what you're storing → tune the conventions to the actual material, not to a template you copied.

Start with one vault, one topic, a handful of sources in the inbox. The map draws itself from there.