A personal knowledge system with two layers: a library catalog of things you've consumed (books, films, music) and a research wiki of things you're thinking about. Unified behind one inbox and six named operations.
A book of instructions for summoning and directing entities. Which, if you squint, is what any knowledge system is — and more specifically, it is what this one is: instructions an LLM follows to route your material into the right place and help you think through it.
The library tells you what you've read; the wiki tells you what you're working through. Neither is complete on its own. The useful part is their intersection — when a concept you're drafting in the wiki resonates with a book you shelved three years ago, or when a pattern in your reading history suggests a synthesis you hadn't named yet. Grimoire makes that intersection cheap.
| Operation | You ask | It does |
|---|---|---|
| inscribe | "process my desk" / "ingest this" | Routes each item in desk/ to the right pipeline |
| bind | "let's do a weekly review" | Deep pass over recent scrolls, updates concepts, flags synthesis candidates |
| consult | "what do I have about X?" | Answers from the wiki with citations |
| divine | "draft a synthesis on X" | Writes a synthesis essay |
| audit | "lint the wiki" | Surfaces contradictions, orphans, stale claims |
| illuminate | "localize images" | Downloads remote images, rewrites refs |
Grimoire is conversational — every operation is you asking an LLM to run it. The files are plain markdown, so Obsidian works well as the reading and editing layer — point a vault at your grimoire folder and you get scrolls and shelves with backlinks, graph view, and search. The Obsidian Web Clipper browser extension sends articles straight into desk/, ready for the next inscribe pass.
A typical rhythm:
Daily-ish: drop new material (articles, bookshelf photos, notes) into desk/. When it's time to process, ask your LLM:
"Inscribe the desk."
Inscribe classifies each item and routes it: book/record/film photos to the Shelves import pipeline, clipped articles and other prose to a new Scrolls source page, anything else into the scroll-source bucket. Nothing sits stale.
Importing a whole shelf at once: take a photo of your bookshelf, film shelf, or record stack and drop it in desk/. Inscribe moves it to the Shelves photo inbox; ask for the extraction pass and vision reads every visible spine, creating a catalog entry for each — title, author (or director, or artist), year, themes, a short synopsis. Hundreds of books in minutes, no typing.
Any time a question comes up:
"Consult: what do I have about X?"
You get an answer synthesized from your wiki, with links back to the source pages and any relevant entries in your library.
Weekly, or when a batch of sources has accumulated:
"Let's do a bind pass."
Bind updates concept and entity pages with recent material, creates topic hubs when warranted, and flags synthesis candidates for you to consider later.
When an essay wants writing:
"Divine a synthesis on X."
Divine proposes the shape first (title, pages to draw from, argument sketch), waits for your approval, then drafts the essay under scrolls/synthesis/.
Maintenance, roughly monthly:
"Audit the wiki." / "Illuminate — pull remote images local."
Audit surfaces contradictions, orphans, and stale claims. Illuminate downloads remote images into raw/assets/ before the URLs rot, rewriting references to the local copies.
Grimoire doesn't watch your files — nothing runs until you ask.
git clone <this-repo> grimoire
cd grimoire
cp grimoire.example.json grimoire.json
# edit grimoire.json — fill in owner_name, intellectual_profile, etc.The Shelves scripts use only Python standard library — no pip install step. Requires Python 3.9 or later.
Grimoire is primarily an LLM-driven system — most day-to-day interaction is asking an LLM to run an operation. Claude Code with the skills in .claude/skills/ is the reference implementation.
pip install pytest
python3 -m pytest tests/ -vThe tests verify structural invariants: skill metadata, depersonalization, and that the Shelves scripts run clean on an empty catalog.
The depersonalization audit scans system files for strings that shouldn't leak into a public repo. It reads its patterns from tests/fixtures/audit_patterns.json (gitignored) if present, otherwise from tests/fixtures/audit_patterns.example.json (whose placeholders don't match anything real). To make the audit catch your personal data, copy the example to the gitignored file and fill in your name, handles, and system paths:
cp tests/fixtures/audit_patterns.example.json tests/fixtures/audit_patterns.json
# edit tests/fixtures/audit_patterns.jsondesk/ # drop new material here
shelves/ # library catalog + generated context
scrolls/ # wiki pages you and the LLM co-maintain
raw/ # ingested source material
.claude/skills/ # the six operations as Claude Code skills
grimoire.json # your personal config (gitignored)
- Shelves is a fork of mccoyspace/llmbrary.
- Scrolls was inspired by Andrej Karpathy's LLM-wiki gist.
- Operation names (inscribe, bind, consult, divine, audit, illuminate) borrow the grimoire metaphor. If you fork this and rename them, go wild.
MIT (see LICENSE). Fork freely; replace the example content with your own.