Setting Up the LLM-Wiki Pattern in Obsidian

Based on Karpathy's llm-wiki pattern — using LLMs to incrementally build and maintain a persistent, interlinked wiki from raw sources rather than re-deriving knowledge on every query.

Quick Start — Paste This Into Your Agent

Copy the prompt below into Claude Code, Codex, or any LLM agent with file access. It will scaffold the wiki, configure the tooling, and update your CLAUDE.md with the schema. Then you just start ingesting sources.

I want to set up the llm-wiki pattern in my Obsidian vault. Follow this guide exactly:
https://gist.github.com/kennyg/6c45cace2e1c4e424a28fcd51dd6c25b

Here's what I need you to do:

1. SETUP: Configure local image capture (attachment folder, hotkey). Set up the
   Obsidian CLI (symlink into PATH). Install and configure qmd for search.

2. SCAFFOLD: Create the Wiki/ directory structure (sources/, entities/, concepts/,
   synthesis/). Create starter index.md (with unprocessed sources list), log.md,
   and overview.md.

3. SCHEMA: Add an "LLM Wiki" section to CLAUDE.md documenting the architecture,
   wiki structure, page conventions, and the ingest/query/lint operations. This
   is the schema that makes you a disciplined wiki maintainer.

4. CALIBRATE: Ingest 2-3 of the richest sources interactively so I can review
   the templates. Adjust based on my feedback before batching the rest.

Important rules:
- Raw sources (Clippings/, or wherever my sources live) are IMMUTABLE. Never modify them.
- You own Wiki/ entirely. Always update index.md and log.md on every change.
- Use [[wikilinks]] heavily for Obsidian graph view.
- Every wiki page gets type: in frontmatter and wiki/* tags.
- Keep source summaries factual. Interpretation goes in concept/synthesis pages.
- When sources contradict each other, note it explicitly — don't silently overwrite.

Start by exploring my vault structure to understand what raw sources I have,
then walk me through the setup step by step.

Part 1: Local Image Capture

Clipped articles often have remote image URLs (e.g. pbs.twimg.com) that break over time. Download them locally to keep your vault self-contained.

1a. Set attachment folder path

In Obsidian: Settings > Files and links > Attachment folder path → set to assets/

Or edit .obsidian/app.json directly:

{
  "attachmentFolderPath": "assets/"
}

1b. Bind a hotkey for downloading attachments

Create or edit .obsidian/hotkeys.json:

{
  "editor:download-attachments": [
    {
      "modifiers": ["Ctrl", "Shift"],
      "key": "D"
    }
  ]
}

1c. Workflow

After clipping an article with Obsidian Web Clipper:

Open the clipped note in Obsidian
Hit Ctrl+Shift+D to download all remote images to assets/
Image references are rewritten to local paths automatically

Find existing clippings with remote images:

grep -rl '!\[.*\](http' Clippings/

Part 2: Obsidian CLI Setup

The Obsidian CLI (built into Obsidian 1.12+) lets agents read, create, search, and manage notes programmatically.

2a. Register the CLI

In Obsidian: Settings > General > Command line interface > Register CLI

2b. Symlink into PATH

The binary lives inside the app bundle but isn't in your PATH by default:

ln -s /Applications/Obsidian.app/Contents/MacOS/obsidian ~/.local/bin/obsidian

Verify:

obsidian help

(Ignore harmless FATAL:electron warnings — it works fine.)

2c. Useful CLI commands

obsidian read file="My Note"              # read a note
obsidian create name="New Note" content="# Hello" silent
obsidian append file="My Note" content="New line"
obsidian search query="search term" limit=10
obsidian tasks todo                        # list open tasks
obsidian task ref="Inbox/Tasks.md:2" toggle  # toggle a task
obsidian daily:read                        # read today's daily note
obsidian property:set name="status" value="done" file="My Note"

Part 3: Wiki Architecture

Three layers:

Raw sources — immutable clippings and captures. LLM reads, never modifies.
The wiki — LLM-generated, interlinked markdown. Summaries, entity pages, concept pages, synthesis.
The schema — instructions in CLAUDE.md (or AGENTS.md for Codex) telling the LLM how to maintain the wiki.

3a. Directory structure

Wiki/
├── index.md          # content catalog — LLM reads this first
├── log.md            # append-only operation log
├── overview.md       # high-level synthesis of everything
├── sources/          # one summary per ingested source
├── entities/         # people, tools, orgs, repos
├── concepts/         # ideas, patterns, techniques
└── synthesis/        # query answers filed back into wiki

Create the scaffold:

mkdir -p Wiki/{sources,entities,concepts,synthesis}

3b. Special files

index.md — content catalog with tables for sources, entities, concepts, and synthesis pages. Also tracks unprocessed sources. The LLM reads this first when answering queries or starting any operation.

log.md — append-only record. Each entry starts with ## [YYYY-MM-DD] operation | Title so it's parseable with grep:

grep "^## \[" Wiki/log.md | tail -5

3c. Page conventions

Every page has type: in YAML frontmatter (source-summary, entity, concept, synthesis)
Tags use wiki/ prefix namespace (wiki/source, wiki/entity, etc.)
Heavy [[wikilinks]] everywhere for Obsidian graph view
date_updated: tracked in frontmatter
source_count: on entity and concept pages
[key::value] inline metadata for Dataview queries
confidence: on concept pages (high/medium/low) tracks how well-supported a claim is

Part 4: Operations

Ingest

Process a raw source into the wiki:

Read the raw source completely
Create a source summary in Wiki/sources/
Create/update entity pages in Wiki/entities/
Create/update concept pages in Wiki/concepts/
Update Wiki/index.md (add to tables, remove from Unprocessed)
Update Wiki/overview.md if the big picture changed
Append to Wiki/log.md

A single ingest typically touches 5-15 wiki pages. Seed with 2-3 rich sources first to calibrate your templates, then batch the rest.

Query

Ask questions against the wiki:

Read Wiki/index.md to find relevant pages
Read relevant wiki pages (not raw sources — the wiki should have what you need)
Synthesize an answer with wikilinks
If the answer is substantial, file it as a new page in Wiki/synthesis/
Update index and log

Filing query answers back into the wiki is key — your explorations compound in the knowledge base just like ingested sources do.

Lint

Health-check the wiki:

Orphan pages (no inbound links)
Broken wikilinks
Stale pages (date_updated older than newest relevant source)
Contradictions between pages
Concepts mentioned in prose but lacking their own page
Missing cross-references

Part 5: CLAUDE.md Schema

Add a "LLM Wiki" section to your CLAUDE.md documenting all of the above — the architecture, wiki structure, page conventions, operation procedures, and rules. This is the schema layer that makes the LLM a disciplined wiki maintainer rather than a generic chatbot.

Key rules to include:

Never modify raw source files
Always update index.md and log.md on every wiki change
Keep source summaries factual; interpretation goes in concept/synthesis pages
When new sources contradict existing wiki content, note contradictions explicitly rather than silently overwriting

Part 6: Search with qmd

At small scale (~25 sources) the index file is enough. As the wiki grows, qmd provides hybrid BM25/vector search with LLM re-ranking, all on-device. It also has an MCP server so your LLM agent can use it as a native tool.

6a. Install

# macOS
brew install tobi/tap/qmd

# or build from source
go install github.com/tobi/qmd@latest

6b. Index your vault

# qmd auto-discovers markdown files
qmd collection add my-vault /path/to/vault "**/*.md"
qmd update
qmd embed

6c. Add context for better results

qmd context add qmd://my-vault/ "Obsidian vault with LLM wiki, raw sources, and notes"
qmd context add qmd://my-vault/Wiki "LLM-generated wiki: source summaries, entities, concepts, synthesis"

6d. Search

qmd query "tools for maintaining state across LLM sessions"
qmd search "MCP"              # BM25 keywords only (no LLM)
qmd vsearch "agent memory"    # vector similarity only

6e. Use as MCP server

Add to your LLM agent's MCP config so it can search the wiki as a tool:

qmd mcp  # starts stdio transport

6f. Keep fresh

After adding new wiki pages:

qmd update && qmd embed

Part 7: Recommended Obsidian Plugins

These plugins complement the wiki workflow:

Dataview — SQL-like queries over frontmatter. Since every wiki page has type:, source_count:, date_updated:, etc., Dataview can generate dynamic tables and dashboards.
Obsidian Web Clipper — browser extension that converts articles to markdown. The fastest way to get sources into your raw collection.
Graph View (core) — the best way to see the shape of your wiki. Hubs, orphans, clusters.
Tasks — if you track ingest tasks or lint findings as checkboxes.

Part 8: Tips

Start small. Ingest 2-3 rich sources manually to calibrate templates before batching the rest.
Let the LLM write everything in Wiki/. You read it; the LLM writes it. Don't hand-edit wiki pages.
File query answers back. The synthesis folder is how your explorations compound.
Lint periodically. Ask the LLM to health-check the wiki every couple weeks.
Raw sources are immutable. Even if metadata is wrong, the wiki layer is where you add clarity.
Use the Obsidian CLI for task management instead of editing markdown files directly — it keeps Obsidian in sync.
The schema evolves. Update CLAUDE.md as you discover what works for your domain. The LLM and human co-evolve it over time.

kennyg/llm-wiki-obsidian-setup.md

Select an option

No results found