Open Source · Obsidian + Claude Code

Your knowledge base.
Self-maintaining.

LLM-Wiki is an Obsidian vault template operated by Claude Code — a Karpathy-style second brain where the AI reads, writes, and maintains your notes alongside you. One vault, built to run multiple ventures and client projects. Not a chatbot. A vault maintainer.

6
Core Workflows
9
Note Templates
6
Domain Hats
MIT
License
Terminal
# open vault in terminal, then:
$ claude

# plain-english commands:
process the inbox
lint the vault
promote Atlas 🧠/1-Literature Notes/RAG.md
what are the tradeoffs between agent memory architectures?

Three components. One operating model.

Most note systems are passive. The LLM-Wiki pattern makes your knowledge base an active participant — an AI that reads, writes, and maintains it alongside you.

01
CLAUDE.md — the contract
The operations schema at the vault root. Defines the folder structure, what each folder means, and exactly how to run each of the 6 core workflows. Not documentation — executable specification. Claude Code reads this at session start and follows it precisely. That precision is what makes the system reliable over months, not just one session.
02
Claude Code — the operator
Opens the vault, reads index.md for navigation, and executes workflows on plain-English command. The difference is the direction of agency: rather than answering questions from the outside, the LLM operates in your knowledge base and builds it with you over time.
03
Obsidian — the interface
Browse notes, explore the graph view, use Dataview queries, write in context. The bundled .obsidian/ pre-registers Dataview and Templater — accept the install prompt and everything is ready. You keep full ownership of your vault in a local markdown folder.
00-Inbox/
Ingest
Literature Note
Promote
Concept Page
MOC
Outputs/

Two streams in. PDFs stay out.

Everything enters through 00-Inbox/, which has two purpose-built streams. Heavy source files never need to live in the vault at all.

✍️ Fleeting
Your own jottings
A half-formed idea, a meeting decision, a quote you want to sit with — written by hand into 00-Inbox/Fleeting Notes/. When you process the inbox, Claude Code expands each one into the vault and then deletes the original: the value has been fully transferred, so there's nothing left to keep.
📎 Clippings
Web articles, one click
Save anything from the browser with the Obsidian Web Clipper extension — it drops the article straight into 00-Inbox/Clippings/. On ingest these become literature notes and the original is archived to Archive/Clippings/, where its source URL and metadata keep independent reference value.
📚 PDFs & books
Read in Zotero, ingest highlights
Large PDFs would bloat the vault, so they stay external in Zotero. You ingest your annotation exports — what you actually highlighted — rather than the raw file, which is also what you'd want distilled. Small reference PDFs can sit in Assets/raw/pdfs/; anything big is pointed at by path and left where it is.

The operations schema in practice.

Six workflows cover the full knowledge lifecycle — from raw capture to synthesised concept pages. All triggered with plain English.

Workflow 01
Ingest
Drop a source into the inbox, get a fully structured literature note placed in the right topic folder. Concept pages are enriched if the source adds something. The original is archived or deleted per type. One command processes the entire inbox.
process the inbox
Workflow 02
Query
Ask a question in natural language. Claude Code reads index.md, locates the relevant concept pages, and synthesises an answer citing them. If the synthesis surfaces something worth capturing permanently, it writes it back.
what are the tradeoffs between agent memory architectures?
Workflow 03
Lint
Six automated health checks: orphan pages, broken wikilinks, index coverage, stale one-liners, contradiction scan, and missing Maps of Content. Run once a month to keep the vault from drifting into abandonment.
lint the vault
Workflow 04
MOC Creation
Build a topic-level Map of Content entry page once a topic crosses ~4 pages. MOCs become the navigation layer — they don't hold knowledge, they aggregate it. Generated on command and kept in sync as the topic grows.
create MOC for agent memory
Workflow 05
Log Rotation
Every operation is logged in log.md. When the log exceeds 30 entries, the oldest are rotated to Archive/archived-logs.md in one step. The active log stays fast to scan; the full history is never lost.
rotate the log
Workflow 06
Promotion
The Zettelkasten distillation step, made explicit. Ask Claude Code to "promote" a literature note: it reads the note, identifies 1–N universal concepts worth standing alone, creates or enriches the corresponding concept pages, and marks the literature note with a promotes: frontmatter field.
promote Atlas 🧠/1-Literature Notes/RAG.md

The 6 Hats as knowledge structure.

I run this vault across multiple ventures and client projects, and the hardest part is context-switching — jumping from a product decision to a client email to a content session in the same day. Each jump costs cognitive load.

Edward de Bono's Six Thinking Hats (1985) offers the fix: instead of trying to think about everything at once, you deliberately put on one hat at a time and work exclusively in that mode. I adapted it for the vault — the six knowledge domains map directly to six hats, so when I sit down in Growth mode I navigate to that domain and find only Growth knowledge. The separation the hats enforce in time, the vault enforces in space.

Cross-venture knowledge stays universal — a concept page on RAG doesn't belong to one product, it belongs to the Product hat. Venture-specific applications get tagged so they're findable, but the underlying idea stays general. It's the same framework behind LexisOps, my task-management product, so a concept page here lives in the same domain as the hat in the app.

"The separation the hats enforce in time, the vault enforces in space."

Hat Domain What you think about
Product 💻 Product & Engineering Building, shipping, architecture, technical decisions, QA
Brand 🎨 Brand & Identity Visual identity, tone of voice, positioning, design system
Content ✍️ Content & Marketing Writing, scripts, SEO, campaigns, social, newsletters
Growth 📈 Business & Growth Revenue, sales, pricing, metrics, partnerships, distribution
CS 🤝 Customer Success & Experience Client relations, user feedback, onboarding, retention
Ops ⚙️ Operations & Admin Processes, workflows, tools, admin, legal, finance

Three page types, enforced by the folder structure.

Zettelkasten's core insight applied explicitly: literature notes are the input layer, concept pages are the output layer — and entities are a third kind of page that accumulates facts about the things you work with. The directories enforce the epistemology.

Input Layer
Literature Notes
Atlas 🧠/1-Literature Notes/

Source-bound. Capture what a source says — tied to a specific book, article, course, video, or SDK docs. Organised by topic, not by medium.

  • Created by the Ingest workflow automatically
  • Marked with a promotes: field once distilled
  • Never deleted — remains as the source record
  • A ## My Take section is yours alone — AI never touches it
Output Layer
Concept Pages
Atlas 🧠/[domain]/

Permanent notes. Capture what you know — synthesised, source-independent, in your own words. Standing alone without needing the source to make sense.

  • Created or enriched by the Promotion workflow
  • Filed in the domain that matches the relevant hat
  • Universal — cross-venture knowledge stays general
  • The pages that get queried and cited going forward
Reference Layer
Entities
Atlas 🧠/Entities/

First-class pages for the things you work with — tools, frameworks, companies, and people. Distinct from concepts: they accumulate facts over time rather than explaining an idea.

  • Three shelves: Tools & Frameworks, Companies & Products, People
  • Grow incrementally as you learn more about each one
  • Linked from the concepts and notes that mention them

On top of these sit the navigation pages — not notes, but maps. index.md is the LLM's flat master map, read first on every session; Maps of Content are the human's curated "start here" pages for the graph view. Two readers, two maps.

Karpathy described the pattern. These are my additions.

Four specific things I built to make the LLM-Wiki pattern work for sustained, multi-venture use.

✍️
## My Take
Every literature note and concept page has a ## My Take section at the bottom — questions I'm sitting with, where I disagree with the source, how a concept applies to a specific venture, what I still need to figure out.

Claude Code writes everything above the line. ## My Take is mine alone. It never touches this section, even when updating an existing note. My perspective survives every AI edit.
📐
Two-layer note model
I made explicit what Zettelkasten implies: literature notes are the input layer (what a source says), concept pages are the output layer (what I know). This distinction is enforced in the vault structure and in CLAUDE.md.

No concept page should live in 1-Literature Notes/. No source-bound note should live in a domain folder. The directories enforce the epistemology.
⬆️
Promotion Workflow
In a traditional Zettelkasten, promotion (turning a literature note into a permanent note) is a manual, effortful process. I automated the pattern: ask Claude Code to "promote" a literature note, and it reads the note, identifies 1–N universal concepts worth standing alone, creates or enriches the corresponding concept pages, and marks the note with a promotes: frontmatter field.
🔄
Log Rotation
A small thing that signals something important about the design philosophy. Every operation is logged in log.md. When the log exceeds 30 entries, the oldest are rotated to Archive/archived-logs.md in one step.

The active log stays fast to scan; the full history is never lost. A system designed to be used for years needs to think about what happens when it fills up.

What's included in the template.

📋
CLAUDE.md — Operations Schema
6 workflows with concrete steps and grep commands. The contract between you and the AI operator. Personalise it to your projects in ~20 minutes.
📄
9 Note Templates
Literature note, fleeting note, permanent note, snippet, project CLAUDE.md, 3 output types, and Zotero integration.
🧪
6 Worked Examples
Step-by-step ingest walkthroughs covering PDFs, books, SDK/framework docs, certifications, online courses, and videos.
🏗️
Full Scaffold
All 6 domain folders, Entities sub-folders, 0-Map of Content (Topics/Books/Videos), two-stream Inbox (Fleeting Notes/ + Clippings/), Archive, Assets, Outputs — empty and ready.
🎯
Demo Content (removable)
A self-contained Deep Work walkthrough: 1 lit note → 1 concept page → 1 entity → 1 MOC + log entry. All marked (demo) for one-command removal.
⚙️
Repo Hygiene
MIT license, .gitignore, GitHub issue + PR templates, CODE_OF_CONDUCT.md, minimal .obsidian/ pre-config with Dataview + Templater.
Vault Structure
├── CLAUDE.md
├── index.md
├── log.md
├── SETUP.md
├── 00-Inbox/
  ├── Fleeting Notes/
  └── Clippings/
├── Atlas 🧠/
  ├── 0-Map of Content/
  ├── 1-Literature Notes/
  ├── Entities/
  ├── 💻 Product & Engineering/
  ├── 🎨 Brand & Identity/
  ├── ✍️ Content & Marketing/
  ├── 📈 Business & Growth/
  ├── 🤝 Customer Success/
  └── ⚙️ Operations & Admin/
├── Outputs/
├── Assets/
└── Archive/

Remove demo content in one command:
find . -name "*(demo).md" -delete

Optional — Power Up with Skills

Out of the box, everything runs on plain-English commands — zero setup. If you want repeatable shortcuts, you can add Claude Code skills (slash commands) in .claude/skills/ — each is just a folder with a SKILL.md carrying name and description frontmatter. That turns "process the inbox" into /wiki-ingest, "lint the vault" into /wiki-lint, "promote this note" into /promote.

The template ships without them on purpose — plain English is friendlier for getting started. Skills are a power-user layer you add once the workflow is second nature.

What this system is — and isn't.

Real costs

This system has real setup cost. Writing a good CLAUDE.md takes time. The first few ingests feel mechanical. The vault doesn't feel intelligent until there's enough in it to synthesise across.

This is not a productivity hack. It's an infrastructure investment. You're building something that pays off over years, not days.

Real payoff

The compounding effect is real. A passive note system degrades over time — old notes become stale, orphaned, irrelevant. An LLM-Wiki actively maintains itself.

The lint workflow catches drift before it becomes abandonment. The promotion workflow turns accumulated literature notes into usable knowledge before they go stale.

If you're a solo founder or independent builder who reads constantly, thinks in writing, and has struggled to turn your notes into something that actually serves your work — this is the system I'd build again.

"Most note systems are passive. The LLM-Wiki pattern makes your knowledge base an active participant — an AI that reads, writes, and maintains it alongside you. Combined with Zettelkasten's note architecture and Tiago Forte's Second Brain philosophy, it becomes something qualitatively different: a knowledge system that compounds."

— From the LLM-Wiki blog post

Clone it. Open it in Obsidian. Run claude.

The template includes the full CLAUDE.md schema, all 9 note templates, 6 worked examples, and demo content to show what the system produces. Follow SETUP.md to personalise it in about 20 minutes.

# 1. Clone the template
git clone https://github.com/tyrozz/llm-wiki

# 2. Install Claude Code
npm install -g @anthropic-ai/claude-code

# 3. Open in Obsidian, then run in terminal
claude
Founder OS — now in pre-order

LLM-Wiki is the open engine. Founder OS is the first paid vertical built on top of it — a calibrated decision journal for solo founders. It captures every real decision you make, reconciles it against what actually happened, and surfaces the patterns you keep repeating across ventures.

See Founder OS → Pre-order · $39 founding price · Ships June 2026