I have 971 files in my knowledge base. It’s organized using a system borrowed from China’s Tang Dynasty bureaucracy.
I’m not joking.
The Problem With Folders
Most people use folders like this: drop a note in, forget about it, never see it again. Three months later, Obsidian looks like a warehouse full of unlabeled boxes.
The deeper problem: a single idea often belongs in multiple categories. A note about “Redis caching for API gateways” is simultaneously about databases, API design, and cost optimization. Folders force you to pick one.
Three Departments, Six Ministries
The Tang Dynasty (618-907 AD) ran the most sophisticated bureaucracy of its era with a “Three Departments, Six Ministries” structure. I stole it.
Three Departments (Decision Layer):
- Central Secretariat — Input planning. What should I learn this month? What’s worth deep-diving? What should I NOT do?
- Chancellery — Quality review. Is this note good enough? Are there logical gaps? Can it be published?
- Department of State Affairs — Execution. Who’s doing what? When is it due? How do we track it?
Six Ministries (Knowledge Layer):
- Personnel — People. Interview prep, character profiles, personality analysis
- Revenue — Finance. Economics notes, investment research, monetization ideas
- Rites — Expression. Writing techniques, content strategy, blog curation
- War — Competition. Tech trends, competitive analysis, information arbitrage
- Justice — Risk. Server security, contract law, backup recovery
- Works — Building. Code projects, server deployment, tooling
A note about “FastAPI rate limiting” gets tagged with both Works (for the code) and Justice (for the security angle). No folder conflicts, no duplication.
The AI Assistants
I trained two AI agents to help run this system:
- The Steward — Auto-classifies new notes, tags them, maintains directory structure. Pure execution.
- The Advisor — Watches for stale content, suggests what to review, flags publishable notes. Strategic.
The key to making this work was writing detailed specifications. Not “help me organize” — but exact formats, structures, good output examples vs. bad output examples. Once the spec was precise enough, the AI output quality jumped an order of magnitude.
The Only Rule That Matters
Don’t aim for perfect archiving. Aim for perfect retrieval.
971 files isn’t impressive. Finding that Cloudflare security analysis from six months ago in 3 seconds — that’s impressive. And that’s what the system is for.