Personal OS Vault — Architecture Kit

A ready-to-open Obsidian vault that turns a folder of notes into a personal operating system an AI assistant runs for you. You talk to one orchestrator agent; it loads the right context, routes the work to the right place, and keeps your records up to date.

This kit gives you the architecture — the structure, the routing brain, and the blueprints to build the agent layer. It contains no personal data. You fill it with yours.

What you're looking at

vault-architecture-kit/
├── README.md                  ← you are here
├── 00-START-HERE.md           ← do this first (5-minute quickstart)
├── CLAUDE.md                  ← master context the agent reads first  (FILL IN)
│
├── _system/                   ← the routing brain + build kit
│   ├── AGENT-CONTEXT.md        ← maps task types → which files to load
│   ├── agent-roster.md         ← the map of your agents (3 tiers)
│   ├── inbox.md                ← dump unsorted notes here; agent processes them
│   ├── installed-capabilities.md ← registry of plugins/skills/tools the agent can use
│   ├── swarming-protocol.md    ← when the agent fans work out in parallel
│   ├── orchestrator-node.md    ← optional graph "breadcrumb" note
│   └── blueprints/             ← HOW THE AGENT LAYER GETS BUILT
│       ├── 00-BUILD-THE-AGENT-LAYER.md   ← the master setup prompt
│       ├── orchestrator-blueprint.md
│       ├── manager-layer-blueprint.md
│       └── specialist-blueprint.md
│
├── Personal/                  ← you: about-me, goals, logs, health, performance
├── Work/                      ← your job: context, deliverables, meetings, SOPs
├── Projects/                  ← side work / clients: context, log, research
│
├── .obsidian/                 ← cleaned Obsidian config (open as a vault)
└── .claude/agents/            ← empty; the setup prompt fills this

Personal / Work / Projects are placeholder domains. Rename them, delete the ones you don't need, or add your own. The architecture doesn't care what the domains are called — only that each has a *-context.md file and its own logs.

The architecture in three ideas

One front door. You only ever talk to the orchestrator. It assesses each request, loads only the relevant files, and routes the work. You never memorize which sub-agent does what.
A routing brain. AGENT-CONTEXT.md maps task types to the exact files to load for them. The agent stays lean — 3 files, not 30 — and gets faster and more accurate as a result.
Three tiers of agents. Orchestrator → Domain Managers → Specialists. You build these from the blueprints, tailored to your own work. Start small (a few specialists in your top domain) and grow on demand.

Plus two habits that compound: an inbox the agent clears at the start of every session, and living logs (decisions, deliverables, weekly) the agent updates as work happens.

Setup in three moves

Open it in Obsidian. Download Obsidian (free, obsidian.md) → "Open folder as vault" → pick this folder. The graph, backlinks, and templates work immediately.
Fill in the template files. Start with 00-START-HERE.md; it lists exactly which files to edit and in what order. ~10 focused minutes.
Build the agent layer. Open the vault in Claude and paste the setup prompt from _system/blueprints/00-BUILD-THE-AGENT-LAYER.md. It interviews you briefly and writes your agents.

Then: summon your orchestrator by name and hand it work.

Requirements

Obsidian — to navigate and edit the vault (free). Not strictly required, but the wikilinks and graph make it click.
An AI assistant that can read this folder's files — Claude Code or Cowork read the .claude/agents/ files automatically. Any file-reading assistant can follow CLAUDE.md + AGENT-CONTEXT.md manually.

No code, no servers, no accounts beyond your AI assistant.

Where this came from

This is the extracted, de-personalized architecture of a vault I actually run daily — one orchestrator, six domain managers, and a roster of specialists handling work across a job, a side business, and school. The structure here is the part that generalizes. The contents were mine and aren't included.

A few things it turns out matter more than they look:

The routing brain earns its keep immediately. Loading 3 relevant files beats loading 30 — on speed, cost, and answer quality. Most "my AI assistant is bad at this" problems are context-selection problems.
Managers exist to keep the roster navigable. One orchestrator picking from 100 specialists routes worse than one picking from six managers who each know their own ten.
The logs are the point. A vault the agent only reads goes stale in a month. A vault the agent writes back to — decisions, weekly capture, project status — compounds.

License

MIT — see LICENSE. Fork it, rename the domains, make it yours.