AI-Native Obsidian Vault Template

A clonable starting point for an Obsidian vault designed to be used by an AI agent, not just by you. Object-OS folders + a two-namespace tag system + a _system/ AI memory layer + 5 specialized subagents + slash commands.

Read the full architecture breakdown before you clone, or jump in.

What this is

A skeleton vault for Obsidian wired up to work with Claude Code (or any Claude Code-compatible CLI). It gives you:

A folder structure (Object-OS) that doesn't fight an AI agent
A _system/ directory that gives the AI persistent memory across sessions
5 specialized subagents (Memory Keeper, Inbox Triager, Knowledge Compiler, Vault Librarian, Research Analyst)
Slash commands for daily/weekly rituals (/triage, /weekly-review, /morning-brief, /health-check, /search, ...)
3 path-scoped rules (daily-note format, frontmatter normalization, research protocol)
A canonical tag system — 2 live namespaces, max 4 tags per note, and a written reason for why it's not 5
Note templates (Daily, Project, Person, Company, Concept, Content)
A handful of anonymized example notes showing a small, connected graph

Requirements

Obsidian (free)
Claude Code CLI (or a Claude Code-compatible client) — required for subagents and slash commands
(Optional but recommended) MCP servers for richer research:
- Readwise MCP — search highlights from books, articles, podcasts
- Context7 MCP — up-to-date library/framework docs
- A personal bookmarks MCP (e.g. Raindrop) if you keep a curated bookmark library

Setup (~10 minutes)

1. Clone into your Obsidian vaults directory

cd ~/Documents/Obsidian   # or wherever you keep your vaults
git clone https://github.com/mariomile/ai-native-obsidian-vault my-new-vault
cd my-new-vault

2. Open in Obsidian

File → Open vault → select my-new-vault. Trust the workspace when prompted.

3. Customize the system files

Open these files and replace the {{placeholders}}:

CLAUDE.md — {{user}}, {{vault_name}}
_system/vault-context.md — fill in who you are, active projects, current state
_system/memory/preferences/preferences.md — fill in your communication style
_system/_tags.md — add or rename #domain/* tags to match your life
LICENSE — replace {{user}} with your name

You don't have to fill everything immediately — the AI will tolerate {{placeholders}} and prompt you when it needs them.

4. Run Claude Code from the vault root

claude

5. Run the health check

/health-check

This validates that everything is structurally OK — folders exist, frontmatter is canonical, no broken references.

6. Create your first daily note

Use Obsidian's Daily Note core plugin (configure it to use Resources/Templates/Daily.md and Journal/Daily/ with DD-MM-YYYY filename).

Start capturing. Run /triage weekly. Run /weekly-review Sundays. Run /morning-brief on Mondays.

That's the loop.

The mental model (90-second version)

Folders define what a file IS. Tags define what it's ABOUT.

Folder	What lives here
`_system/`	The AI's memory layer — context, rules, decisions, indices, templates
`_inbox/`	Catch-all for untriaged stuff. Should trend toward zero.
`Input/`	Material you didn't author — Readwise, Books, Articles, Courses
`Knowledge/`	Permanent knowledge — organized by domain (`Product/`, `AI and Agents/`, ...)
`CRM/`	People and Companies you deal with
`Active/`	Projects in progress + Goals
`Output/`	Deliverables — `_artifacts/` for agent-generated HTML; everything else stays in its project, tagged `#type/output`
`Journal/`	Daily / Weekly / Monthly / Yearly notes
`Resources/`	Reusable tools — Templates, attachments
`.archive/`	Inactive items (preferred over deletion)
`.claude/`	Claude Code configuration — rules, agents, commands

One declared exception: Knowledge/ is organized by domain, not just tagged with one — because there, domain is intrinsic and permanent (a note about pricing strategy stays about pricing forever). Everywhere else, domain is a tag, resolved at query time.

Tags use 2 canonical namespaces. Max 4 per note.

tags:
  - type/concept         # what the file IS
  - domain/product        # what it's ABOUT — cross-folder
status: active            # PROPERTY, not a tag — only for project/experiment/course

Why only 2 (not 5, like most PKM templates ship): #status/*, #stage/*, #scope/* all sounded useful and all turned out redundant with the folder path or with a single frontmatter property — each one is documented as an anti-pattern, with the reason, in _system/_tags.md. Read that file before you're tempted to add a third namespace; it's the fastest way to avoid rebuilding a mistake that's already been made and undone once.

The 5 subagents

Defined in .claude/agents/. The main AI dispatches to these via the Agent tool.

Subagent	Owns
🧠 Memory Keeper	Learnings, decisions, rule promotion, quality gate
📥 Inbox Triager	Route captures into Object-OS folders (never moves without OK)
📚 Knowledge Compiler	Promote notes into `Knowledge/`; create MOCs
🏛️ Vault Librarian	Indices, broken links, staleness, tag audit
🔍 Research Analyst	Multi-source research with citations

Full guide: _system/agents-guide.md.

Slash commands

Defined in .claude/commands/. Type /command-name in Claude Code.

Daily / weekly rituals

/morning-brief — daily boot: yesterday, today, blockers
/standup — quick status of active projects
/weekly-review — pre-filled weekly digest
/letswork — pull next-actionable tasks

Capture & triage

/triage — process _inbox/ and daily-note captures
/meeting — start a structured meeting note

Research

/search — multi-source deep search with citations
/crm-lookup — surface everything the vault knows about a person/company

Maintenance

/health-check — full vault audit
/find-orphans — surface unlinked, under-tagged notes
/tag-audit — audit tags vs. canonical namespace
/refactor-note — promote a note into Knowledge/
/archive — move a completed project to .archive/
/update-context — refresh a project's context.md

Pipeline

/enrich-daily-notes — wikilink + summary enrichment for recent dailies
/link-readwise — graph-additive linking of Readwise highlights
/new-project — scaffold a fresh project
/export — export to PDF / DOCX / HTML

The 8 Sacred Rules

Loaded at every session boot from CLAUDE.md:

Respect .claude/ boundaries (CLI-managed vs user-authored)
Sync-owned folders (e.g. Input/Readwise/) are graph-additive only — never touch synced content
Never create files or folders at vault root — and a folder that appears there is a signal of intrusion, not a mess to tidy
Always tag new notes (1–4 canonical tags)
Ask before overwriting (unless explicitly told to edit)
Normalize frontmatter on touch
Use wikilinks for vault-internal references
Search before asking (don't interrupt for googlable things)

Full rationale: CLAUDE.md.

Customizing for your workflow

Add a domain

Edit _system/_tags.md under #domain/*. The point is to have a finite, stable, orthogonal set — don't add a domain you'll only ever use twice.

Add a type

Edit _system/_tags.md and _system/object-schema.md. Define the type's required and optional frontmatter properties.

Add a rule

Create .claude/rules/{name}.md and reference it from CLAUDE.md. Use path-scoping in the rule frontmatter if it should only apply to specific files.

Add a subagent

Create .claude/agents/{name}.md with proper frontmatter:

---
name: subagent-name
description: When to invoke this subagent.
tools: All tools
model: sonnet # or haiku, opus
---

Add it to _system/agents-guide.md and to the table in CLAUDE.md.

Add a slash command

Create .claude/commands/{name}.md. The file body is the instruction the AI executes when you type /name.

Thinking about a company/team vault instead of personal?

Same primitives, different governance — see the "Personal brain vs company brain" section of the architecture breakdown. Short version: one extra structural requirement — living documents (specs, playbooks, decision logs) need a declared owner and a review date in frontmatter. Everything else in this template stays the same.

What's intentionally NOT included

Your notes. This is a skeleton, not a content library. A handful of anonymized examples ship to show formatting and the graph.
Personal MCP tokens / API keys. settings.example.json is the template; copy to settings.local.json and add your own secrets (gitignored).
Specific community plugin configs. Configure Obsidian plugins (Daily Notes, Tasks, Bases, Templater) to your taste.
A graph theme / CSS snippets. Use any Obsidian theme. The structure is theme-agnostic.

Credits & inspirations

This system stands on the work of many people:

Tiago Forte — Building a Second Brain, PARA, CODE method
Sönke Ahrens — How to Take Smart Notes, Zettelkasten / fleeting → literature → evergreen
Andrej Karpathy — the "LLM as compiler" framing
Steph Ango (kepano) — Obsidian skills, folders-as-types philosophy
Anthropic — Claude Code, subagents protocol, skills protocol
The Obsidian community

License

MIT. Fork it, remix it, ship it. If you build something better, I'd love to see it — open an issue or @ me with what you made.

Contributing / feedback

Found a bug in the skeleton? Want to propose a new subagent or command pattern? Open an issue.

This template will evolve. If you want to be notified, watch the repo or subscribe to the newsletter.