How to Use AI in Obsidian Without Losing Your Vault

To use AI in an Obsidian vault safely, decide what AI is allowed to do before you install anything. Write two files in your vault root — a USER.md that tells every AI tool who you are, and an AI-Policy.md that names what it can and cannot touch. Then work through four zones in order: let AI own the janitorial work (tags, YAML, folder structure), let it assist with summaries and transcripts, use it as an interrogator of your ideas — and keep your original insight notes AI-free. This guide covers the full framework, the setup, and the exact files to install.

There is a thread happening in the LYT community right now that I have not been able to stop thinking about. It started as a question about agentic AI and Obsidian. It turned into something more honest: a community of careful thinkers trying to figure out how to let AI in — without losing what makes their vault theirs.

One practitioner put it most directly: “AI is chatty. PKM is mind-work. You cannot create gold out of thin air. And it is not personal.”

That tension is real. And it is the one nobody is resolving — not the tutorials, not the vibe-coding crowd, not the people dropping OpenClaw setup guides. They are answering the question how do I install this when the actual question is should I, and if so, where does it belong?

This guide is my attempt to answer that question properly. Not as a technologist who wants you to automate everything. As a practitioner who has been building knowledge systems for seven years and has spent the last two living inside Claude Code, Obsidian CLI, and agentic workflows — and has strong opinions about what is worth doing and what is not.

We will cover the philosophy first. Then the practical setup. By the end, you will have a working framework, eight files you can install directly, and a clear sense of what belongs in your vault and what does not.

01 — The Core Problem

The Question Nobody Is Actually Asking

Everyone in the AI + Obsidian space is answering setup questions. What plugins to install. How to connect Claude Code. What OpenClaw’s architecture looks like. These are important. But they are the second question.

The first question — the one that determines whether any of this is worth doing — is: what do you want AI to do in your vault, and what do you want to protect it from?

Without a clear answer to that, you are not building a system. You are collecting tools. And the history of PKM is littered with people who collected tools so beautifully that they never produced anything with them. I call it the Collector’s Trap: the vault grows, the system becomes more elaborate, and the one thing that was supposed to happen — original thinking reaching the world — never does.

AI does not solve the Collector’s Trap. Improperly used, it deepens it. You now have a very sophisticated machine to produce notes that look like yours but are not. The vault fills faster. The feeling of productivity intensifies. The actual output — your thinking, your voice, your compound insight — stays buried.

"The skill that makes AI useful is not prompting. It is knowing what you want before you open the chat window."

This is what I mean by person first, technology second. It is not a platitude. It is a design constraint. You have to define the boundaries of the system before you build it, or the system will define itself — and it will optimize for output, not for you.

The four-zone framework below is how I have been thinking about this. Use it as a decision-making filter before you install anything. When you are uncertain about a new AI tool or workflow, ask which zone it belongs in — and whether you have built the earlier zones first.

This guide is the setup: the files to write, the tools to install, the guardrails to put up first. If you want the zones themselves worked through in full, with what belongs in each one and the mistakes that show up in every zone, that lives in The Four-Zone Method.

Some practitioners are solidly in Zone 1 — using Claude Code for maintenance, underwhelmed by generated text, satisfied with the ROI. That is a completely valid and valuable position. Others span Zones 1 through 3 with local models. Some are deep in Zone 3 with 25+ OpenClaw agents. Others are still in Zone 0 — deciding. All of these are correct positions for different people at different stages.

Before you read another word of setup instructions, know which zone you are entering and why. If you cannot answer that question, start with the two files in the next section.

02 — The Foundation

The One File That Changes Everything

Every AI tool that touches your vault — Claude Code, OpenClaw, Obsidian Copilot, local Ollama models — shares a common failure mode: they do not know who you are. They know your folder structure at best. They do not know your thinking style, your goals, what you protect, what you are building toward.

The solution is not better prompting. Prompting is a session-level fix for a structural problem. The structural fix is a single file that travels with you across every tool, every session, every upgrade.

Think of it this way. Right now, every time you open a new AI session, you are spending the first five exchanges re-establishing who you are. What you are working on. How you think. What you do not want it to do. That is cognitive overhead you pay on every single session, forever — unless you write it down once.

Here is what a minimal but effective USER.md looks like:

# Who I Am (for AI tools)
# Read this before you do anything in this vault.
## My PKM Intent
My vault exists to: ___________
# Complete this sentence before anything else.
# It is the most important line in this file.
## How I Think

I process ideas by writing first, then linking
I prefer questions back over generated content
My notes are personal — do not generate text that sounds like me
When uncertain, do nothing and report the uncertainty

## My Vault Structure

/Inbox     — unprocessed captures
/Notes     — processed permanent notes
/Projects  — active work (PARA)
/Archive   — completed or inactive

## AI Permissions
✓ Can touch:  tags, YAML, folder org, template application
✓ Can assist: summaries, transcripts, source notes
✗ Cannot touch: /Notes/Evergreen, /Fleeting, published work
✗ Cannot write in my voice or generate original insight
## AI Tell
Any note touched by AI carries the tag: #ai-assisted
Any AI-generated text opens with: > 🤖 [AI-generated]

The second file to build before anything else is your AI-Policy.md — a shorter, sharper sovereignty document. Where USER.md describes who you are, AI-Policy.md describes your rules. It names what AI can touch, what it cannot touch, and what happens when it is uncertain. Put both files in your vault root. They work together.

Some practitioners already do this instinctively — noting that AI always adds a tag when it touches a note. The USER.md and AI-Policy.md formalize that instinct into something portable across every tool you will ever use. The five minutes these files take to write will save you hours of correcting AI behavior downstream.

03 — The Practical Setup

Claude Code Inside Obsidian in 30 Minutes

This is the setup that several practitioners in the LYT community are running variants of. It requires no coding knowledge. The Terminal plugin made it dramatically simpler in late 2025 — what used to take configuration now takes about three clicks.

There are three main paths depending on your comfort level. I will walk through all three from lowest friction to most powerful. Choose the one that fits where you are today — you can always move up later.

Path A — The Terminal Plugin (Recommended for Most People)

1. 01
   Install Claude Code. Requires Node.js 18+. Check with node --version in your terminal. Then: npm install -g @anthropic-ai/claude-code. Takes about two minutes. A Claude Pro or Max subscription works directly — run claude setup-token once and you are authenticated.
2. 02
   Install the Terminal plugin in Obsidian. Settings → Community Plugins → Browse → search "Terminal" by polyipseity. Install and Enable. This embeds a full terminal directly in your vault sidebar — no window switching, no context loss.
3. 03
   Add the Obsidian Skills folder. Download the .claude folder from the obsidian-skills GitHub repo and drop it in your vault root. This teaches Claude Code exactly how Obsidian files work — the difference between Claude guessing and Claude knowing your vault structure.
4. 04
   Drop in USER.md and AI-Policy.md. Build these first (Section 02). Put them in your vault root. When Claude Code starts a session, it reads these files automatically. Every output is calibrated to your permissions before you type a word.
5. 05
   Open terminal, navigate, launch. Click the Terminal icon in Obsidian. cd to your vault folder. Type claude. You are in. Run your first three commands inside Zone 1: add YAML to your Inbox, tag untagged notes, apply templates. Safe, reversible, immediately useful.

Path B — The Sidebar Plugin (Zero Terminal)

If you would rather not use a terminal at all, the obsidian-claude-code community plugin embeds Claude as a persistent sidebar assistant. You get the same read, write, and edit access, the same skill loading, plus @[[filename]] autocomplete to reference any note directly by name. Install via community plugins or GitHub.

Best for: Writers and thinkers who want AI capabilities without touching a terminal. Slightly less powerful than Path A for automation — excellent for Zone 2 and Zone 3 work done conversationally inside Obsidian.

Path C — Local AI with Ollama (Zero Cloud, Maximum Privacy)

This is what many LYT practitioners are asking for: AI in your vault without a single byte leaving your machine. Ollama makes this straightforward on Mac.

Privacy note: Local AI means your notes stay completely on your device. The tradeoff: local models (phi4, llama3.2, mistral) are slower and less capable than Claude for complex synthesis and Zone 3 sparring. They are excellent for Zone 1 and most Zone 2 work. For deep thinking partner sessions, cloud models still win significantly.

Install Ollama from ollama.ai, pull a model (ollama pull phi4 for a balanced setup, or ollama pull llama3.2:3b for speed on older hardware), then install the AI Providers and Local GPT plugins in Obsidian and point them to http://localhost:11434. You are running fully local AI on your notes.

04 — Keeping Your Voice

The Thinking Partner Skill

This is the section I care about most. And it is the one missing from every other Claude Code + Obsidian guide.

The default behavior of AI is to generate. You give it a rough idea, it gives you a polished version. The problem: polished is not the same as yours. And in PKM, if it is not yours, it does not compound.

Here is what I mean by compound. Your vault is not a storage system. It is a thinking system. Every note you write in your own language, with your own connections, builds up a layer of conceptual residue that makes the next note sharper, faster to write, more densely linked. This is the actual ROI of PKM — not retrieval efficiency but cognitive compounding. When AI writes your notes for you, the compounding stops. You get outputs that look productive. The vault fills. The thinking does not grow.

"You do not want AI to think for you. You want AI to help you think more than you would have alone."

The Thinking Partner Skill is a Claude Code skill file that changes the default behavior. Instead of generating, it interrogates. Instead of writing, it questions. Here is the exact file — save it as vault/.claude/skills/thinking-partner.md:

# Thinking Partner Skill
## When to Activate
When user pastes raw notes, fleeting thoughts, or rough ideas
and asks for help developing, expanding, or thinking through them.
## Your Role
You are NOT a writer. You are an interrogator of ideas.
Never rewrite the user’s text.
Never generate new content from scratch.
Your only job: ask the questions that help the user
find what they already know but have not said yet.
## The Process

Read the raw note completely — do not skim
Identify the single most interesting tension or claim
Ask: “What do you mean when you say [X]?”
After their answer: “What is the strongest objection to that?”
After objection: “What does this tell you about [original idea]?”
Repeat until user says: “Oh — that is what I actually meant.”

## Hard Rules
NEVER suggest vague directions to “explore”
NEVER generate a bullet list as output
NEVER summarize what the user already wrote
NEVER say "great insight!" or use affirmations
## Close Every Session With
“In one sentence — what did you figure out today
that you did not know when we started?”

That closing question is not decorative. It forces you to crystallize the session’s output into a single, ownable sentence. That sentence becomes your permanent note. Your voice. Your compound interest.

This is Zone 3 in practice. The AI is not writing for you — it is revealing what you were trying to say. Every serious PKM practitioner should install this skill before they do any generative AI work in their vault.

05 — Agentic AI

What OpenClaw Is and Whether You Need It

OpenClaw is worth understanding because it represents where the entire space is heading — but it is not where most people should start.

The short version: OpenClaw is a self-hosted agent runtime that runs on your machine (or a Raspberry Pi, or a Mac Mini), connects to your messaging apps, and acts on your behalf without being prompted. It has persistent memory, a heartbeat that checks in every 30 minutes, and a skill ecosystem that grew to 700+ community-built extensions. It went from zero to 214,000 GitHub stars in under four months — the fastest growth of any open-source project in that period.

Some practitioners run 25+ agents, each handling one micro-task, cycling through a REFI loop (Routines, Emergencies, Flows, Ideas) — sophisticated and genuinely impressive. It takes months to build and requires comfort with Node.js, workspace files, and troubleshooting agent behavior. It is not a Day 1 setup. It is not a Day 30 setup for most people.

If you want to experiment with OpenClaw and Obsidian today, start with one safe, read-only agent and a minimal ~/.openclaw/HEARTBEAT.md:

# Heartbeat Instructions
# Runs every 30 minutes while active
Check my Obsidian Inbox folder.
If there are unprocessed notes (no YAML front matter),
generate a 2-sentence summary and add it as a “summary” property.
Do NOT move files. Do NOT delete anything.
Do NOT write in any note that has the tag #protected.
Log what you did in Agent-Log.md with a timestamp.
If uncertain about any action, do nothing and log the uncertainty.

This is low-risk and immediately useful. It processes your inbox while you sleep and never touches anything you have marked as yours. That is Zone 1 agentic AI — the right entry point. Let it run for two weeks before you add a second agent. Understand what it actually does before you scale it.

06 — The Eight Files

Install These Eight Files in Order

Eight weeks of deliberate practice produces eight files. Here they are in the order that makes each one useful before the next — and in the order that builds the safest possible foundation.

1. 01
   AI-Policy.md — Your sovereignty document. What AI can and cannot touch. The explicit rules that every tool in your vault must follow. Build this before any installation — it is the foundation everything else rests on.
2. 02
   USER.md — Your PKM identity file. Works with Claude Code, OpenClaw, Obsidian Copilot, and local models simultaneously. Write it once, keep it updated, and every AI tool you ever use gets smarter about who you are.
3. 03
   AI-Commands.md — Thirty tested Claude Code commands for vault maintenance. Drop it in your vault as a reference. Stop reinventing the same prompts every session — the commands that work are the ones worth repeating.
4. 04
   Capture-Pipeline.md (skill) — A Claude Code skill for YouTube → transcript → processed note. A systematized, repeatable pattern that runs in under five minutes from raw input to vault-ready note.
5. 05
   Thinking-Partner.md (skill) — The interrogator skill above. Your insurance against AI taking over your voice. Install before you do any Zone 3 work. Use it every time you have a rough idea you are trying to develop into a permanent note.
6. 06
   HEARTBEAT.md (agent) — OpenClaw starter config. One agent, one task, read-only, logged. Your first agentic experiment with explicit safety rails built in. Run it for two weeks before adding anything else.
7. 07
   Vault-Review.md — A weekly AI-generated health check. Surfaces where AI crept past your policy, notes without proper tagging, and Inbox items that have been sitting too long. Your maintenance and accountability layer.
8. 08
   My-AI-Stack.md — Your personal stack doc. Which tool for which zone, why you made each choice, what you would tell someone starting from scratch. Teaching it to someone else compounds what you know. Write it last — you will know what to say.

07 — The Honest Assessment

What Is Worth Your Time and What Is Not

I have been inside AI tooling deeply enough to tell you directly: most of what you see online about AI + Obsidian is people showing you their setup, not their results. The setup is impressive. The results are often modest. The mistake is treating configuration as a proxy for capability.

Here is my honest assessment of where the real value is, by zone:

Zone 1 is underrated. Automating YAML, tags, template application, and folder structure with Claude Code or a simple agent will save a serious Obsidian user 2–4 hours per week, indefinitely. That is real — practitioners who start here are consistently the most satisfied with their AI integration. It is not glamorous but it compounds — every hour saved in maintenance is an hour recovered for Zone 4 thinking.

Zone 2 is hit or miss. AI synthesis of your vault depends entirely on whether your vault has good structure to synthesize from. If your notes are consistent and well-tagged, it is powerful. If they are messy, AI makes the mess bigger and faster. Fix Zone 1 completely before you try Zone 2.

Zone 3 is the most misunderstood. Most people use AI in Zone 3 the wrong way — they ask it to write, and then they edit. That is backwards. The right way is to use it as an interrogator and write yourself. The Thinking Partner Skill above is the entire game here. If you take only one thing from this guide, take that skill file.

Zone 4 needs no AI. Full stop. Your original insights, your fleeting notes, your evergreen notes, your published voice — these are yours. The entire point of the AI layer in Zones 1 through 3 is to protect your time and thinking so you have more capacity for Zone 4, not to replace it. If you ever catch AI writing your Zone 4 notes, stop everything and rebuild your policy.

"Build the zones in order. Do not skip to agents before you have cleaned your inbox. Do not run a heartbeat before you have written a USER.md. Depth before breadth."

The LYT community is one of the few places online where people hold this tension carefully. The thread that inspired this guide is better than 90% of what is on YouTube about AI + Obsidian — because the people there care about whether the output is theirs, not just whether it is impressive.

That is the foundation everything in this guide is built on. Use AI where it reduces friction and protects your thinking time. Protect everything that makes your vault irreplaceable. The system you trust is the one you actually use — and the system you trust is the one that still sounds like you.

Common Questions

Frequently Asked Questions

Is it safe to use AI inside my Obsidian vault?

Yes — with a policy in place. The risk is not that AI will corrupt your files; it is that AI will gradually replace your thinking if you operate without clear boundaries. Write AI-Policy.md and USER.md before installing anything. With those in place, AI becomes a maintenance assistant and thinking partner — not a ghostwriter.

What is the four-zone PKM framework?

Zone 1: AI owns janitorial work (YAML, tags, templates). Zone 2: AI assists with synthesis and transcripts — you review. Zone 3: AI sparring partner — it asks questions, you write. Zone 4: Yours alone — original insights, published voice. Build in order. Never skip ahead.

What is a USER.md file?

A persistent identity file in your vault root that every AI tool reads before doing anything. It contains your PKM intent, how you think, your vault structure, and explicit AI permissions. You write it once, update it rarely, and every AI session starts with context about who you are.

Do I need OpenClaw to use AI in Obsidian?

No. Start with Claude Code in the terminal (Path A) or the sidebar plugin (Path B). Use OpenClaw only after you have completed Zones 1 and 2 and want autonomous agents running continuously. The minimal safe entry is one read-only agent with a simple HEARTBEAT.md — not 25 agents on Day 1.

How do I stop AI from writing in my voice?

Three things: write a USER.md with an explicit “cannot write in my voice” rule, install the Thinking Partner Skill so Claude interrogates instead of generates, and protect your Zone 4 folders with a #protected tag in your AI-Policy. The Thinking Partner Skill is the most important — it changes Claude’s behavior from output to interrogation by design.

What eight files should I install first?

In order: AI-Policy.md, USER.md, AI-Commands.md, Capture-Pipeline.md, Thinking-Partner.md, HEARTBEAT.md, Vault-Review.md, My-AI-Stack.md. Each one depends on the earlier ones — build them in sequence and you will have a complete, principled AI layer in your vault within two weeks.