July 26, 2026

Claude Code and Obsidian: How to Turn Your Vault Into a Queryable Knowledge Base

Claude Code reads an Obsidian vault with zero setup. Here is how to synthesize, draft from, and audit thousands of markdown notes with an agent.

Yes, with no setup at all. An Obsidian vault is a folder of plain markdown files, and plain markdown on disk is exactly what Claude Code reads natively. You open the vault folder as your working folder and you are done. No plugin. No integration. No export step.

Which means the thing you have been calling a note archive has been a queryable knowledge base the entire time. It just never had a reader.

Most vaults are where notes go to die. Three thousand files. A gorgeous graph view. Four years of thinking you have never once retrieved on demand. The problem was never your note-taking system. The problem was that retrieval depended on you remembering what you wrote and where you filed it.

An agent does not have that limit. It reads all of it, on every question. If you are new to working this way, the broader system lives in Claude Code for non-developers. This post goes deep on the vault.

Can Claude Code read an Obsidian vault?

Yes, natively, because nothing proprietary sits in the way. Obsidian writes .md files into a normal folder on your drive. Open that folder and every note is readable text.

The features that make Obsidian feel like a database are also just text:

  • Wikilinks. [[Client Onboarding]] is literally those characters inside the file. The agent can search them, follow them, and build a picture of what points at what.
  • Tags. #pricing is text in the body, or a value in frontmatter. Searchable either way.
  • Properties. Obsidian properties are YAML frontmatter at the top of the file. Agents read YAML well, so status: draft becomes a real filter.
  • Folders. Directory structure, which the agent reads as directory structure.

The one real exception is Canvas. A .canvas file is JSON, not prose. The agent can parse it and tell you which notes a canvas references, but do not expect it to reason about the spatial layout the way your eye does.

What can you actually do with a vault an agent can read?

Four things, and each is a single prompt rather than a project: synthesis, drafting, checking your own prior thinking, and maintenance. The specific prompts matter more than the categories, so here they are.

How do I get an agent to synthesize notes I forgot I wrote?

Ask for the synthesis and demand citations back to source notes. The citation requirement is what separates a useful answer from a confident one.

Cross-vault synthesis prompt

Prompt
Search the vault for every note that touches pricing, packaging, or how I charge clients. Include notes where the topic is secondary rather than the main subject.

Then write a synthesis of my actual position on pricing, in my own words where possible.

Rules:

- Cite the source note filename after every claim, like this: (Pricing Ladder Rewrite.md)
- Where two notes disagree, say so and cite both. Do not resolve the contradiction for me.
- Flag anything I wrote more than a year ago that I have not revisited since.

What comes back is a few paragraphs of position with filenames trailing each claim, plus a short list of places you argued with yourself.

The two rules that make this beat the obvious version: including notes where the topic is secondary, and forbidding the agent from resolving contradictions. A default prompt returns tidy consensus. Your best thinking is usually sitting in the disagreement, inside a note you filed under something else entirely.

How do I turn meeting notes into a client deliverable?

Hand it the raw note and an actual example of the output format. A format example beats a format description every time.

Meeting notes to deliverable prompt

Prompt
Read Meetings/2026-07-14 Acme Kickoff.md.

Turn it into a client-facing scope summary using Templates/Scope Summary.md as the exact structure.

Rules:

- Every deliverable, date, and owner must come from the meeting note. If something is not in the note, put it under an OPEN QUESTIONS heading at the bottom instead of guessing.
- Strip my shorthand and internal comments out of the client version.
- Save it to Clients/Acme/ as a new file. Do not edit the meeting note.

The open questions rule is the one doing the work. Without it, the agent fills the gaps with plausible dates and invented owners, the document reads clean, and you send it.

How do I find what I already think about a topic before I write about it?

Run this before you draft, not after. It is the difference between publishing your third-best take and finding the one you already had.

Prior thinking prompt

Prompt
I am about to write about [TOPIC].

Before I start, search the vault for anything I have already written that touches this: half-finished drafts, meeting notes, book highlights, voice memo transcripts, daily notes.

Return:

1. My strongest existing angle, quoted directly, with the filename
2. Any claim I made before that I would now have to defend or retract
3. Three specific examples or stories from my own notes I could use
4. What I have clearly never written about here

Do not write the piece.

Item four is the one people leave out, and it is usually the most valuable line in the output. The hole in your own coverage is the post.

How do I audit my vault for orphans and dead notes?

Ask for a report, not a cleanup. The agent should never be the thing deciding what dies.

Vault audit prompt

Prompt
Audit this vault and write the results to _audit/vault-report.md. Change nothing else.

Report:

- Orphan notes: no inbound wikilinks and no outbound wikilinks
- Stubs: under 100 words with no links, grouped by what they seem to be about
- Near-duplicates: notes covering the same topic under different titles
- Contradictions: two notes stating opposing things, quoted side by side
- Tag sprawl: tags used fewer than three times, and which existing tag each could fold into

For each item, recommend keep, merge, or archive. Do not act on the recommendation.

Near-duplicates and contradictions are the valuable half. An orphan list is easy and mostly boring. Two of your own notes arguing with each other means your thinking moved and you never wrote down that it moved.

How do you keep the agent from wrecking your vault?

Duplicate the folder before any structural change, let sync finish before you start a session, and ban deletion in writing. The risk is not that the agent goes rogue. The risk is that "clean this up" is ambiguous, and it resolves the ambiguity differently than you would have.

Four rules cover almost everything:

  1. Copy the vault folder before any session that reorganizes, merges, or renames. A text vault is tiny. The copy takes seconds and it is the only backup that is genuinely instant to restore.
  2. Let sync settle first. Obsidian Sync, iCloud, and Dropbox each resolve conflicts on their own schedule. An agent writing files mid-sync is how you end up with conflicted copies scattered across folders.
  3. Turn on File Recovery. Obsidian ships a core plugin that keeps periodic local snapshots of note contents, which covers most single-file mistakes. Check its settings before you rely on it, since snapshot frequency and retention are configurable and the defaults are not generous.
  4. Ban deletion outright and route to quarantine instead. A _quarantine/ folder with a log file means nothing is ever gone, only moved.

Should you put a CLAUDE.md in your vault?

Yes, and it looks nothing like a project one. A code CLAUDE.md describes architecture, commands, and conventions for building software. A vault CLAUDE.md describes permissions, naming, and what to do when a file is ambiguous. The general pattern is covered in using CLAUDE.md as a business operating system. Here is the vault-shaped version.

Starter vault CLAUDE.md

Prompt
# Vault Rules

## What this vault is

Personal and business knowledge base. These notes are the source of truth for how I think. This is not a scratchpad.

## Never

- Never delete a file. Move it to _quarantine/ and log it in _quarantine/log.md with the reason.
- Never edit anything in Archive/ or Journal/. Those are read-only.
- Never rewrite a note's wikilinks without listing every change first and waiting for approval.
- Never create a note outside the folder I named.

## Naming

- Notes: Title Case, no dates in the filename except in Meetings/ and Journal/
- Meetings: YYYY-MM-DD Client Name Topic.md

## New note format

Frontmatter with type, status, created, and tags. Body opens with a one-sentence summary before the first heading.

## When uncertain

If it is unclear which folder a note belongs in, which tag applies, or whether two notes are duplicates, stop and ask. Do not guess. Do not create a new tag.

That last block is the one that saves you. Left to itself, an agent will invent a tag rather than admit ambiguity, and now #pricing-strategy is sitting next to #pricing forever.

This file also solves a smaller version of the problem in why AI agents forget. The agent starts every session with no memory of the last one. The vault rules file is the part that persists, so your conventions survive the reset even though the conversation does not.

What does this change about how you take notes?

Capture quality starts to matter more than organization quality. Folder trees and tag taxonomies were always a retrieval workaround, a system for helping future-you find one thing without rereading everything. The agent does retrieval now. So write more and file less.

In practice:

  • Stop deliberating over folders. Drop it in an inbox folder and move on. The decision was costing you more than it returned.
  • Stop pruning tags. Tag sprawl is a retrieval problem, and retrieval is handled.
  • Write longer captures. Voice memo transcripts and messy half-thoughts are assets now, because an agent can pull structure out of prose that would have been unusable to you six months later.
  • Write the why, not just the what. "Chose vendor B" is worth almost nothing. "Chose vendor B because A's contract terms were bad on data portability" is a note that answers a real question two years from now.

One honest caveat. This only pays off if there is something in the vault. An agent pointed at forty notes hands you forty notes back, rearranged. The value curve is steep, and it starts late.

An agent with no history behind it is a stranger. A vault is the fastest history you can hand one, because you have been writing it for years without meaning to. That gap between a stranger and something that already knows your context is the problem I am building Focus Pilot to close, in public and not finished.

Does this work with Notion, Apple Notes, or Roam?

It works with anything that stores plain text or markdown on disk. Everything else needs an export step first.

  • Obsidian. Native. Markdown files in a folder, no conversion, nothing to maintain.
  • Logseq. Native. Also markdown on disk.
  • Roam. Export required. Roam exports to Markdown, JSON, or EDN. The Markdown export lands close to an Obsidian-style vault.
  • Notion. Export required. Notion is a database with a page interface, not files on your drive. Its Markdown and CSV export produces a folder of markdown files plus attachments, which an agent reads fine.
  • Apple Notes. The hardest of the four. Notes live in a local database rather than as files, so getting them out means an export script or a third-party tool. Doable, not casual.

The tradeoff worth naming plainly: an exported knowledge base is a dead copy. It is accurate the day you export it and stale the day after. Obsidian is the only option on that list where the agent reads the same file you are editing, so a note you fix right now shows up in the next question you ask. That live loop is most of the value.

Export sanity check prompt

Prompt
This folder is a Notion export, not a native vault.

Before anything else, report:

- How many files, and how deep the folder nesting goes
- Whether internal Notion links survived as relative links or broke into URLs
- Which files are database CSVs rather than notes
- Any filenames still carrying the Notion hash suffix

Then propose a flattening and renaming plan. Do not execute it.

Your vault was never a filing problem. It was a reading problem, and the reading problem is solved.

The folder is already sitting on your drive. If you want the rest of the system that runs on top of it, the free masterclass is where I walk through it.

Frequently Asked Questions

Can Claude Code read an Obsidian vault?
Yes, with no setup. An Obsidian vault is a folder of plain markdown files on your drive, which is the format Claude Code reads natively. You open the vault folder as your working folder and start asking questions.
Do you need an Obsidian plugin to use Claude Code with your notes?
No. There is no plugin, no integration, and no export step. Obsidian stores notes as .md files in a normal folder, so the agent reads the same files Obsidian edits.
Will Claude Code follow the links between my notes?
Yes. A wikilink is literally the characters [[Note Name]] inside the file, and a tag is text in the body or in YAML frontmatter. Both are searchable and followable, so the agent can trace what points at what.
Is it safe to point an AI agent at my Obsidian vault?
It is safe with three habits: duplicate the vault folder before any structural change, let sync finish before starting a session, and ban deletion in the vault's CLAUDE.md so uncertain files go to a quarantine folder instead.

Related Articles

I'm building this in public. Come build with me.

The Sprint: Focus Pilot, live weekly mentorship, and a community of operators who ship with AI.

agentic aiknowledge managementnote takingclaude codepersonal knowledge base
Matt Ganzak

Matt Ganzak

Founder, The Sprint & ScaleUp Media

25+ years building software companies. Multiple SaaS exits. Bestselling author of The Million Dollar Plan. Writes about running AI agents for real operational work.