Video summary

Nick Milo Reads My Obsidian Vault Like a Doctor

Main summary

Key takeaways

Technology

Core premise: Why this vault matters

The creator explains that people repeatedly ask the same questions about using Obsidian + AI, especially around ownership/control of data. The video features Nick Milo analyzing a real Obsidian vault (not a demo) and advising how to structure it for AI use.

What Obsidian is at a technical level

  • Obsidian does not “hold” notes the way Evernote/Notion do. It’s best understood as a UI/skin over a local filesystem.

  • The vault structure shown is intentionally simple, using one main folder with subfolders such as:

    • Clippings
    • Copilot
    • Readwise
    • Web Clipper
  • Notes are plain text Markdown files:
    • Standard files you can export/port
    • Not locked into a proprietary database format

AI safety principle: “Back up before granting access”

A key risk discussed is that even if AI is intended to be “read-only,” there are reports (as the speaker notes) that content can still be modified.

Mitigation:

  • Always create backups
  • Or use duplicates before pointing an LLM at your vault

Two approaches to giving AI access to notes

1) The user’s rule (strict separation)

  • AI should never touch the originals
  • Make a duplicate vault
  • Point the AI at the copy

2) Nick’s approach (more nuanced)

  • Back up first
  • Use access more flexibly depending on the creator’s risk tolerance
  • Still emphasizes safety and control

Duplicating vaults: workflow and feasibility

Duplicating is presented as extremely straightforward:

  • Use Command + D
  • Add a date to the copy

Duplicates are treated like “throwaway experiments” because the vault size is shown to be small (tens to ~hundreds of MB), making copy-and-test practical.

Important limitation:

  • Duplicates are snapshots
  • They become stale unless re-copied/updated

Granular privacy: limiting AI scope via folders

The video emphasizes that privacy is less about encryption and more about scope control.

  • With tools like C0 / pilot-co-work (as referenced), you can restrict the AI to only a subfolder.
  • The “best practice” is to use the smallest folder that still contains the context the task needs.

Key idea:

  • If the AI can see the whole vault, the “folder boundary” concept stops meaningfully constraining what it can read/touch.
  • Therefore, folders act as a permission container for scope.

Taxonomy of AI integrations (and a preference)

Multiple integration methods are mentioned:

  1. Point an AI tool directly at the vault/folder/directory
  2. Use Obsidian CLI
  3. Install AI-related plugins (e.g., Copilot, Smart Connections)

Nick’s preference:

  • Point AI directly at a folder
  • Avoid relying on many plugins and moving parts

Why:

  • Fewer dependencies means fewer “breaking points” if plugins stop updating or maintainers change.

The “translation layer” concept (three files)

Architecturally, the key idea is that AI works best when you provide a simple “operating manual” between your notes and the LLM.

Nick describes a system using three text files:

  • Identity file (MI file): who you are / your context
  • Vault map file: folder structure + note creation templates (e.g., project notes vs idea notes)
  • Skill map: what AI skills/actions to apply next (a “just-in-time” chain)

This is positioned as a route to digital sovereignty:

  • Your notes remain the core system
  • Models are treated as interchangeable “robots” that connect at the edges

How note linking/structure improves model reliability

Nick argues that as vaults grow large, LLMs tend to sample rather than truly read everything.

  • Links and structure act as signals for importance (similar to PageRank: importance inferred from connections)

  • A cited rule-of-thumb: around 500 notes is when structure (folders/links/metadata) becomes especially important to guide sampling.

Practical “exact reference” tip

When asking AI to refer to specific notes:

  • Don’t describe them vaguely
  • Use the note’s exact file path (via search/hotkey, e.g., Command/Ctrl+P and copy paths)

Motivation:

  • Reduces token waste
  • Reduces the chance of choosing the wrong reference

“Sherpa system” for style transfer (examples-based training)

Rather than asking the model to describe a style abstractly, Nick’s approach is:

  • Provide 15–20 exemplar maps of content the creator likes
  • Let the system reverse-engineer what makes them good
  • Produce a style guide for future outputs

Workflow described:

  • Run “Sherpa”
  • Then use a questionnaire / skill chain to determine topic needs and output in the target style

Cognitive load warning: “cognitive debt”

The video critiques some AI-driven note workflows because AI outputs can create cognitive debt:

  • extra decisions
  • extra reading requirements
  • extra paths you must later process

Personal rule given:

  • Don’t generate anything you don’t have time/energy to read

Ideation vs organization

Obsidian organization is framed as more than filing—it’s a meaning-making / ideation process (mapping relationships).

  • Organizing may or may not be necessary depending on use case
  • AI search can reduce some reasons to organize
  • But AI can’t fully replace the human meaning-making benefits

Preventing AI “landfill” inside your vault (segmentation)

When AI is allowed to write new notes, the recommended governance pattern is:

  • Create an AIOS (AI workspace) folder
  • Inside it, create an AIOS history area for AI outputs

Rules:

  • Configure AI to write only into AIOS history (or specific templates only when explicitly instructed)

This preserves separation between:

  • your notes
  • AI-generated artifacts

Later, you can archive/clean older low-value AI outputs.

Testing how well AI handles Obsidian links

The video demonstrates:

  1. Generating a dossier/project note
  2. Having AI:
    • find terms corresponding to existing notes
    • create double-bracket links
    • build a Map of Content (MOC) section following Nick’s guidance

Observed behaviors:

  • Links appeared mostly literal (less “creative interpretation” and more direct mapping)
  • One step required copying a file into the vault via shell/commands because writing directly from the environment used wasn’t possible

Evaluation: what AI did well vs what was missing

  • AI produced an accurate and organized project dossier (e.g., about the creator’s book)
  • But the creator wanted novel insights, not just summaries of existing knowledge

For marketing ideation:

  • AI was asked to generate/derive ideas using limited email topic inputs

Critique:

  • AI may overindex on the provided “working files,” assuming they represent the full truth → warning about blind spots and incomplete input

Takeaway:

  • Use discernment and “build in rules” to avoid confident overgeneralization from partial datasets.

Overall takeaway: the value proposition

The vault becomes:

  • not only a “second brain” for the human
  • but also a second brain for the AI

paired with:

  • a translation layer
  • clear boundaries between your land and AI land

Finally, Obsidian’s biggest benefit is framed as file portability (“file over app”)—reducing lock-in as AI models/tools evolve over time.

Main speakers / sources

  • Nick Milo — primary expert; repeatedly interviewed/advising on Obsidian + AI workflow
  • The video creator/host — the person whose vault is reviewed; provides the rules/experiments and critiques

Original video