Video summary

Nick Milo Reads My Obsidian Vault (The Full Session)

Main summary

Key takeaways

Technology

Overview

Thiago interviews Nick Milo about how Nick structures his Obsidian vault and how he uses AI (primarily Claude via cloudco-work) while preserving data control and privacy.

A recurring theme is a “translation layer” approach: your notes/files stay the core, and AI is treated as a tool that reads specific folders to produce results safely and usefully.

What Nick “diagnoses” when opening an Obsidian vault

Nick describes how he assesses a vault visually and structurally:

  • Theme (first impression)
  • Sidebar/folder structure to infer organization habits
  • Links and graph view:
    • used more for understanding “shape” than for precise utility
  • Graph observations:
    • identifies a major project (e.g., “Life in Perspective” / “LIP content”)
    • sees how Readwise-imported notes connect as supporting material
  • Placeholder notes:
    • notes can be created implicitly via linking (e.g., “faded”/empty gray notes gain backlinks before creation)
    • so you’re not starting from blank pages

He also checks scale:

  • In the demo: ~1662 files and 4 folders
  • In his main vault: ~17,000 notes, growing slowly

Key technological/product principles: AI + Obsidian

1) Do not let AI modify canonical notes directly

Nick’s core practice:

  • Keep AI’s generated work separate from original notes.
  • Use backups before pointing AI at note folders.

Risk/controls discussed:

  • Even if an integration claims read-only, Nick recommends assuming the model/tool could write/delete/change.

Backups strategy:

  • Duplicate the vault (e.g., “command D”)
  • Keep additional backups elsewhere (e.g., cloud drive / external zip / external drives)

2) Obsidian is “just files”

Nick’s framing:

  • Obsidian “doesn’t hold anything”—it’s a UI (“skin”) over a plain-text file folder on your computer.
  • This makes integrations safer because AI can read folders as Markdown/text, not proprietary database structures.

He demonstrates:

  • Duplicating the vault
  • Showing folder size is manageable (text-only baseline; media may increase size)

3) Recommended AI entry points: prefer folder-scoped reads

Nick discusses AI integration categories:

  • Point AI at a folder/directory (simplest, safest)
  • Plugins (examples mentioned: Copilot, Smart Connections)
    • He’s cautious about relying on third-party plugin ecosystems long-term due to maintenance and breaking changes.
  • Notes about avoiding Obsidian CLI unless needed (he doesn’t want to code)

AI tool discussed:

  • Claude via cloudco-work / Copilot-like “Copilot” sessions (Claude is his daily driver here)

“Granular privacy” by scoping to subfolders

Nick argues for folder-based control:

  • Share AI only what you choose (e.g., a private subfolder rather than the whole vault).
  • This makes folder organization essential—“Folders are back, baby.”

The “translation layer” concept (3-layer model)

Nick introduces AIOS (his naming for a framework) and explains a 3-layer idea:

  1. Core layer (your knowledge / Second Brain) Your actual notes and system.

  2. Middle/translation layer A small set of “orientation”/instructions that teach the AI how to work with you (he suggests this can be done with just a few files).

  3. Outer layer (the AI model/tool) Claude/OpenAI/Gemini/etc. (ephemeral, changeable)

He uses a specific “three-note” idea for orientation:

  1. Me file (identity/persona)
  2. Vault map file (folder structures + templates / note types)
  3. Skill map (what actions/skills the system supports)

This helps him start new AI sessions with consistent behavior and workflow guidance.

Diagnostic/exploratory prompting technique: “dossier”

For exploration, Nick uses prompts like:

  • “Scan all of these notes and write a dossier about me”

Similarly for projects/books:

  • “Read through this entire Obsidian vault… write a project dossier” for the book (“Life in Perspective”)

He notes:

  • It can take a few minutes.
  • While waiting, he avoids deep work and checks email/calendar/YouTube.

Handling large vaults: token limits + indexing fidelity

Nick warns that:

  • With vaults beyond roughly ~500 notes (depending on note length), AI may not index with full fidelity.
  • AI then makes assumptions, which can cause hallucinations or overconfidence.

He also suggests a heuristic:

  • Important notes have more “gravity” because they have many links; AI can prioritize via link structure (compared to PageRank-like graph influence).

Map of Content (MOC) vs table of contents

Nick distinguishes:

  • Index / TOC: more static
  • Map of content: fluid, relationship-driven, supports ideation

Goal:

  • Not a decorative artifact—it should help you think and create meaning.

He frames mapping as:

  • “architect” vs “gardener” approaches (mapping helps generate understanding and relationships)

AI output lifecycle: strict separation + “AIOS/History” folder

A key workflow feature:

  • AI generated artifacts go only into an AIOS folder, specifically AIOS/History.

This enforces a “very clear line of separation”:

  • canonical notes/thoughts remain untouched
  • AI outputs are stored in a controlled location

When AI creates content, Nick instructs it to:

  • save to AIOS/History
  • prepend timestamps + title

He prefers automation so artifacts are easy to find later (not lost in chat history).

Critique of chat-only artifacts:

  • If artifacts aren’t saved into files, they’re hard to retrieve later.

Experiments shown: dossier + linking + MOC generation

Nick demonstrates a workflow where AI:

  1. Generates a project dossier with sections like:
    • snapshot/title
    • publishing/business
    • key people
    • research/foundation
    • source material
    • status
    • key files
  2. Creates backlinks/links to existing notes (he checks the note sidebar afterward)
  3. Optionally:
    • links terms in the dossier to corresponding existing notes
    • creates a map of content section for the project

Caution from the experiment:

  • Generated “links” in a MOC may appear as literal external links (possibly to Evernote-like sources), so outputs should be inspected for correctness/meaning.

Marketing use case: turn book research into a marketing plan

Nick uses vault data for a marketing task:

  • Feed existing book-related notes + marketing plan.
  • Ask AI to produce marketing execution guidance and identify gaps/blind spots.

AI strengths highlighted:

  • AI can analyze large amounts of source material to find:
    • audience patterns
    • avatars/segments
    • surprising connections
    • what resonates most based on testimonials and audience-building notes

Design problem noted:

  • AI can over-focus on one thread (“overindex” on what it sees).
  • It’s not “truth”—it reflects the provided context.

Important critique about AI summaries (“summary of summary” problem)

Nick strongly dislikes:

  • relying on summaries made from other summaries, because the signal becomes diluted and the work stops compounding.

He wants grounding:

  • cite sources / reference specific “killer lines”
  • use verbatim quotes where required

Instruction emphasis in templates:

  • “do not rewrite me” (avoid altering his phrasing when asking about what he wrote)

Goal:

  • Ask AI to add value by finding what your source material pushes hard on that the current plan underweights/emits—not just restating content.

When context is missing across tools

Nick notes real projects are often scattered across:

  • Google Drive, ClickUp, email systems, etc.

AI needs context pulled together; otherwise it makes poor decisions. He frames this as:

  • “pulling together context” = a meta-skill
  • not a one-click “cloud AI” benefit

Main speakers/sources

  • Nick Milo: primary source/expert; demonstrates Obsidian + AI workflow, explains AIOS/translation layer, and folder-scoped Claude usage
  • Thiago: host/interviewer; questions Nick and guides through experiments

Original video