Context Engineering VIP 2026-06-07

AI Agents Are Creatures That Live on Your Disk

Why does an AI agent struggle with data locked inside Notion but happily read a markdown file on your disk? That single difference is going to shape your data-keeping habits for the next ten years.

Use an AI agent for a while and you meet a strange moment. Two sets of material, equally well-organized. For one, the agent dives straight in and starts working. For the other, it says, "I can't access that data." Why? The content isn't the problem. The location is.

That small difference is going to shape how you store data for the next decade. Today we unpack the principle. The example is Obsidian, but the principle applies to any storage method. Slowly.

Agents Read Local Files

Start with one fact. AI agents natively read and write files on your computer. Claude Code, Gemini CLI, ChatGPT's agent mode — all of them work the same way. Point them at a folder and they open, edit, and create files directly inside it.

Where does most of your material live these days? In cloud services — Notion, Google Docs, Dropbox Paper. That data exists only on the web. It's not sitting as a file on your disk. For an agent to reach it, you need an API — a dedicated connector that logs into an external service — set up separately. That means auth keys, permission scopes, quota management, and debugging when errors come.

Same content as a markdown file in a local folder? The agent just reads it. No setup, no auth, instantly. Same material, different location, and the work speed is ten times apart.

Example — One Brief, Two Locations

Concretely. Imagine a YouTube video brief, one page long. The content is identical. You store it in two places.

Location A — a Notion page

  1. Issue a Notion API key (10 min)
  2. Configure MCP in Claude Code (10 min)
  3. Find the page ID (5 min)
  4. Ask: "Read this brief and write the script"
  5. The agent pulls the page through the API and dumps the whole thing into the chat
  6. About 30 repeats and the chat dies ("Prompt is too long")

Location B — a markdown file (project/idea.md)

  1. Run claude in the terminal
  2. Ask: "Read project/idea.md and write the script"
  3. Agent reads the file and starts
  4. Done

Same material. Different location. A 10-minute gap at small scale. At big scale it widens. A 100-file refactor hits the Notion API limit and stalls. The local markdown version just finishes.

Analogy — Fish and Water

Think of habitats. Fish live in water. Drop them on land and they die fast. Birds live in air. Put them underwater and they can't move.

AI agents behave the same way. The agent's habitat is the local file system. In that environment the movements feel natural. Read files, edit files, create files, restructure folders, even hand files to another agent — everything runs without friction.

Drag the agent into a cloud service and it becomes a fish out of water. It can still move, but it struggles, slows, and stops often. The AI isn't broken. The habitat is wrong.

So these days I judge storage with one question. When I can't decide where to put a piece of data, I ask: "Am I likely to hand this to an AI later?" If yes, I start it local. If no, anywhere is fine.

A Number to Pin It Down

I measured my own productivity around switching the note app.

Item Notion-centric Local markdown-centric
AI integration setup time ~3 hours total 0 min
Response time per request ~30 sec avg ~3 sec avg
100-file refactor success rate ~40% (API limits) 100%
Monthly cost Notion $10 + API fees Free

Setup time already runs infinitely wide in the ratio. Each single request is 10× faster. At 100-file scale, the success rate itself flips.

Where you keep your data isn't a format question. It's a question of which habitat you're inviting the agent into.

Apply It — One Question

Think about the data you currently keep. Ask yourself one thing.

"A year from now, will I want to hand this to an AI?"

Three possible answers.

  • Yes → move it local, or start new material local from day one
  • Not sure → default to local. Moving to cloud later is easy; the reverse is painful
  • No → put it wherever is convenient. Real-time team edits belong in the cloud

Answers one and two cover most of your material. So I set local as the default, cloud as the exception. Start from the opposite direction and the migration later eats months.

A Real Example — Three Tools, One Folder

Here's the combo I actually use. One folder, three tools mounted on it at once.

my_project/
├── notes/          ← Obsidian reads it (for readability)
├── scripts/        ← VS Code reads it (for structure)
├── data/           ← Claude Code reads it (for work)
└── handoffs/       ← session-to-session handoffs
  • Obsidian renders the markdown nicely. I use it when I want to read like a book.
  • VS Code shows folder structure at a glance. Good for moving or bulk-editing.
  • Claude Code reads the same folder from the terminal and makes actual changes.

Three tools looking at the same files. I edit a note in Obsidian and three seconds later Claude Code reads that change. Claude Code creates a file, and Obsidian shows it instantly. No sync needed. Same disk, same file.

Notion can't do this. You'd copy-paste or bridge it through APIs. Local just works.

Commands You Can Use Today

Run this right now.

# 1. Create a project folder
mkdir -p ~/Projects/my-vault
cd ~/Projects/my-vault

# 2. Point Obsidian at the same folder
# → Settings → Open another vault → pick ~/Projects/my-vault

# 3. Open the same folder in VS Code
code ~/Projects/my-vault

# 4. Enter the same folder with Claude Code
cd ~/Projects/my-vault
claude

That's it. Obsidian, VS Code, and Claude Code all look at the same folder. No additional setup.

If you already have a pile in Notion, you don't have to migrate all at once. Start writing new material locally. A year later the center of gravity naturally moves local.

Wrap-Up

Today's shape.

AI agents natively read local files. Cloud data requires an API bridge, and that bridge is 10× slower and far less reliable. The location of your data defines your relationship with the AI.

The example was Obsidian, but the principle fits any storage method. Three years from now, when Obsidian is gone and a new note app shows up, the rule "local-file-based or not" still works. Because the agent's habitat itself is the local file system. Tech changes. Habitats move slowly.

Change one habit starting today. Whenever you add new material, ask "might I hand this to an AI later?" If the answer is foggy, start it local. How easily your five-year archive moves around is decided by today's choice. The tech changes. The principle doesn't.

Data local. Tool agent. Folder shared.

Edit Section