Keep AI Out of Your (Obsidian) Vault
mehdio's contrarian essay on why wiring AI into an Obsidian vault is "a dead end." Obsidian's local, open markdown makes AI integration as easy as it gets, but the author argues that should not mean generating notes or connections inside your personal knowledge base.
The core argument
Long AI summaries and auto-generated notes become "AI slop": over time you can't tell your own writing from the model's, and your own — far more valuable — thoughts get diluted. Searching a vault full of generated text means fighting through noise; a vault of only your own writing is all signal, because every note was written for a reason ("a conviction, idea, or something that moved you").
Where AI does earn its keep
- Vibe-coding agents: use the Obsidian CLI — faster file ops, search, and grep than the app.
- Retrieval: finding related notes ("maybe the best use case") via Smart Connections, Graph Analysis, or Omnisearch — vector/similarity search out of the box.
- Voice-to-text: only when you're driving, and often more work than typing.
- Isolation: if you do generate, put it in a dedicated PARA "AI" folder hidden from search, and mark AI text clearly.
What to keep AI out of
- Tagging and organization — connections an agent makes "won't count for anything, since they aren't made by you."
- Generating your system ("whose knowledge system is it?") — a second brain is where your thinking happens.
- Long summaries — Kepano: "A summary of a PDF is noise. An insight I had from reading the PDF is signal." A 1–2 sentence intro is fine.
The forward-looking case
"Your notes will be prompts or libraries for AI tomorrow, but not if you generate them." The author wants human-curated knowledge preserved at scale — better training data than synthetic, "just all equally similar" output — and treats Karpathy's auto-generated LLM wiki as a cautionary tale ("hard to finish an idea that is not yours").