I keep a plain Obsidian vault — about 1.4 MB of markdown across notes, projects and an inbox — and I let coding agents (the Claude Code / Cursor kind) work inside it: searching, summarizing, drafting. The first thing I measured changed how I set all of this up: an agent that reads the whole vault isn't being thorough. It's drowning.
Every file it opens costs context. Read enough files and the actual question gets buried under files that were never relevant. Three boring fixes, each measured on my own setup:
1. A derived index instead of the raw vault
A small script walks the vault and writes one line per note: title, what it's about, size, last change. That map is ~9 KB standing in for ~1.4 MB — the same answer to "what exists here and where," for about 150× less context. It's generated, never hand-maintained: I rerun the script after heavy edits, because a stale map is worse than none — it asserts confidently things that stopped being true.
2. A reading ladder, cheapest rung first
Measured cost per question on my setup:
- keyword lookup against the index: ~0.5 KB
- ranked snippet search that returns excerpts with a file → section trail: ~1 KB
- outline of the one file that matters: ~1.1 KB
- reading that file whole: ~35 KB
Most questions die on the second rung and never pay the 35 KB. The specific tools don't matter; the point is that "just read everything" is the most expensive possible answer to most questions.
3. A writing rule the agent can actually see
Reading is free; writing is gated. A plain AGENTS.md at the vault root states it: the index is the front door, the agent may read anything, and it writes nothing without asking — with folder ownership explicit (my notes vs. a staging area it may use). This is the piece that stops an assistant from quietly reorganizing your knowledge base overnight.
If you'd rather DIY
Part 1 is an afternoon of Python (walk the folder, pull first heading + gist + size + modified date, one line per note). Part 2 done well is the fiddly one — ranking by term rarity and returning section trails took real iteration to stop returning junk. Part 3 is one text file.
Full disclosure: I ended up packaging all of this — the vault starter (notes/people/projects/inbox with example notes teaching each folder's pattern), the two tools, the rules file, a 10-minute tutorial and a short ebook on the method, including a what-NOT-to-do section — as a US$15 one-time template. It's my product; the link is below. The three ideas above work fine with nothing but your own scripts — build your own index and you're 90% there.
Second Brain Starter — US$15 one-time
Happy to answer questions about the measurement method, or about where this approach breaks.
Cross-posted from Oroboro Labs — this post is the canonical version; the original carries the full method notes.
