← SOP Library

SOP: Set Up an Obsidian Second Brain Wired Into Your AI Coding Agent

Version: 1.0 Last Updated: 2026-06-30 Owner: Sondra (run with her AI operator in Claude Code)

Purpose Stand up a personal "second brain" in Obsidian that your AI coding agent (Claude Code, Codex, or similar) can read, write, and maintain for you — so your notes, decisions, learning, and daily activity all live in ONE searchable place that the agent keeps current. The point is not a prettier notes app. The point is a brain your agent reads at the start of every session, so it actually knows your business, your projects, and what you decided last week — and so nothing you capture ever gets lost.

Scope Run this once to set up a fresh vault, then it runs itself. Best for someone working inside an AI operating system style workspace — a folder on their computer that their coding agent lives in. You do NOT need to be technical. You need Obsidian, an AI coding agent, and about an hour. This documents how Sondra's own brain is built; copy it as-is or take the parts that fit.

Tools & Resources Required

your vault. This is the "librarian" that compiles and maintains the brain.

INSIDE your agent's workspace folder and that folder is already a git repo, you get this free.

protects everything.

(Granola/Fathom), a community exporter — anything that can drop a daily file into your inbox.

Procedure

  1. Install Obsidian and create the vault inside your agent's workspace. Download Obsidian, choose "Create new vault," and put it INSIDE the folder your AI agent works in (e.g. a Brain/ subfolder). This one decision is what makes the agent able to read and write it. A vault is just a folder of plain markdown files — no lock-in, no database.
  1. Build the four-folder structure (this is the whole system). Inside the vault, create exactly four folders. Keep it this simple on purpose:

screen logs. Messy is fine. Add subfolders as feeds appear (raw/daily-notes/, raw/meetings/, raw/screenpipe/, raw/llm-exports/).

inbox. This is the source of truth. Sub-folders: wiki/brand/, wiki/projects/, wiki/people/, wiki/learning/, wiki/memory/.

The flow is always one direction: content comes IN to raw/, gets compiled UP to wiki/, and finished work goes OUT from wiki/ to output/. Raw is disposable; wiki is gold.
  1. Write the "librarian" instructions file. Put an instructions file at the top of the vault that tells your agent how to behave as the brain's keeper — name it whatever your agent reads (CLAUDE.md for Claude Code, AGENTS.md for Codex). It should say: wiki/ is yours to maintain, raw/ is the inbox, read the master index + change-log first every session, classify each raw file before extracting, link related notes, and log every change. This file is what turns Obsidian from "my notes" into "a brain my agent runs." (A working example is at the end of this SOP.)
  1. Install the three community plugins that actually earn their keep. In Obsidian → Settings → Community plugins → Browse. Resist installing 20. These three pull real weight:

break when a link dies. Non-negotiable if you paste screenshots.

vault. Great for capturing ideas away from the desk.

You do NOT need the "Obsidian Git" plugin if your vault lives inside an agent workspace that's already a git repo — the parent repo versions it for you, which is simpler and one less moving part. Only reach for the Obsidian Git plugin if your vault stands alone.
  1. Turn on version history + backup. If the vault is inside a git repo, commit it regularly (your agent can do this) — that's your undo button across time. Separately, run a whole-machine backup like Backblaze so the gitignored and binary stuff is safe too. Two layers: git for "what changed," backup for "my computer died."
  1. Wire the compile workflow. "Compile" = the agent reads new/changed files in raw/, classifies each one (brand fact, project update, decision, learning, person, idea), writes or updates the right wiki/ article, cross-links related notes, and appends one line to a change-log. Set it to run once per day at the start of a session (Sondra's runs automatically inside her /prime session-start routine). The agent should skip if it already compiled today, so it's cheap to leave on.
  1. Connect your capture feeds (so the inbox fills itself). The brain is only as good as what flows in. Point automatic feeds at raw/:

Each feed is optional — start with one, add more later. Even just daily voice notes is enough to make this worth it.

  1. Adopt the weekly review habit. Once activity is flowing in (especially the Screenpipe screen log), do a weekly pass: ask your agent to read the week's raw activity and tell you what you actually worked on, what shipped, what stalled, and what you said you'd do but didn't. This is the payoff loop — the brain reflects your real week back at you instead of you guessing. Pick a fixed day (Sunday or Friday).
  1. Verify it actually works (don't assume). Drop a test note in raw/, run a compile, and confirm a wiki/ article was created or updated and the change-log got a line. Then ask the agent a question only the brain would know ("what did I decide about X?") and confirm it answers from the wiki — not from thin air.

Definition of Done

output/, templates/).

wiki/ article, with a change-log entry.

How to Use Your Brain — What to Put In It and How

Setup is the easy part; the value is in the habit. Higher-level ways to actually use it:

whole point is that YOU never have to sort — the librarian does. Lower the bar to capture.

become the agent's richest source of decisions, tasks, and ideas. This is the #1 input.

into raw/. The agent turns them into wiki/learning/ articles and wiki/people/ profiles of the experts you follow — so your learning compounds instead of evaporating.

you walk into any conversation with context.

lives, what's next, what to watch out for. This is what stops "wait, where did I leave that?"

wiki/memory/. The agent reads these every session, so it stops making the same wrong call twice — it learns how YOU work.

on X," "what did that expert say about Y," "what were my open questions last month." The agent reads the index, finds the article, answers.

agent can write emails, posts, and sales copy that actually sound like you — pulling from the brain instead of starting blank.

the agent reads it back so you can SEE your week and course-correct. Reflection without effort.

Start with ONE habit (daily voice notes) and ONE feed. Add the rest only when the first sticks.

Common Mistakes to Avoid

the whole system falls apart. Vault goes inside the workspace folder.

agent was supposed to solve. Dump to raw/, let it compile.

with what it compiles. Put new input in raw/ instead.

dies or a binary/gitignored file is lost. Run a real whole-machine backup too.

wiki. The compile is what makes it a brain, not a junk drawer — run it daily.

trickle in and the weekly review, not the initial setup.

This SOP documents how Sondra's "Sondra Brain" vault is built and run. Vault structure: four folders (raw/wiki/output/templates), three plugins (Local Images Plus, Style Settings, Telegram Inbox), versioned via the parent workspace git repo + Backblaze, compiled daily by the agent at session start.