Second Brain Starter

An agent-maintained second brain built on Obsidian (plain Markdown, local-first) and Claude Code (or any agent that can read/write files on a schedule).

The idea: you work normally with your AI agents, and the vault maintains itself in three layers —

1. Per-session capture (event-driven). Agents follow a capture-by-default rule: any session that produces a durable lesson, decision, or reusable process writes a timestamped note into 01 Inbox/Agent Captures/. Trivial Q&A gets skipped.
2. Daily digest (scheduled, e.g. 5:00 AM). A scheduled task reads all of yesterday's and today's agent session transcripts and rebuilds a digest section in each day's daily note — so even sessions that didn't self-capture get summarized. Follow-ups carry forward day to day until resolved.
3. Weekly cleanup (scheduled, e.g. Monday 5:07 AM). Sorts inbox captures into permanent homes, regenerates an AI-readable Vault Index, maintains a single cross-project Open Follow-ups list, and flags projects that have gone stale.

Agents get context the same way every time: read AI Start Here → check the Vault Index → read the relevant project dashboard → read the newest linked notes. Freshness is tracked with last_verified dates instead of assumed.

Repo layout

vault/                          # Obsidian vault skeleton — point Obsidian at this
  00 Home/                      # Entry points: AI Start Here, Vault Index, Open Follow-ups, daily notes
  01 Inbox/Agent Captures/      # Where agents drop new captures
  02 Projects/                  # One folder + dashboard note per project
  03 Areas/                     # Ongoing responsibilities (never "done")
  04 Resources/                 # Reference material
  90 Archive/                   # History — never treated as current truth
  99 Sensitive Review/          # Quarantine for sensitive content — gitignored, agents never read it
  _templates/                   # Note templates (project, area, capture, daily, decision, resource)
  _system/sort-manifests/       # Machine-written logs from the weekly cleanup

claude/
  skills/second-brain/          # The always-available skill: retrieval order, capture rules, safety
  scheduled-tasks/daily-digest/     # Prompt for the daily digest job
  scheduled-tasks/weekly-cleanup/   # Prompt for the weekly cleanup job

Prerequisites

• Obsidian (free) — the app you'll use to read and write notes. The vault is plain Markdown, so the automation below works on raw files either way; Obsidian is the human-facing layer (wikilinks, graph, daily notes).
• Claude Code — runs the skill and the two scheduled jobs. Any agent that can read/write files on a schedule can substitute, but the included prompts are written for Claude Code.
• ripgrep (optional) — makes the bundled search scripts fast; the PowerShell script falls back to built-in search without it.

Setup

1. Create your vault

Copy the vault/ folder wherever you keep your notes (or clone this repo and use it in place), then open that folder as a vault in Obsidian. Optionally git init inside it — the included .gitignore keeps 99 Sensitive Review/ and Obsidian workspace state out of history. If you version your vault, keep that repo private.

2. Install the skill

Copy claude/skills/second-brain/ into ~/.claude/skills/. Then open the three files and replace every <vault-path> placeholder with your vault's absolute path (e.g. C:\Users\you\SecondBrain or /Users/you/SecondBrain).

This skill is what makes agents capture by default and retrieve context in a consistent, safe order.

3. Schedule the two jobs

Copy both folders from claude/scheduled-tasks/ into ~/.claude/scheduled-tasks/, and replace <vault-path> in each SKILL.md. Then wire them to a schedule — either:

• Claude Code scheduled tasks (desktop app): ask Claude Code to "run the daily-digest task every day at 5:00 AM" and "run the weekly-cleanup task every Monday at 5:07 AM", or

• Your own scheduler (cron / Windows Task Scheduler): run headless, e.g.

  claude -p "$(cat ~/.claude/scheduled-tasks/daily-digest/SKILL.md)"

4. Seed the vault

Fill in 00 Home/AI Start Here.md (it has placeholders), create a dashboard note in 02 Projects/ for each active project using _templates/Project Note.md, and let the automation take it from there.

Conventions that make it work

• Frontmatter contract — every note uses the same small property set (type, status, created, updated, last_verified, next_action, source, …). See claude/skills/second-brain/references/SCHEMA.md. This is what lets the weekly job regenerate the Vault Index mechanically.
• Current truth vs. history — a dashboard is only "current" if last_verified is recent and its sources support it. Archive notes are history, never current state. Agents are told to say "this looks stale" instead of guessing.
• Follow-ups never silently drop — daily digests carry unchecked items forward with their origin date; the weekly job promotes them to Open Follow-ups.md; items only resolve with cited evidence.
• Safety rails — agents never read 99 Sensitive Review/, never copy secrets into notes, never delete notes, and cap how many files a cleanup run may touch.

Customizing

• Rename/renumber the folders freely — just keep the skill, the two scheduled tasks, and the templates consistent with each other.
• Using a second agent (Codex, a local model, etc.)? Give it the same skill and have it write its own ### <Agent> Digest section under ## Agent activity in the daily note — the schema is shared and each automation only ever replaces its own section.
• The 21-day staleness threshold, the 25-file cleanup cap, and the digest bullet caps are all just numbers in the prompts. Tune them.

License

MIT — see LICENSE.