second-brain-starter
A Claude-Code-native second brain, packaged as five skills plus a vault skeleton. It's built around two use cases — learning a technical/industry domain over time, and running a structured search-and-track pipeline for roles or opportunities — but the underlying architecture (capture → digest → compound → self-review) is domain-agnostic. Pairs with Obsidian, optionally: the vault is plain Markdown + frontmatter, so any editor works, but Obsidian's graph view and the optional mcpvault MCP server make cross-linking and search nicer.
Industry learning: /daily-digest captures raw material (text, links, screenshots) from a session; /summarize and /process-call turn long-form content (videos, articles, PDFs, recorded calls) into study digests that compound onto existing entity pages instead of piling up as disconnected notes.
Career OS: /network-sync and /job-pipeline apply the same capture → digest → compound → self-review architecture to contacts and job postings, routing them into person and company pages with a tracked pipeline of open roles.
Skills
| Command | Input | Output |
|---|
/daily-digest | Whatever you typed/pasted/shared this session | Unclassified raw note in 00_RAW/notes/ |
/summarize | YouTube URL, article URL, PDF, EPUB | Study digest in 10_WIKI/Summaries/ + compounded updates to existing wiki pages |
/process-call | Recorded call (audio file or transcript) | Standard transcript in 00_RAW/voice/ + compounded updates to person/company/concept pages |
/network-sync | (reads 00_RAW/notes/ raw files from /daily-digest) | Routes job-search-relevant people into 10_WIKI/Entities/People/{Classmates,Network}/ |
/job-pipeline | (reads 00_RAW/clips/ job posting clips) | Company insight updates in 10_WIKI/Entities/Companies/ + a tracked role note in 30_OTHER_WORLDS/Career_Story/_Pipeline/ |
All writes follow vault/CLAUDE.md's hard rules: archive-before-overwrite, attribution required on investment signals, a 10_WIKI/Log.md entry on every new/overwritten file.
Requirements
| Tool | Required? | Needed for |
|---|
| Claude Code | Required | Everything — these are Claude Code skills |
Obsidian + Local REST API plugin, or the mcpvault MCP server | Optional | Faster reads/writes against the vault. Every skill falls back to plain file tools (Read/Write/Edit/Glob/Grep) if it's not installed or a call errors — nothing in this vault requires Obsidian to function |
yt-dlp | Optional | /summarize's YouTube path (metadata + captions) |
pandoc | Optional | /summarize's EPUB path |
pdftotext (poppler) | Optional | /summarize's fallback for long PDFs |
ffmpeg + a local Whisper model | Optional | /summarize's caption-less YouTube path. Without them, /summarize asks you to paste a transcript instead; captioned videos don't need either |
| A speech-to-text service with speaker diarization (e.g. ElevenLabs Scribe) + its API key | Optional | /process-call on a raw audio file (Lane B). Without it, /process-call still works on an existing text transcript (Lane A) |
| A local transcriber (e.g. whisperX) | Optional | /process-call's automatic fallback if the cloud speech-to-text call fails at runtime. Without it, there's no fallback when the cloud call fails |
A missing tool in any optional row just means that one input type doesn't work (e.g. no pandoc → /summarize can't take EPUBs). Nothing else breaks.
Install — easy mode
Open Claude Code in any directory and paste this:
Install the second-brain-starter pack from https://github.com/bondaehsu5814/second-brain-starter.
1. Clone the repo to ~/src/second-brain-starter.
2. Ask me where I want the new vault to live (default: ~/second-brain).
3. Copy the repo's vault/ folder to that location.
4. Copy templates/ from the repo into the new vault's templates/ folder.
5. Run `git init` inside the new vault folder and make an initial commit of everything
in it (the skills assume the vault is a git repo).
6. Create ~/.claude/skills/ if it doesn't exist, then symlink daily-digest, summarize,
process-call, network-sync, and job-pipeline from ~/src/second-brain-starter into it.
Claude will: 1. Clone the repo 2. Ask where to put the vault 3. Create the vault at that path 4. Copy the page templates into the new vault 5. git init the vault and make an initial commit 6. Symlink the five skills into ~/.claude/skills/.
Then quit Claude Code, cd into the new vault folder (e.g. cd ~/second-brain), and launch Claude Code again from there — every skill treats the directory you opened Claude Code in as the vault root, and restarting is what picks up the new skills. Then run /daily-digest or /summarize <url> to try it.
Recommended: point this at a new, dedicated vault rather than an existing one. The skills write according to a specific folder contract (00_RAW/, 10_WIKI/, 20_ACE/, 30_OTHER_WORLDS/, 99_AIOS/) defined in vault/CLAUDE.md — dropping them into a vault with a different structure will produce misplaced notes until you reconcile the two schemes.
Install — individual skill only
mkdir -p ~/src && cd ~/src
git clone https://github.com/bondaehsu5814/second-brain-starter.git
mkdir -p ~/.claude/skills
ln -s ~/src/second-brain-starter/summarize ~/.claude/skills/summarize
Swap summarize for any of daily-digest, process-call, network-sync, job-pipeline. An individually-installed skill still expects the vault folder contract above — copy vault/CLAUDE.md, the relevant 99_AIOS/ files, and templates/ into wherever you point it, make sure that folder is a git repo (git init + an initial commit) if you'll use /process-call or 👀 New Progress, and open Claude Code from inside that folder.
Usage
Capture scattered thoughts from a session
/daily-digest
Pulls everything typed, pasted, or shared in the current session into one clean, unclassified raw file under 00_RAW/notes/. Deliberately doesn't decide what category anything belongs to — that's a later pass's job.
Digest a video, article, PDF, or EPUB
/summarize https://www.youtube.com/watch?v=...
Produces a distilled study digest (not a transcript) plus edits to any existing wiki pages the content touches, so insights compound instead of sitting in isolated files.
Process a recorded call
/process-call ~/Downloads/call-2026-09-30.m4a
Transcribes and speaker-identifies the call, then fans the content out to the person/company/concept pages it concerns, with honest source-vs-self conviction recorded on anything investment-relevant.
Route today's people mentions into contact pages
/network-sync
Reads the raw files /daily-digest produced, pulls out people worth tracking as professional contacts, and writes or updates their person pages.
Route a saved job posting into company + role pages
/job-pipeline
Splits a clipped job posting into a company-level insight update (shared with other research on that company) and a role-level tracking note for that specific posting.
Architecture philosophy
This starter kit is less about the five skills and more about the governance layer underneath them — the rules that make it safe to let an agent write into your notes unsupervised. Each idea below names the exact file it's implemented in, so you can go verify it rather than take this README's word for it.
Immutable raw zone. 00_RAW/ is a capture zone: capture skills (/daily-digest, /process-call) add new files there, and /daily-digest may back up and merge into that same day's own file on a second run. Beyond that, existing raw content is never rewritten by any skill — the one other sanctioned write is flipping a processed flag. Capture and interpretation are kept as separate steps: a digest can misread a raw note, but the raw note itself is never at risk of being silently altered by a later processing pass. Defined in vault/CLAUDE.md, hard rule 1.
Append-only log. Every new or overwritten file under 10_WIKI/ or 20_ACE/ gets one line appended to 10_WIKI/Log.md — who, what, input, output. This gives a plain-text audit trail of every agent write without needing git blame or a database. Defined in vault/CLAUDE.md, hard rule 5, with 10_WIKI/Log.md as the actual log file.
Archive-before-overwrite. Before any skill overwrites or deletes an existing file, it copies the old version to 99_AIOS/_archive/ first. The vault is git-tracked, so this is partly belt-and-suspenders, but it makes recovery obvious even outside git (a stray manual edit, a non-git export of the vault). Defined in vault/CLAUDE.md, hard rule 7.
Staleness checks. 99_AIOS/Hot.md carries a Last Updated date; if it's more than 14 days old, an agent reading it is required to flag it as history rather than present state and offer a refresh, instead of quietly reasoning from stale context. Defined in vault/CLAUDE.md, hard rule 6.
Read-on-demand token discipline. Most of 99_AIOS/ is intentionally not loaded into context by default. vault/CLAUDE.md keeps a small table mapping task type to the one governance file worth reading for it (e.g. processing raw clips → Processing_Flow.md; choosing whether to escalate → Judgment_Rubrics.md), so an agent pulls in exactly the context a task needs instead of front-loading everything on every run.
Judgment rubrics for agent autonomy. vault/99_AIOS/Judgment_Rubrics.md turns otherwise-fuzzy calls — when to escalate to a bigger model, when a task is actually done, when to stop and ask the user instead of guessing, when to change approach instead of retrying — into checkable tests with a passing and a failing example each. The goal is that a smaller/cheaper model gets the same decision right as a larger one would, because the judgment call has been reduced to a rubric rather than left to intuition.
License
MIT. See LICENSE.
