Fastest way to scaffold your LLM powered personal knowledge base - Obsidian, Claude Code, qmd, Github
BRAIN_NAME
built using claude-second-brain
Your notes don't compound. This wiki does.
Drop in a source. Claude reads it, extracts what matters, cross-links it to everything you already know, and files it. Query it six months later and get cited answers — not a list of files to re-read.
Inspired by Andrej Karpathy's approach to LLM-powered knowledge management.
Quick start
# 1. Generate vector embeddings (first run downloads ~2GB of GGUF models — once per machine)
pnpm qmd:reindex
# 2. Open Claude Code
claude
qmd collections are already registered by the CLI during scaffolding — pnpm qmd:reindex just needs to build the index.
Then open this folder in Obsidian. The Git plugin is pre-configured — enable it in Obsidian settings and your wiki syncs automatically.
Your Claude Code skills
Daily workflow
/brain-ingest — Add a file to raw-sources/articles/, raw-sources/pdfs/, or raw-sources/personal/, then run /brain-ingest. Claude summarizes the source, asks what aspects matter most, updates related wiki pages, flags contradictions, and logs everything.
/brain-search — Ask anything: what do I know about [topic]? Claude searches the wiki semantically and returns a cited answer. If it synthesizes multiple pages in a useful way, it offers to file it as a permanent wiki/qa/ entry.
/lint — Health-check the wiki. Finds orphan pages, broken links, unresolved contradictions, and data gaps. Reports findings and fixes what it can.
Maintenance
/brain-refresh — Re-scan the vault for new or changed files and regenerate vector embeddings. Run after a bulk ingest session or manual edits. Pass force to re-embed every chunk.
/brain-rebuild — Destructive. Redesigns the qmd schema: analyzes the wiki, proposes new collections and contexts, waits for your approval, then patches scripts/qmd/setup.ts, drops the old index, and rebuilds embeddings from scratch.
How it works
Sources flow in on the left, Claude synthesizes them into the wiki, qmd indexes every page into a local hybrid search index, and GitHub makes the whole vault editable from Obsidian and Claude Code on any device.
flowchart TB
subgraph Sources["raw-sources/ — raw, immutable"]
direction LR
S1[articles/]
S2[pdfs/]
S3[personal/]
end
subgraph Skills["Claude Code skills"]
direction LR
I["/brain-ingest"]
Q["/brain-search"]
R["/brain-refresh"]
end
subgraph Wiki["wiki/ — cross-linked synthesis"]
direction LR
W1[overview.md]
W2[topic / entity pages]
W3[sources/]
W4[qa/]
end
QMD[("qmd.sqlite<br/>vector + BM25<br/>hybrid index")]
subgraph Remote["GitHub — source of truth"]
GH[private repo]
end
subgraph Read["Read / edit anywhere"]
direction LR
OB["Obsidian<br/>graph · backlinks · mobile"]
CC["Claude Code<br/>desktop + mobile"]
end
User((you))
Sources -->|read| I
I -->|write pages, cross-link,<br/>flag contradictions| Wiki
R -.->|chunk + embed<br/>changed files| QMD
Wiki -.->|indexed by| QMD
User -->|ask a question| Q
Q -->|hybrid search| QMD
QMD -->|top-k pages| Q
Q -->|cited answer| User
Wiki <-->|"obsidian-git<br/>auto commit / pull"| GH
GH --> OB
GH --> CC
Obsidian Mobile
If this repo lives inside your iCloud Drive folder, Obsidian Mobile reads it with no extra setup. Graph view, backlinks, offline access — all working. The Git plugin handles sync automatically when you commit and push.
Wiki structure
Five page types, all with YAML frontmatter:
| Type | File | Purpose |
|---|---|---|
overview |
wiki/overview.md |
Evolving high-level synthesis |
topic |
wiki/[concept].md |
A concept, domain, or idea |
entity |
wiki/[name].md |
A person, tool, company, or project |
source-summary |
wiki/sources/[slug].md |
One page per ingested source |
qa |
wiki/qa/[slug].md |
Filed answers to notable queries |
All pages cross-link with Obsidian [[wikilinks]]. Contradictions are flagged with [!WARNING] callouts. Full schema in CLAUDE.md.
Directory layout
claude-second-brain/
├── CLAUDE.md ← The schema. Claude reads this every session.
├── raw-sources/ ← Your raw inputs. Claude never modifies these.
│ ├── articles/ ← Web articles saved as markdown
│ ├── pdfs/ ← PDFs or extracted text
│ └── personal/ ← Brain dumps, rough notes
├── wiki/ ← Claude owns this entirely.
│ ├── index.md ← Master index
│ ├── log.md ← Append-only activity log
│ ├── overview.md ← Evolving synthesis
│ ├── sources/ ← One summary per ingested source
│ └── qa/ ← Filed Q&A answers
└── scripts/qmd/ ← Semantic search setup and re-indexing
Installing and updating skills
Skills are slash commands Claude Code loads from .claude/skills/[name]/SKILL.md in this vault. The wiki ships with /brain-ingest, /brain-search, /brain-refresh, /brain-rebuild, and /lint pre-installed.
Update built-in wiki skills
Pull the latest skills from the upstream template:
# Install or update all 5 wiki skills
npx skills add https://github.com/jessepinkman9900/claude-second-brain/tree/main/template/.claude/skills -a claude-code -y
# Or update a specific skill
npx skills add https://github.com/jessepinkman9900/claude-second-brain/tree/main/template/.claude/skills --skill brain-ingest -a claude-code -y
Once installed via npx skills, future updates are a single command:
npx skills update -a claude-code
Re-indexing
After a bulk ingest session, re-index to keep search current:
/brain-refresh
This wraps pnpm qmd:reindex — you can also run that command directly if you're not inside Claude Code. Pass force to /brain-refresh to re-embed every chunk (e.g. after changing the embedding model).
Managing brains
From anywhere (csb is the short alias — use npx claude-second-brain if not installed globally):
csb ls # list all brains (default marked with *)
csb use <name> # switch the default brain
csb rm <name> # remove a brain
csb path qmd # print the default brain's qmd index
csb qmd query -c wiki "<terms>" # run qmd against the default brain
csb exec -- pnpm qmd:reindex # run any command inside the default brain
csb doctor # verify tools, config, and each brain
Run csb help <command> (or csb <command> --help) for per-command details.
License
MIT