/second-brain — Karpathy LLM Wiki workflow

Wraps an existing Second Brain in Obsidian (or any markdown-based vault). The wiki vault's schema doc (CLAUDE.md or AGENTS.md) is the authoritative schema — the skill orchestrates the operations the user has been doing manually.

Mental model

Three layers, each with a clear role:

raw/      →  wiki/         →  outputs/
sources      compiled         generated
                              artifacts

• raw/ — unprocessed source material. Articles, highlights, ideas, braindumps, tweets. Type-prefixed (article-, idea-, highlights-, braindump-, note-, resource-, tweet-). Never deleted — source of truth.
• wiki/ — AI-compiled topic pages. One page per concept, not per source. Interlinked via [[wikilinks]]. INDEX.md at root.
• outputs/ — generated artifacts from queries: research summaries, analyses, slide decks. Named descriptively.

Folders to leave alone during wiki ops: Projects/, Daily/, Templates/, Inbox/, Notes/, Tasks.md, Kanban.md, Home.md.

Step 1 — Load vault config + schema

1. Read references/vault-config.md for the vault path (default: ${SECOND_BRAIN_VAULT:-$HOME/Documents/SecondBrain}/)
2. Read the vault's schema doc — <vault>/CLAUDE.md or <vault>/AGENTS.md, whichever exists (if both, read both) — for the authoritative schema. If present, trust it over references/schema.md — the user's vault is the source of truth.
3. If neither exists, fall back to references/schema.md.

Step 2 — Parse mode

Step 3 — Run the mode

capture

Inputs: URL, pasted text, file path, or screenshot.

1. Detect type from content:
   • URL → article-
   • Pasted text with quoted highlights → highlights-
   • User's own thoughts / brainstorm → braindump- or idea-
   • Single tweet / X post → tweet-
   • PDF, video, podcast → resource-
   • Quick reference (recipe, command, fact) → note-
   • If ambiguous, ask.
2. Generate a descriptive filename: <type>-<kebab-case-topic>.md (e.g., article-andrew-wilkinson-tiny-manual.md). Use the source title or topic — not the URL slug.
3. Add metadata to the top:

   source: <URL if applicable>
   captured: YYYY-MM-DD

4. Save to <vault>/raw/.
5. If the source is a URL, fetch the article content (your agent's URL fetch, or agent-browser for auth-walled pages) and save the readable text — not just the URL.
6. Report path + a one-line summary of what was saved.

Don't compile into the wiki here — capture is fast intake. Compilation is a separate, deliberate pass.

compile

The expensive but valuable operation. Process unprocessed raw files into wiki pages.

1. Find unprocessed raw files: grep wiki/*.md for Sources sections; the raw files NOT listed are unprocessed.
2. Read each unprocessed raw file + the existing wiki/INDEX.md.
3. For each raw file:
   • Extract key concepts, facts, insights
   • Default: merge into an existing wiki page if the topic overlaps. Only create a new page if the concept doesn't fit anywhere.
   • One page per concept, not per source.
   • Use [[wikilinks]] for every related concept
   • Add the raw file under the wiki page's ## Sources section with a one-line note on what was drawn from it
4. Update wiki/INDEX.md:
   • Add new pages under their category (Creative / Health & Longevity / Faith & Personal Growth / Business / Personal Growth / Tech / Hobbies / Sci-Fi / Pets — or new category if needed)
   • One line per entry: - [[Page Name]] — brief description
5. Connections section is mandatory on every wiki page. If a new page has no connections, find one before saving.
6. Quality > quantity. If a page would be <100 words, hold the raw file for now and ask the user if it should be merged into an adjacent page.

Output: list of pages created/updated, what merged where, anything held for clarification.

query

Answer a question using ONLY the wiki/raw corpus. Different from deep-research (which goes external).

1. Read wiki/INDEX.md to identify potentially relevant pages
2. Read those pages + traverse [[wikilinks]] 1–2 hops
3. Compose the answer:
   • Cite wiki pages by name: "Per [[Microplastics Detox]]..."
   • If the wiki contradicts itself, surface both sides
   • If the wiki doesn't contain the answer, say so and offer to run /deep-research to expand
4. Save to outputs/<YYYY-MM-DD>-<question-slug>.md with:
   • The original question
   • The answer
   • List of wiki pages consulted
5. Show the answer in chat + path to the saved output
6. Optional render: if --render pdf or --render html was passed, pipe the output through pandoc using the shared stylesheet. See references/schema.md → "Publishing alternatives" for the commands.

lint

Health check the wiki.

Check:

1. Orphan pages — wiki/*.md that aren't in INDEX.md
2. Connection orphans — pages with no [[wikilinks]] to other pages
3. Unprocessed raw — raw files not listed under any wiki page's Sources
4. Stale pages — most recent source >6 months old AND topic is volatile (AI, marketing, finance, health protocols)
5. Topic gaps — concepts mentioned in 3+ pages without their own dedicated page
6. Contradictions — wiki pages making opposing claims without flagging it
7. Missing connections — pages on clearly related topics with no [[wikilink]] between them (suggest /sb connect)

Output: prioritized list. Most important first (broken structure beats stale content).

connect

Find pages that should be linked but aren't.

1. Build a topic map from INDEX.md + page summaries
2. For each page, find 2–5 other pages with thematic overlap
3. Check whether each candidate is already linked
4. Suggest the missing links — and if the user approves, edit the pages to add them to their ## Connections sections

search

Quick grep across wiki/ + raw/ for a term. Return matching files with a 2-line excerpt around the match. Faster than query when the user knows what page they're looking for.

Multi-writer git sync (remote agents)

When the vault is git-backed to a hosted remote (GitHub/GitLab), the remote becomes a capture API for agents that don't have filesystem access — cloud agents, scheduled jobs, other machines. Any agent that can reach the git host's API (directly, or through an MCP integration layer like Executor) can read the wiki and capture into raw/ by committing to the default branch.

The discipline that keeps writers from diverging:

1. Local sessions pull before writing: git pull --rebase --autostash at the start of any vault work, push after committing. Never assume local is current — a remote agent may have committed since the last session.
2. Obsidian users: install the community Git plugin with auto-pull on an interval (~10 min) and pull-on-startup, but leave its auto-commit/auto-push off — sessions and agents own commits, which keeps history semantic instead of a stream of "vault backup" noise.
3. Remote agents commit append-mostly: new files in raw/ with descriptive commit messages. Append-mostly writes to distinct files make conflicts rare, and rebase absorbs interleaved writers cleanly.

Verify the loop once end-to-end when setting it up: remote commit via API → local pull → file appears in the vault.

Composes with

• deep-research — when query finds gaps in the wiki, route to deep-research to expand from external sources. Deep-research output can be captured back into raw/ for future compilation.
• paste — capture content cleanly into raw/ (especially for terminal/CLI captures).
• business-brainstorm — checks Portfolio of Businesses and Entrepreneurship & Startups wiki pages for relevant context before brainstorming.
• decide — pull from Personal Philosophy / Productivity & Systems wiki for principles when scoring Q34 ("what principles are we bending"). New: a Decision Log wiki page accumulates the narrative form of decisions over time (the decide archive is the structured form; the wiki page is the story).
• jab-hook — a Content Ideas wiki page hoppers hooks, frameworks, and stories. /jab-hook drafts pull candidates from there.
• slide-deck — content drafted in outputs/ becomes deck source; speaker notes can reference relevant wiki pages.
• pm — Projects/ folder in the vault is off-limits to second-brain; pm owns it. But a Workflow Docs wiki page captures operational patterns that show up across multiple projects.

Sibling implementations (reference)

Two other systems following the same raw → wiki → outputs pattern. Both are worth watching as upgrade paths.

• Hermes' llm-wiki skill — off-the-shelf implementation of the 3-folder pattern. Pre-built workflows for compile / query / lint. Useful for comparing schema decisions.
• Gbrain by Garry Tan — much more sophisticated. Treats the brain as a database (Postgres or PGLite) with synthesis, graph traversal, gap analysis, scheduled cron maintenance, and MCP integration. Powers a 146K-page deployment with 24K people entities. If the user's vault outgrows the markdown-only pattern, Gbrain is the upgrade direction. Borrows worth adopting today even without migrating: people-as-entities (the person- raw type + People wiki page) and scheduled maintenance (wire compile and lint to fire on a recurring schedule via the loop or compound-engineering:schedule skill).

Notes on quality

• Quality over quantity. Fewer well-connected wiki pages beat many thin ones. Hold raw files for clarification if compilation would produce a thin page.
• Don't flatten nuance. If two raw sources contradict, the wiki page should note the disagreement, not pick a side silently.
• Connections section is mandatory — every wiki page must link to at least one other page.
• Never delete raw files after compilation. They're the source of truth.
• Never modify files in Projects/, Daily/, Templates/, Notes/, Tasks.md, Kanban.md, Home.md, or Inbox/ during second-brain operations. Those belong to other workflows.