Brain

What This Skill Does

Manages Sarthak's personal second brain — a Karpathy-style LLM knowledge base at personal/second-brain/. Raw materials go in, interconnected wiki articles come out.

Command Detection

Parse the user's message to determine the command:

• ingest: "brain ingest", "ingest this", "add to my brain", "process the raw folder"
• compile: "brain compile", "recompile", "rebuild the wiki"
• query: "brain query ...", "what does my brain say about ...", "search my knowledge"
• lint: "brain lint", "check my wiki", "health check"
• connect: "brain connect", "connect the dots", "find patterns", "cross-domain insights"
• status: "brain status", "how big is my brain", "brain stats"
• add: "brain add ...", "save this to brain", "remember this insight" (adds quick note to raw/notes/)

If no command is clear, ask the user what they'd like to do.

Setup

The knowledge base lives at: personal/second-brain/

personal/second-brain/
├── CLAUDE.md              ← Schema (rules for the LLM)
├── raw/                   ← Source materials (human curates)
│   ├── articles/          ← Web articles, blog posts
│   ├── papers/            ← Research papers, PDFs
│   ├── tweets/            ← Interesting tweets/threads
│   ├── conversations/     ← Key insights from conversations
│   ├── notes/             ← Quick brain dumps
│   └── media/             ← Images, screenshots, diagrams
├── wiki/                  ← Compiled knowledge (LLM writes)
│   ├── _index.md          ← Master index
│   └── _changelog.md      ← What changed and when
└── outputs/               ← Query responses, reports

Always read personal/second-brain/CLAUDE.md first — it contains the full schema with article format, rules, and domain tags.

Command: INGEST

1. Read the schema at personal/second-brain/CLAUDE.md
2. Scan raw/ for new or unprocessed materials
3. For each source:
   • Read it fully (never chunk)
   • Extract key concepts, insights, decisions, mental models
   • Create or update wiki articles following the article format in the schema
   • Add backlinks to related articles
   • Add source attribution
4. Update wiki/_index.md
5. Append to wiki/_changelog.md
6. Report: N sources processed, N articles created, N articles updated

If the user provides content inline (not as a file), save it to raw/notes/YYYY-MM-DD-<slug>.md first, then ingest.

Command: COMPILE

1. Read the schema
2. Read wiki/_index.md for current state
3. Scan ALL raw/ materials
4. Identify gaps and stale articles
5. Regenerate/update articles
6. Rebuild index
7. Run lint
8. Report results

Command: QUERY

1. Read the schema
2. Read wiki/_index.md
3. Identify relevant articles based on the question
4. Read those articles fully
5. Synthesize answer grounded in wiki content
6. If insufficient info, suggest what raw material would help
7. Save response to outputs/query-YYYY-MM-DD-<slug>.md

Command: LINT

1. Read the schema
2. Scan all wiki articles checking for:
   • Broken [[backlinks]]
   • Orphan articles (nothing links to them)
   • Missing source attributions
   • Missing or incorrect domain tags
   • Stale content
3. Auto-fix safe issues
4. Write report to outputs/lint-report-YYYY-MM-DD.md
5. Show summary to user

Command: CONNECT

This is the magic — finding cross-domain patterns.

1. Read the full index and all article summaries
2. Look for non-obvious connections:
   • A trading lesson that applies to product building
   • An engineering pattern that mirrors a business strategy
   • An AI technique useful for content creation
   • Mental models that span domains
3. Write "Connections Report" to outputs/connections-YYYY-MM-DD.md
4. Present the most surprising/valuable connections to the user

Command: STATUS

1. Count files in raw/ by subdirectory
2. Count articles in wiki/
3. Count articles by domain tag
4. Show total word count
5. Show last ingest/compile date from changelog
6. Present a clean summary

Command: ADD

Quick capture — save an insight without full ingestion:

1. Take the user's text
2. Save to raw/notes/YYYY-MM-DD-<slug>.md with basic frontmatter:

   ---
   added: YYYY-MM-DD
   source: conversation
   ---
   
   <user's text>
   

3. Confirm saved, suggest running /brain ingest later to compile it into the wiki

Important Rules

• Always read the CLAUDE.md schema before any operation
• Never fabricate knowledge — every wiki claim traces to a source
• Write for future-Sarthak who has no memory of original context
• Cross-link aggressively — connections are the value
• Use Obsidian-compatible markdown ([[backlinks]], standard frontmatter)
• Keep the index under 200 entries