Personal Knowledge Base

One of many experiments. This one inspired by Andrej Karpathy's idea that a wiki maintained by an LLM compounds knowledge in ways a human-maintained one hardly sustains. Source

---

What this is

A personal knowledge base where Claude maintains a persistent, interlinked wiki from raw sources — articles, notes, transcripts. The LLM handles all filing, cross-referencing, and synthesis. You handle sourcing and questions.

The wiki gets richer with every source added and every query answered. Nothing is rediscovered from scratch.

Architecture

┌─────────────────────────────────────────────────────────┐
│                        YOU                              │
│          clip articles · ask questions                  │
└────────────────┬─────────────────┬───────────────────── ┘
                 │                 │
          new article           query
                 │                 │
                 ▼                 ▼
┌────────────────────┐   ┌─────────────────────┐
│   Obsidian Web     │   │     Claude Code      │
│   Clipper          │   │   + qmd (MCP)        │
└────────┬───────────┘   └──────────┬──────────┘
         │                          │
         ▼                          │ qmd search → ranked pages
  raw/articles/                     │ synthesize → propose file-back
         │                          │
         ▼                          ▼
  fswatch detects          ┌─────────────────┐
  new .md file             │   wiki/         │
         │                 │                 │
         ▼                 │  Findings/      │ ← cross-source claims
  scripts/ingest.sh        │  source-notes/  │ ← one per source
         │                 │  Meta/          │ ← architectural docs
         ▼                 │  index.md       │ ← navigation
  Claude Code              │  log.md         │ ← append-only record
  (--print mode)           └────────┬────────┘
         │                          │
         ├── writes source note     │
         ├── updates index.md       │
         ├── appends log.md    ─────┘
         ├── increments counter
         │
         ▼
  counter >= 5?
         │
    yes  ▼
  Synthesis runs inline
  Findings pages written
  counter reset to 0
         │
         ▼
  git commit (atomic)
         │
         ▼
  qmd reindex

How it works

Three layers:

• /raw/ — immutable source documents, never modified
• wiki/ — everything Claude writes: source notes, Findings, Meta, index, log
• CLAUDE.md — the schema: architecture, page types, format templates, conventions
• .claude/skills/ — the procedures: one skill file per workflow, loaded on demand

Three wiki page types:

• Source Notes — one per source, the non-obvious insight a first-pass reader would miss
• Findings — cross-source claims that only emerge when holding multiple sources simultaneously. Named after the assertion, not the topic.
• Meta — architectural documentation about how this wiki works

Four workflows (each a skill, loaded only when invoked):

• Ingest (source-ingestor) — triggered automatically when a new file lands in raw/. Dedup check → quality gate → write → synthesis trigger → commit
• Synthesis (synthesis-builder) — runs inline at every 5th ingest, or on demand. Reads all source notes, identifies cross-source claims, writes or extends Findings pages
• Query (query) — qmd retrieves semantically relevant pages, Claude synthesizes, proposes filing back if the answer produces something new
• Lint (lint) — periodic health check: name collisions, orphan pages, stale links, CLAUDE.md drift

The synthesis trigger:

scripts/ingest_counter  →  increments every ingest
                        →  synthesis fires when value >= 5
                        →  resets to 0 after synthesis

The file-back loop (Karpathy's key insight):

query → qmd retrieves wiki pages → Claude synthesizes
      → new cross-source insight? → propose filing back
      → filed page gets indexed by qmd
      → next query starts from richer foundation

Stack

• Claude Code — LLM that maintains the wiki (--print --dangerously-skip-permissions)
• qmd — local hybrid BM25 + vector search with LLM reranking, wired as MCP server
• Obsidian — local markdown editor and graph viewer
• Obsidian Web Clipper — clips web pages to raw/articles/
• fswatch + launchd — watches raw/ for new files, auto-triggers ingest
• git — versions schema, skills, and scripts; wiki content stays local

Authentication

Background ingest runs on Claude subscription (no API key). Auth is handled via a long-lived OAuth token stored outside the repo:

~/.claude_ingest_token   ← generated by: claude setup-token
                         ← never committed to git

ingest.sh reads this file and sets CLAUDE_CODE_OAUTH_TOKEN before calling claude --print. ANTHROPIC_API_KEY is explicitly unset so API billing never occurs.

Watch out: The long-lived token will eventually expire. When ingest starts failing with 401 errors in scripts/watcher.log, run claude setup-token in your terminal, save the output to ~/.claude_ingest_token, and you're done.

What's tracked in this repo

CLAUDE.md                        Schema — architecture, page types, formats, conventions
.claude/skills/source-ingestor/  Ingest workflow
.claude/skills/synthesis-builder/ Synthesis workflow
.claude/skills/query/            Query workflow
.claude/skills/lint/             Lint workflow
README.md                        This file

Wiki content, raw sources, assets, and config are local only — not pushed.

Inspiration

Andrej Karpathy's gist on LLM-maintained knowledge bases: https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f