AI-Powered Learning System — Turn any topic into a structured, high-density curriculum with interactive study tools, all living in your Obsidian vault.
Built on Claude Code + NotebookLM + Obsidian + YouTube ASR.
You say: "帮我研究 distributed systems"
You get:
├── 00-Overview.md ← Framework + study plan + architecture diagram
├── 01-Consensus-Algorithms.md ← Teaching note with mermaid/code/tables
├── 02-Distributed-Transactions.md
├── 03-...
├── materials/
│ ├── flashcards.md ← NLM-generated flashcards
│ ├── quiz-answers.md ← Quiz with separate answer key
│ └── overview-infographic.png
└── _deep-notes/ ← Created as you study, linked back to source
├── Raft-vs-Paxos.md
└── 反思-CAP-Theorem.md
Five specialized roles produce progressively denser content:
| Phase | Role | What it does |
|---|---|---|
| 1 | Resource Collector | YouTube ASR + Academic Papers + 知乎/CSDN (parallel) |
| 2 | Topic Discoverer | Reads all ASR transcripts → extracts sub-topics + entities |
| 3 | Framework Architect | Designs learning path with NLM → atomic headings + multimodal marks |
| 4 | Content Writer | Fills framework with dense content (parallel agents) |
| 5 | Deep-Diver | Enriches thin sections with examples, code, edge cases |
| 6 | Learning Path Reviewer | Audits learning journey + content density → directly fixes issues |
Multi-modal content: Quiz (interactive checkboxes) · Flashcards · Mermaid diagrams · Infographics · Data tables · Code blocks
Multi-source fusion: YouTube ASR transcripts + NotebookLM synthesis + Academic papers (Semantic Scholar/arXiv) + Web search (知乎/CSDN) + Claude knowledge
Copy any sentence from your study notes → get instant explanation, deep research, or Socratic discussion.
You paste: "HNSW索引的层级结构"
Claude: [200-word structured explanation in <5s]
You: "还有其他例子吗"
Claude: [follow-up with 3 concrete examples]
You: "我觉得IVF比HNSW好因为..."
Claude (Professor mode): "你凭什么这么说？IVF的跨簇丢失问题你考虑了吗？"
You: "ok"
→ Auto-creates _deep-notes/HNSW-vs-IVF.md linked back to exact source line
Features:
- Fluid role switching (expert ↔ professor) within one conversation
- Block reference (
^block-id) for precise bidirectional navigation - Sub-document dedup: same term from different docs → shared annotation
- Context-efficient: <2000 tokens per 5-round exchange
All notes auto-interlink via [[wikilinks]]. The more you research, the denser your graph.
- Backlinks panel shows every note linking to current one
- Block references for line-level precision
- Unresolved links reveal knowledge gaps
- Obsidian (free)
- Claude Code (requires Anthropic API)
- Python 3.10+
- Chrome browser (for YouTube subtitle access)
# Clone the repo
git clone https://github.com/YOUR_USERNAME/azio-vault-os.git
cd azio-vault-os
# Run setup (installs yt-dlp, notebooklm-cli, etc.)
./scripts/setup.sh
# Authenticate NotebookLM (opens browser)
notebooklm login
# Open vault in Obsidian
# File → Open Vault → select vault-template/# In Claude Code, navigate to your vault
cd /path/to/vault-template
# Start researching
# Just type naturally:"帮我研究 AI Agent架构设计，我要准备面试"
Claude will:
- Ask you to confirm the framework
- Search YouTube + download ASR transcripts
- Search academic papers
- Create NotebookLM notebooks
- Write teaching notes (parallel agents)
- Generate quiz, flashcards, infographics
- Output a complete project folder in your vault
While reading any teaching note, copy a sentence and ask:
"ANN 索引算法对比 — HNSW vs IVF, 详细点"
Or share your understanding:
"我觉得 TCC 比 Saga 好因为一致性更强"
Claude switches to professor mode and challenges your reasoning.
vault-template/
├── CLAUDE.md ← Constitution: note rules, folder routing, tool paths
├── Home.md ← Dashboard
├── .claude/skills/ ← AI skills (the engine)
│ ├── research-pipeline-v2/
│ │ ├── SKILL.md ← Main workflow (8 phases)
│ │ └── references/
│ │ ├── roles.md ← 6 role definitions + NLM prompting
│ │ └── templates.md ← Teaching note templates
│ ├── study-assist/
│ │ ├── SKILL.md ← Interactive learning companion
│ │ └── scripts/locate.py ← Fast vault content locator
│ ├── notebooklm-notebook/ ← Direct NotebookLM control
│ └── youtube-search/ ← YouTube search with metadata
├── projects/ ← Research outputs live here
├── concepts/ ← Atomic concept notes
├── sources/ ← Video/article reference notes
├── daily-notes/ ← Daily journal
└── inbox/ ← Quick captures
The system defaults to Chinese (中文) for content generation. To switch:
- Edit
CLAUDE.md→ change behavioral guidelines - Edit
research-pipeline-v2/references/roles.md→ change NLM prompting language
Edit research-pipeline-v2/SKILL.md Phase 1 to add new platforms. Current channels:
| Platform | Status | Access Method |
|---|---|---|
| YouTube | ✅ Full ASR | yt-dlp + --cookies-from-browser chrome |
| Bilibili | 🟡 Needs login | yt-dlp + chrome cookies |
| 知乎/CSDN | 🟡 Search snippets only | WebSearch (full text blocked by anti-scraping) |
| arXiv | ✅ Abstracts | WebFetch |
| GitHub | ✅ READMEs | WebFetch |
| NotebookLM AI Research | ✅ Full | Built-in web research |
In SKILL.md Depth Adaptation table:
- "快速了解" → 3 sub-topics, skip deep-dive
- "深度研究" → 5+ sub-topics, full pipeline
- "面试准备" → adds interview Q&A sections
┌─────────────────────────────────────────────────────────┐
│ User (Claude Code) │
│ "帮我研究 X" ──→ research-pipeline-v2 │
│ "这是什么意思" ──→ study-assist │
└────────────┬────────────────────────────┬────────────────┘
│ │
┌───────▼───────┐ ┌─────────▼────────┐
│ Research Phase │ │ Study Phase │
│ │ │ │
│ YouTube ASR │ │ Locate (ripgrep) │
│ Papers (S2) │ │ Explain/Research │
│ NotebookLM │ │ Reflect (Socratic)│
│ Web Search │ │ Create sub-doc │
└───────┬───────┘ └─────────┬────────┘
│ │
┌───────▼────────────────────────────▼────────┐
│ Obsidian Vault │
│ projects/ → Teaching notes + Quiz + Cards │
│ _deep-notes/ → Study annotations │
│ concepts/ → Atomic knowledge nodes │
│ [[wikilinks]] → Knowledge graph │
└───────────────────────────────────────────────┘
MIT License — see LICENSE.
- Claude Code by Anthropic
- NotebookLM by Google
- Obsidian
- yt-dlp
- Semantic Scholar API