learning-lab

A skill-based learning pipeline for Claude Code. Feed it any source — YouTube video, web page, PDF, or local file — and it produces a structured learning package. Final knowledge gets exported to an Obsidian vault.

NotebookLM is the primary analysis engine. Heavy AI analysis runs on Google at zero token cost. Claude Code orchestrates the pipeline and formats the output.

---

How It Works

Source (YouTube / web / PDF / text)
        │
        ▼
  [learn-source]
        │  1. Detect & ingest (yt-dlp / defuddle / Read)
        │  2. Send to NotebookLM (create → add source → generate)
        │  3. Structure output into 8-file learning package
        │  4. Run adversarial logic check
        ▼
  workspace/{slug}/          ← German, working drafts
        │
        ├── [fill-gaps]        →  Research answers for open questions
        ├── [rebuild-project]  →  projects/{slug}/    (optional, tutorials only)
        │
        └── [save-to-vault]    →  The Vault/research/{Note}.md  (English, final)

---

Skills

Skill	Trigger	What it does
lab-master	"lab run", /lab --auto <url>	Master skill that chains all skills. --auto = zero interaction.
learn-source	URL or file path	Ingest → NLM analysis → 8-file learning package
fill-gaps	"fill gaps", "fülle Lücken"	Web search for open questions → integrate answers
rebuild-project	"rebuild", "baue nach"	Build a minimal MVP from a tutorial source
save-to-vault	"save to vault", "speichere im Vault"	Export to Obsidian vault with full asset bundle

---

Learning Package (8 files, German)

Each source produces a workspace/{slug}/ folder:

File	Contents
00_zusammenfassung.md	Executive summary + core thesis
01_kernkonzepte.md	Key concepts with definitions
02_schritt_fuer_schritt.md	Step-by-step walkthrough
03_uebungen.md	Flashcards and exercises
04_projekt_rebuild.md	Rebuild blueprint (tutorial sources only)
05_offene_fragen.md	Open questions and gaps
06_notebooklm_artefakte.md	NLM deliverable paths and notebook info
07_logik_check.md	Bias analysis, missing counter-arguments

---

Source Types

Type	Tool
YouTube URL	yt-dlp → transcript → NLM
Web page URL	defuddle → clean markdown → NLM
PDF file	Claude Code Read → NLM
Local .md / .txt	Claude Code Read → NLM

---

Prerequisites

pip install yt-dlp notebooklm
npm install -g defuddle
notebooklm login   # browser OAuth — run once in a separate terminal

# Token-efficiency tools:
cargo install --git https://github.com/rtk-ai/rtk  # or download rtk binary
winget install Buzz  # offline whisper transcription

---

Repository Layout

learning-lab/
├── .claude/
│   ├── settings.json          # vault_path + tool permissions
│   └── skills/
│       ├── lab-master/
│       ├── learn-source/
│       ├── fill-gaps/
│       ├── rebuild-project/
│       └── save-to-vault/
├── docs/
│   ├── vault-format-reference.md
│   └── workflow-spec.md
├── scripts/
│   ├── run_state.py           # run.json state machine
│   └── vault_sync.py          # Vault asset export + migration
├── tests/                     # pytest suite for scripts
├── sources/        # gitignored — raw transcripts and NLM output
├── workspace/      # gitignored — 8-file learning packages
├── projects/       # gitignored — MVP project rebuilds
├── CLAUDE.md
└── README.md

---

Vault Integration

Polished notes are exported to a separate Obsidian vault. Each export produces:

• research/{Note-Title}.md — English research note with full YAML frontmatter and wiki-links
• research/assets/{source_type}/{slug}/Source Bundle.md — index of all source artifacts
• research/assets/{source_type}/{slug}/transcript.txt — raw source transcript
• research/assets/{source_type}/{slug}/study-guide.md — NLM study guide
• daily-notes/{YYYY-MM-DD}.md — daily note entry (marker-based, deduped)

The vault path is configured in .claude/settings.json → vault_path.

Assets are organized by source type (youtube/, web/, pdf/) to avoid a flat graveyard in research/assets/.

---

Language Convention

• Workspace files — German (working drafts)
• Vault research notes — English (final, analytical)
• Config and docs — English