Content — research + course pipeline

A self-contained repo that bundles:

• Claude Code skills for end-to-end YouTube research (search → transcribe → structure → NotebookLM → synthesize).
• A worked example: a full Playwright course (slides + video + narrated deck) distilled from 15 community tutorials.
• Seed sources used to bootstrap the pipeline design.

Everything needed to reproduce the example — or run the pipeline on a new topic — is tracked here.

---

Repo layout

Content/
├── .claude/
│   ├── skills/                      # Claude Code skills (project-scoped)
│   │   ├── youtube-research-pipeline/   # orchestrator (/youtube-research-pipeline)
│   │   ├── youtube-research-from-urls/   # orchestrator for specific URLs (/youtube-research-from-urls)
│   │   ├── youtube-transcript/          # captions + whisper fallback
│   │   ├── transcript-to-markdown/      # structure raw transcripts
│   │   ├── notebooklm/                  # NotebookLM CLI wrapper
│   │   └── research-to-course/          # bundle → curated reveal.js deck
│   ├── setup.sh                     # recreates Python venvs from requirements
│   └── settings.local.json          # per-clone settings (gitignored)
├── playwright-beginners-course/     # the example output (reveal.js deck + assets)
│   └── README.md                    # how to render and present the deck
├── sources/                         # seed material (design notes, reference docs)
├── .gitignore
└── README.md                        # (this file)

---

Prerequisites

The skills shell out to two binaries and pull Python deps via uv:

• uv — for fast venv + pip installs brew install uv (macOS) or curl -LsSf https://astral.sh/uv/install.sh | sh
• ffmpeg — required only when transcribing videos without captions (whisper path) brew install ffmpeg (macOS) or apt install ffmpeg
• Claude Code — to invoke the skills

---

First-time setup

git clone <this-repo>
cd Content
./.claude/setup.sh

setup.sh creates two project-local venvs:

• .claude/skills/youtube-transcript/.venv/ — yt-dlp, faster-whisper (~190 MB)
• .claude/skills/notebooklm/.venv/ — notebooklm-py (~140 MB)

Both are gitignored. Re-running the script is safe and upgrades in place.

---

Running the pipeline

Open the repo in Claude Code. From any conversation:

Research by subject (discover videos)

/youtube-research-pipeline <subject>, top <N>, save to <output-path>

Example:

/youtube-research-pipeline Kubernetes operators, top 15, save to ./k8s-operators

Research from specific URLs

/youtube-research-from-urls --subject "My Research Topic" <url1> <url2> ..., save to <output-path>

Or with a JSON file:

/youtube-research-from-urls --file urls.json, save to ./my-research

The pipeline will:

1. Search & rank — yt-dlp pulls ~30 candidates, ranks by views + 10 × likes, keeps top N
2. Transcribe — captions when available, faster-whisper (tiny) otherwise
3. Structure — each transcript becomes a well-sectioned Markdown doc
4. Upload to NotebookLM — all docs become sources in a fresh notebook
5. Synthesize — Claude reads the whole corpus and writes a cross-cutting summary
6. Generate deliverables — NotebookLM produces a slide deck + video overview

Every stage is idempotent; re-invoking with the same output path resumes from the first incomplete stage.

See playwright-beginners-course/README.md for a complete worked example — the 15-video Playwright research bundle that shipped with this repo.

Optional follow-up: build a presentable course deck

The pipeline output is a research bundle (transcripts, master summary, NotebookLM artifacts) — useful for reference but not a polished presentation. To turn the same corpus into a hand-curated reveal.js slide deck (~120-180 slides), invoke the research-to-course skill:

"Turn the markdown files in ./playwright-beginners-course/ into a reveal.js course deck."

You'll be asked four questions up-front (audience, depth, language scope, adjacent topics), then Claude synthesizes the corpus into a single course.md plus any SVG diagrams. Render with npx reveal-md <bundle>/course.md --watch.

Kept separate from the pipeline because the synthesis is heavy (~1 hour of model time) and not every research project warrants a full deck.

---

Skill overview

Skill	Invoked as	What it does
youtube-research-pipeline	/youtube-research-pipeline	Top-level orchestrator; searches YouTube by subject and chains the other skills
youtube-research-from-urls	/youtube-research-from-urls	Orchestrator for specific YouTube URLs provided by the user
youtube-transcript	(internal)	Fetches YouTube captions; whisper-transcribes if none
transcript-to-markdown	(internal)	Structures raw transcripts into readable Markdown
notebooklm	(internal + CLI)	Wraps the notebooklm-py CLI for notebook management
research-to-course	(intent-based)	Curates a research bundle into a reveal.js course deck

The internal skills are directly invokable too — check each skill's SKILL.md for standalone usage.

---

Design notes

• Portable paths. All skills reference ${CLAUDE_PROJECT_DIR} (set by the Claude Code harness) instead of $HOME. Clone this repo anywhere and it works without editing.
• Python deps via uv. venvs are recreated from requirements.txt rather than committed. setup.sh is the single entry point.
• Claude Code discovery. Skills in .claude/skills/ are auto-discovered by Claude Code when working in this directory. Project copies take precedence over anything in ~/.claude/skills/.
• Gitignored state. Pipeline artifacts (.notebook-id, .tasks.json, errors.log) and Claude Code local state (settings.local.json, statsig/) are excluded from the repo. The output folders themselves (e.g. playwright-beginners-course/) are intentionally committed — they are the research deliverables.

---

License & attribution

Source transcripts are derivative works of the original YouTube creators — credit them when reusing content. The skill code, pipeline design, and synthesized write-ups in this repo are original work.