extract-notes

name	extract-notes
model	sonnet
description	Deep-research a book and propose video angles
disable-model-invocation	false
argument-hint	<book-slug>

Extract Book Notes

Research a book deeply using NotebookLM as a center hub, then propose video angles for selection.

The NotebookLM notebook is the research center hub for each author — not a book dump. Every book gets surrounded by wiki, academic criticism, adaptations, Vietnamese context, reader discussions, and competitive landscape. The richer the hub, the sharper the angles.

Context

• When to use: New book ready — need to extract insights and choose video angle.
• Input: Book slug + primary sources (text, URL, or file)
• Output: notes.md (structured notes + chosen angle)
• Reference: See projects/ai-book-video/books/atomic-habits/notes.md for the standard format.
• MCP tools: NotebookLM MCP (notebook_create, source_add, notebook_query, chat_configure)

Steps

Step 0: Vault Context

Before any research, check what the Knowledge Vault already knows. This surfaces cross-book connections and prevents redundant angles.

1. Read projects/ai-book-video/knowledge-base/library.md — what books/authors are already covered
2. Check if projects/ai-book-video/knowledge-base/authors/<author-slug>.md exists for this book's author
   • If yes: read it — we already know their thesis, strengths, blind spots
3. Glob for projects/ai-book-video/knowledge-base/concepts/*.md — scan theme files for overlapping concepts
4. Read relevant concept files where this book's themes might connect
5. Read projects/ai-book-video/knowledge-base/connections/ files (contradictions, agreements, evolutions) for existing cross-book tensions

Vault summary (include in notes.md later):

• "We've covered N books. Author [X] is [known/new]. Overlapping themes: [list]. Existing connections: [list]."
• If vault is empty (first book), note: "First book in vault — no cross-references yet."

Step 0b: Cross-book Query

If the "Bookie: Library" Master notebook exists in NotebookLM:

1. notebook_list → find "Bookie: Library" notebook
2. If found, notebook_query: "What themes from previously added books relate to [this book's topic]? What contradictions or tensions might arise?"
3. Present findings alongside vault context — these inform angle selection later

If Master notebook doesn't exist or query fails, skip — vault files are source of truth.

Step 1: Validate

Check that $ARGUMENTS is provided. If missing, ask Hai for the book slug (e.g., "atomic-habits"). Set SLUG=$ARGUMENTS. Check that projects/ai-book-video/books/$SLUG/ exists.

Step 2: Check existing notes

Read projects/ai-book-video/books/$SLUG/notes.md. If it already has content beyond the scaffold template, warn and ask Hai before overwriting.

Step 3: Build Research Center Hub

This is the most important step. The NotebookLM notebook becomes a research hub that enables deep cross-referencing between the book, its author, scholarly analysis, cultural context, and audience reception.

3a. Create notebook + add primary book text

• notebook_create with name "Bookie: [Author Name]" (per-author convention — multiple books by same author share 1 notebook)
• If author already has a notebook from a previous book, use that notebook instead of creating a new one
• Check projects/ai-book-video/books/$SLUG/ for existing text files (.txt, .pdf, .md). If found, source_add (type=file) for each — no need to ask Hai.
• If no text files found in the folder, ask Hai: "Book text chưa có trong folder. Cung cấp file path, URL, hoặc paste text?"

3b. Enrich: 7 source categories

Systematically search and add sources across all categories. Run searches in parallel where possible.

Read references/search-categories.md for Categories 1-6 (Wikipedia, Academic, Adaptations, Vietnamese Context, Reader Reviews, Author Deep Dive) with search queries and targets per category.

Read references/youtube-search-patterns.md for Category 7 (YouTube Analysis) with yt-dlp commands and selection criteria. Save the yt-dlp search metadata — reused in Step 4 for competitive analysis.

3c. Verify hub completeness

After enrichment, verify:

• Minimum 15 total sources (primary + reference). If under 15, search harder.
• At least 5 of 7 categories have sources. If a category is empty, do one more targeted search.
• Vietnamese sources present — this is non-negotiable for Bookie's audience.
• YouTube sources present — at least 1 YouTube video added. If no analysis videos exist for this book, note that as a competitive gap.

Report the hub status to Hai: "Notebook has N sources across K categories: [breakdown]."

3d. Configure chat lens

Configure chat with Bookie analytical lens via chat_configure:

Bạn là trợ lý nghiên cứu sách cho kênh Bookie (Việt Nam). Khi phân tích:
1) Tìm mâu thuẫn với các cuốn sách đã thêm trước đó
2) Phát hiện thiên lệch văn hóa (Western vs Vietnamese)
3) Tìm patterns xuyên suốt nhiều nguồn
4) Gợi ý kết nối bất ngờ giữa các tác giả/ý tưởng
5) Đánh giá tính ứng dụng trong bối cảnh Việt Nam (20-35 tuổi, self-improvement)
Trả lời bằng tiếng Việt.

3e. Deep queries

Now that the hub has rich context, query for structured insights. Read references/notebooklm-queries.md for the base query set and conditional multi-book queries. Detect the book configuration (single book, twin/multi-book same author, multi-book different authors) and run the matching query set.

If any MCP tool fails, fall back: tell Hai to use NotebookLM web UI (notebooklm.google.com) and paste results manually.

Step 4: Competitive analysis

Two sources: yt-dlp for YouTube landscape, WebSearch for non-YouTube (blogs, podcasts, courses).

4a. YouTube landscape (yt-dlp)

Reuse the yt-dlp search metadata from Category 7 (Step 3b). If you need broader coverage, run additional searches:

# Broader search — top 10 results
yt-dlp "ytsearch10:[Vietnamese book title] phân tích review tóm tắt" --dump-json --flat-playlist 2>/dev/null | jq -r '[.title, .channel, .view_count, .duration, .id] | @tsv'

If multi-book video: also search for comparison content ([book A] vs [book B], [author] hai cuốn).

From the results, note:

• Which angles have been done (summary, review, criticism, comparison)
• View counts and engagement (actual numbers from yt-dlp — not estimates)
• Top channels covering this book/genre
• Gaps — what angles are MISSING
• Whether anyone has done a similar multi-book or author portrait angle

4b. Non-YouTube competition (WebSearch)

Search multilingual — same principle as YouTube. Foreign blog posts, English literary essays, and original-language criticism can reveal angles absent from Vietnamese content.

Use WebSearch across languages:

Vietnamese:

1. "[Vietnamese book title]" review phân tích blog
2. "[Vietnamese book title]" podcast cảm nhận

English: 3. "[English book title]" analysis essay blog 4. "[author]" "[English book title]" literary criticism

Original language (if applicable): 5. "[Original title]" analyse critique blog

Note angles and depth — written content often goes deeper than YouTube summaries. Foreign-language articles with unique angles are especially valuable as potential Bookie adaptations.

Step 5: Structure notes

Read references/notes-template.md for the full output format. The notes should reflect the depth of the research hub — not just book text, but academic insights, author bio, Vietnamese context, and adaptations. See projects/ai-book-video/books/atomic-habits/notes.md as a reference example.

Step 6: Propose angles

Based on notes + competitive analysis + vault context, propose 3 ranked video angles. All Vietnamese output (working titles, hooks, key points) MUST use proper diacritics (có dấu). Each angle:

• Working title (Vietnamese)
• Hook — opening sentence that creates curiosity in 3 seconds
• Core message — 1 sentence summary of what the video is about
• 2-3 key points to cover
• Why this angle works — audience appeal + differentiation from existing videos
• Recommended template — which narrative template from templates/narrative-templates.md fits this angle (Contrarian Analysis, Hidden Connection, Meta-Pattern, Author Portrait, The Tension)

Prioritize angles that:

• Solve a specific pain point (not generic book summary)
• Have stories/examples that visualize well
• Are focused enough for 5-8 minutes
• Differentiate from existing YouTube VN coverage

Vault-enabled angles: After generating angles, cross-reference templates/narrative-templates.md and vault state:

• If vault has entries from other books on overlapping themes → highlight "Hidden Connection" option
• If vault has 3+ books on same theme → highlight "Meta-Pattern" option
• If vault has author profile with 2+ books → highlight "Author Portrait" option
• If vault has high-tension contradictions entry → highlight "The Tension" option
• Always note which templates are currently ACTIVATED by vault state

Step 7: Decision point

Present 3 angles with template recommendations. Ask Hai to choose, modify, or suggest a different angle.

Step 8: Write output

After angle selection, append "## Chosen Angle" section with the selected angle details + recommended template. Write to projects/ai-book-video/books/$SLUG/notes.md

Step 9: Next step

Tell user to run /create-storyboard $SLUG (storyboard will use the recommended template as scaffold).

Fallback

If NotebookLM MCP tools are unavailable or fail:

1. Tell Hai to create a notebook manually at notebooklm.google.com
2. Ask Hai to paste the raw notes/insights
3. Continue from step 6 (structuring) onward — the skill still handles analysis, angle proposal, and file writing