Notebooklm Book Pipeline

NotebookLM Book Pipeline

Split a full book into high-signal parts, feed each part into NotebookLM as an independent source, and generate reusable material packs.

Workflow

1. Split source book into parts:

python3 skills/notebooklm-book-pipeline/scripts/split_book_for_notebooklm.py \
  --book "/path/to/book.epub" \
  --target-parts 4 \
  --out-dir state/runtime/book_pipeline/hard-thing

This command creates:

• chunks/*.md
• split_manifest.json

2. Generate materials from NotebookLM:

python3 skills/notebooklm-book-pipeline/scripts/run_notebooklm_book_pipeline.py \
  --split-manifest state/runtime/book_pipeline/hard-thing/split_manifest.json \
  --notebook-id "<notebook_id>" \
  --artifacts report,quiz,flashcards,video \
  --output-dir state/runtime/book_pipeline/hard-thing/outputs

Re-run only failed parts:

python3 skills/notebooklm-book-pipeline/scripts/run_notebooklm_book_pipeline.py \
  --split-manifest state/runtime/book_pipeline/hard-thing/split_manifest.json \
  --notebook-id "<notebook_id>" \
  --artifacts report,quiz,flashcards \
  --part-ids P02,P03 \
  --rpc-retries 4

3. Test iteratively before real run:

python3 skills/notebooklm-book-pipeline/scripts/run_notebooklm_book_pipeline.py \
  --split-manifest state/runtime/book_pipeline/hard-thing/split_manifest.json \
  --notebook-id "<notebook_id>" \
  --artifacts report,quiz,flashcards \
  --limit-parts 1 \
  --dry-run

Use --dry-run first, then remove it for real generation.

Output Contract

• Split step prints:
  • RUN_ID
  • BOOK_TITLE
  • CHAPTERS
  • PARTS
  • OUT_DIR
  • MANIFEST
• Pipeline step prints:
  • RUN_ID
  • STATUS
  • NOTEBOOK_ID
  • PARTS_TOTAL
  • PARTS_FAILED
  • OUTPUT_DIR
  • RUN_MANIFEST

Treat run_manifest.json as the single source of truth for per-part status.

Operational Rules

• Keep each part around 12000 to 60000 chars by default; the splitter auto-widens upper bound when --target-parts is small.
• Use one source per part for better attribution and error isolation.
• Generate report first; add video only after report quality passes.
• On failures, rerun with --limit-parts 1 for the failed part range.
• Prefer targeted retry with --part-ids and increased --rpc-retries when NotebookLM RPC is unstable.

References

• Prompt patterns: references/prompt_templates.md
• Quality gates: references/quality_gates.md