Skip to content

moonlight-lupin/agent-skills / research/notebooklm-mode

notebooklm-mode

Verified

Source-grounded research pipeline inspired by NotebookLM. Collect verbatim source extracts into a temporary workspace, then answer questions and create deliverables grounded in those sources. Two grounding modes: strict (vault-only) and augmented (vault + labeled [background] knowledge). Trigger when the user asks to research a topic from sources, wants NotebookLM-style grounded answers, asks to build a resource vault, or says "notebooklm mode".

references/book-study.md

View on GitHub
# Book Study Mode

Study a **primary text** alongside one or more **reference texts** (commentaries,
study guides, criticism, teacher's manuals) using the notebooklm-mode vault
infrastructure with structured reference parsing.

## When to use

- "Study Chapter 12 of Grudem with the study guide"
- "Study John 3:16 with Calvin's commentary" (Bible — needs alias file)
- "What does the criticism say about Chapter 5 of Moby Dick?"
- "Study the theme of predestination across the primary text and commentary"
- Daily reading plan with companion commentary + bilingual text

## Vault setup

A book-study vault is a regular notebooklm-mode vault with **tagged sources**:

```
<vault_path>/
├── vault_index.md
├── sources/
│   ├── 001_primary.md              # source_role: primary
│   ├── 002_primary-alt.md          # source_role: primary (translation/edition)
│   └── 003_reference.md             # source_role: reference
├── rag_index.db
└── outputs/
    └── study_YYYY-MM-DD.md
```

### Source file format (extends notebooklm-mode)

Add a `source_role` field to the source file metadata table:

```markdown
# Source Title

| Field | Value |
|-------|-------|
| URL | ... |
| Type | book / bible / commentary / criticism |
| Source Role | primary / reference |
| Retrieved | YYYY-MM-DD |
```

The agent uses `source_role` to distinguish "the text under study" from
"the commentary on it." This is a convention — notebooklm-mode's RAG treats
all sources the same; the role only affects how the agent presents results.

### Using library-rag if books are already indexed

If the primary and reference texts are already in `~/.hermes/library/` (indexed
by library-rag), skip re-indexing into the vault. Instead:
1. Use `parse_reference.py` to resolve the exact passage from the markdown file
2. Use `mcp_library_rag_search` to find semantically related passages in the
   reference text, filtered by `source_type`
3. Compose the study output from the exact passage + semantic search results

This is the fast path — no vault setup or re-indexing needed.

### Per-session standalone RAG (for new book pairs)

If the texts are NOT in library-rag:
1. Create a vault per the notebooklm-mode workflow
2. Convert the primary and reference texts to markdown (EPUB/PDF via library-rag
   converters, or use existing markdown)
3. Add sources with `ingest_source.py`, tagging each as primary or reference
4. Use `search_vault()` for semantic search across both

## Two modes

### Passage-first mode

User provides a specific passage reference (e.g. "Ch 12", "John 3:16",
"§3.2"):

1. **Resolve the passage** — use `parse_reference.py` to extract the exact text
   from the primary source markdown:
   ```bash
   python3 scripts/parse_reference.py "Ch 12" --source /path/to/book.md
   python3 scripts/parse_reference.py "John 3:16" --source /path/to/bible.md
   ```
2. **Find commentary** — semantic search the reference text(s) for passages
   related to the primary passage:
   - If in library-rag: `mcp_library_rag_search(query="<key phrases from passage>",
     source_type="<reference-source-type>")`
   - If in vault: `search_vault(vault_path, "<key phrases from passage>")`
3. **Compose study session** — see output format below

### Topic-first mode

User provides a topic or question (e.g. "what does the commentary say about
predestination?", "study the theme of faith in Chapter 8"):

1. **Search both corpora** — semantic search primary + reference for the topic:
   - Primary: `mcp_library_rag_search(query="<topic>",
     source_type="<primary-source-type>")` for relevant passages
   - Reference: `mcp_library_rag_search(query="<topic>",
     source_type="<reference-source-type>")` for commentary
2. **Pair results** — match primary passages with their commentary
   (by chapter/section overlap or semantic similarity)
3. **Compose study session** — see output format below

## Study session output format

```markdown
<!-- Study: {reference} | Date: YYYY-MM-DD | Primary: {primary} | Reference: {reference} -->

# Study: {Reference}

**Date:** YYYY-MM-DD
**Primary text:** {primary_source_name}
**Reference:** {reference_source_name}

## Primary Text

### {Chapter/Section/Verse reference}

{Exact passage text from parse_reference.py}

{If multiple editions/translations, present the second here}

## Commentary

**{Commentator/critic name} on {reference}:**

{Relevant commentary passages, with references if available. Quote verbatim
from the reference text. Cite the source file.}

## Insight

{Synthesize the primary text and commentary into 2-3 key observations:
- What is the main theme or argument?
- What does the commentary highlight that isn't obvious from the text alone?
- How do the primary text and commentary interact?}

## Application / Reflection

{1-2 paragraphs of personal reflection:
- How does this passage speak to current situations?
- What action or attitude change does it call for?
- Keep this grounded in the text, not generic commentary.}
```

> **Note:** The `Application / Reflection` section is optional and adapts to
> the book type — a Bible study might replace it with (or add) a short prayer,
> while a literary-criticism study may end with reflection questions. Omit or
> adapt sections that don't fit the text.

## Multi-edition study

For studying a text with two editions/translations (e.g. NIV + CUV Bible,
or an original-language text + translation):
1. Resolve the passage from both editions using `parse_reference.py`
2. Present both side by side in the Primary Text section
3. Commentary is typically in one language only
4. Reflection can be in either language per user preference

## Cron / recurring study

For a daily study habit (e.g. a reading plan through a book):
1. Create a cron job that runs the study workflow on a schedule
2. The cron prompt includes the reading plan (which passage for which day)
3. Output is saved to a studies directory
4. Use `context_from` to pass the previous day's study for continuity

Example cron prompt (generic):
```
You are a book study assistant. Today's passage is {passage_ref}.
Study this passage using {reference_name}.

## Steps
1. Use parse_reference.py to get the exact text from the primary source: {passage_ref}
2. If multi-edition, resolve from the second source as well
3. Search library-rag for {reference_name} commentary on this passage
4. Compose the study session per references/book-study.md output format
5. Save to {output_dir}/Day-{N}_{slug}_{user}.md

Primary: {primary_path}
Reference: {reference_name} (search library-rag source_type={reference_type})
```

## Pitfalls

- **parse_reference.py needs the exact markdown path** — know where your
  source files live. For single-file books (e.g. a whole Bible in one .md),
  pass that file. For per-chapter files (library-rag's usual structure),
  pass the specific chapter file.
- **Book name aliases are optional** — place an alias JSON next to
  `parse_reference.py` (or pass `--aliases path.json`) to map short names to
  canonical headings (e.g. "Psalm" → "Psalms", "Rom" → "Romans"). Without
  an alias file, only title-casing is applied. Create your own for non-Bible
  books, or omit it entirely.
- **Commentary may not cover every passage** — if semantic search returns
  nothing relevant, say so — don't fabricate commentary.
- **Reference structure varies** — commentaries use inline prose references,
  not structured verse markers. Semantic search is the only reliable way to
  find relevant commentary passages.
- **Multi-edition alignment** — translations may have different chapter/verse
  or section numbering. If passages don't align perfectly, note the discrepancy.
- **Large files** — `parse_reference.py` reads the full file each call. For
  a cron doing one passage/day this is fine. For batch processing, cache the
  file content.
writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deploymentStars: 322kUpdated Claude CodeCodexVerifiedInstall
xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved.Stars: 198kUpdated Claude CodeCodexVerifiedInstall
shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy.Stars: 119kUpdated Claude CodeCodexVerifiedInstall
childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release.Stars: 118kUpdated Claude CodeCodexVerifiedInstall