name: nlm-deepdive description: Deep dives into a subtopic within an existing NotebookLM notebook. Decomposes the subtopic into sequential learning units and creates per-unit infographics and video overviews in parallel, then renames videos into playlist order. Use when user says "nlm-deepdive", "deep dive into", or "drill down into". user-invocable: true

NLM Deep Dive — Subtopic Learning Unit Generator

Deep dives into a subtopic within an existing NotebookLM notebook using the NLM CLI (/Users/user/.local/bin/nlm). Decomposes the subtopic into sequential learning units and generates per-unit infographics and video overviews, then renames videos into playlist order.

Input Format

/nlm-deepdive <subtopic> [--notebook <notebook_id>]

• <subtopic> is required
• --notebook is optional; if omitted, notebooks are listed and matched by topic relevance

Workflow

Execute the following 5 phases in order. Announce each phase as you enter it.

Phase 1: Identify Notebook

If --notebook <id> is provided:

Verify the notebook exists:

nlm notebook get <notebook_id>

If it fails, list all notebooks and ask the user to pick (see fallback below).

If --notebook is NOT provided:

List all notebooks:

nlm notebook list --title

Match the subtopic against notebook titles to find the most relevant one:

• If exactly one good match → confirm with the user and proceed
• If multiple candidates or no clear match → present the top 3 options and ask the user to pick
• If no notebooks exist → tell the user to create one first (e.g., with /nlm-new-topic)

Phase 2: Topic Decomposition

Step 2a — Understand notebook content:

nlm notebook describe <notebook_id>

Step 2b — Check source count to calibrate unit count:

nlm source list <notebook_id>

Use source count to set target unit range:

• Fewer than 5 sources → 3-4 units
• 5-9 sources → 4-5 units
• 10+ sources → 5-7 units

Step 2c — Decompose the subtopic into learning units:

nlm notebook query <notebook_id> "Break the subtopic '<subtopic>' into sequential learning units of roughly equal information density. Each unit MUST be mutually exclusive — no concept should appear in more than one unit. Each unit should have a short title (3-6 words), a one-sentence description of what it covers, and a one-sentence boundary note stating what it does NOT cover (i.e., what belongs to adjacent units). Order them logically: foundational concepts first, then intermediate, then advanced. Return as a numbered list with format: 'N. Title — Description. Boundary: <what this unit excludes>'. Target count: <target_count> units."

Parse the response to extract for each unit:

• Title (used for video naming)
• Description (what the unit covers)
• Boundary (what the unit excludes)

All three are used to build the --focus parameter in Phase 4.

Phase 3: Source-to-Unit Mapping

Map notebook sources to relevant learning units so each unit's artifacts only draw from topically relevant sources. Reuse the source list already fetched in Step 2b.

If the notebook has only 1 source, skip this phase entirely — all artifacts use that single source.

Step 3a — Collect source summaries:

From the source list obtained in Step 2b, parse to get all source IDs and titles. Then run ALL in parallel:

nlm source describe <source_id> --json   # for each source

Collect id, title, summary (truncate to ~100 chars), and keywords (first 5). If source describe fails for a source, use only its title in the mapping prompt.

Step 3b — Map sources to units:

nlm notebook query <notebook_id> "Given these learning units:
1. <title> — <description>
2. <title> — <description>
...
And these sources:
- <source_id> '<title>': <summary_snippet> [keywords: kw1, kw2]
- <source_id> '<title>': <summary_snippet> [keywords: kw1, kw2]
...
For each learning unit, list ONLY the source IDs directly relevant to that unit's specific topic. A source may appear in multiple units. Format:
Unit 1: <id>, <id>, ...
Unit 2: <id>, <id>, ..."

Step 3c — Parse the mapping:

Build map: unit_number → [source_id1, source_id2, ...]

Fallback rules:

• If a unit maps to 0 sources → use ALL sources for that unit (omit --source-ids)
• If the mapping query returns unparseable output → use ALL sources for all units (degrades gracefully to current behavior)

Phase 4: Per-Unit Artifacts (Parallel)

Fire ALL per-unit artifact creation commands in parallel (maximize throughput):

For every learning unit, fire both commands simultaneously using an enriched focus prompt that scopes content and prevents overlap with neighboring units. Include --source-ids from the Phase 3 mapping to restrict each artifact to relevant sources only:

nlm infographic create <notebook_id> --focus "Unit <N> of <total>: <unit title>. <unit description>. Does NOT cover: <boundary>." --source-ids <id1>,<id2>,<id3> -y
nlm video create <notebook_id> --focus "Unit <N> of <total>: <unit title>. <unit description>. Does NOT cover: <boundary>." --source-ids <id1>,<id2>,<id3> -y

Build the focus string from the parsed decomposition output (title + description + boundary). Use the source IDs from the Phase 3 mapping for each unit. If a unit has no mapped sources (fallback), omit --source-ids entirely for that unit. For example, with 5 units this fires 10 commands in parallel.

Error handling — retry with backoff:

If any command fails, retry up to 3 times with increasing backoff:

1. Retry 1: Wait 30 seconds, then retry
2. Retry 2: Wait 60 seconds, then retry
3. Retry 3: Wait 120 seconds, then retry

If a command still fails after all retries and the error contains "Try again later":

• Stop retrying that artifact type (e.g., all remaining infographics)
• Tell the user which artifacts failed and that the cause is likely a daily creation limit imposed by NotebookLM on their account tier
• List the exact commands the user can run manually later (tomorrow) to create the missing artifacts
• Continue with any other artifact types that are still succeeding

Phase 5: Verify, Order & Report

Step 5a — Check artifact status:

nlm studio status <notebook_id> --json

Poll up to 3 times with 60-second intervals until all new artifacts show complete status.

Step 5b — Determine deep dive sequence prefix:

From the studio status --json output in step 5a, inspect existing video names for the pattern DD<number>-. Find the highest existing deep dive number and increment by 1. If no existing deep dive sequences are found, start at 1.

• No existing DD*- videos → prefix is DD1-
• Existing DD1- videos → prefix is DD2-
• Existing DD1- and DD2- videos → prefix is DD3-

Step 5c — Rename videos in sequence order:

Use nlm studio rename to create a numbered playlist with the determined prefix:

nlm studio rename <unit1_video_id> "DD<X>-01 - <unit1 title>"
nlm studio rename <unit2_video_id> "DD<X>-02 - <unit2 title>"

Where <X> is the deep dive sequence number from step 5b. Continue for each unit video in learning order. The DD<X>- prefix (Deep Dive #X) distinguishes these from the original new-topic sequence (01, 02, ...) and from other deep dive sequences.

Step 5d — Present final report:

## Deep Dive Complete: <subtopic>

**Notebook ID:** <notebook_id>
**Notebook URL:** https://notebooklm.google.com/notebook/<notebook_id>

### Learning Units (Video Playlist Order)
- **DD<X>-01 - <unit1 title>** — Infographic + Video (<N> sources)
- **DD<X>-02 - <unit2 title>** — Infographic + Video (<N> sources)
- ...

### Failed Artifacts (if any)
- <list any failures>

### Download Commands
nlm download video <notebook_id>
nlm download infographic <notebook_id>

Error Handling

Key NLM CLI Commands Used