notebooklm

NotebookLM research using Agent Teams — two-phase: Opus strategist plans, then right-sized team (scout + N workers + synthesizer) executes. Best for YouTube videos, podcasts, audio content, and media Claude cannot access directly.

Haiku scout for Agent Teams-based NotebookLM research. Spawned as a teammate by the notebooklm-research command. Reads the strategist's plan, finds the best YouTube videos, podcasts, and articles for each notebook's topic area via web search, and writes sources.md for workers to consume. <example> Context: Strategist has written strategy.md with search guidance for 2 notebooks. user: "Find the best YouTube videos and podcasts for the notebooks in strategy.md" assistant: "I'll read the strategy, execute searches for each notebook's topic area, vet accessibility, and write sources.md." <commentary> Scout reads strategy.md, executes searches per notebook's 'search guidance for scout' field, writes sources.md with ## Sources for Notebook A/B/C convention. Task completion unblocks workers. </commentary> </example>

Opus strategist for NotebookLM research. Pre-team planning agent — reads EM context, designs the optimal research strategy including notebook topology, worker count, question design, source strategy, and custom instructions. Writes strategy.md to disk for the EM to consume. NOT a teammate — dispatched as a regular Agent in Phase 1. <example> Context: EM has scoped a research topic and written em-context.md. user: "Design the research strategy for 'AI agent architectures' on free tier, 12 queries used today" assistant: "I'll read the EM context, assess topic breadth, and design a quota-aware strategy with 1 worker covering the topic." <commentary> Strategist reads em-context.md, applies NLM best practices and tier limits, writes strategy.md with worker_count: 1. </commentary> </example> <example> Context: EM needs a broad multi-angle investigation on Plus tier. user: "Design strategy for 'future of work and AI' — broad topic, Plus tier, 50 queries used today" assistant: "Broad topic warrants 2-3 workers. I'll design notebooks covering different angles with 7-8 questions each." <commentary> Strategist splits the topic into 2-3 notebook clusters, writes strategy.md with worker_count: 2 or 3. </commentary> </example>

Opus synthesizer for Agent Teams-based NotebookLM research. Spawned as a teammate by the notebooklm-research command. Blocked until all worker tasks complete, then reads findings from disk, cross-references across notebooks, writes the final polished research document, and deletes all notebooks. <example> Context: All workers have completed their notebooks and written findings. user: "Synthesize findings from 3 NotebookLM notebooks into a final research document" assistant: "I'll wait for all DONE messages, read the findings files, cross-reference, synthesize, and clean up the notebooks." <commentary> Synthesizer waits for DONE messages from all workers, reads {letter}-findings.md files, produces polished output, then deletes notebooks using IDs from the findings metadata. </commentary> </example>

Sonnet worker that executes NotebookLM MCP operations as a teammate in an Agent Teams research session. Blocked by the scout until sources are ready, then creates its own notebook, ingests assigned sources, runs queries, and writes findings. Sends DONE to synthesizer when complete. <example> Context: Scout has written sources.md. Worker is assigned Notebook B. user: "Execute NotebookLM research for Notebook B on 'agent evaluation frameworks'" assistant: "I'll check my task is unblocked, read strategy.md and sources.md for Notebook B, bootstrap MCP, create the notebook, ingest sources, run queries, and write findings." <commentary> Worker checks TaskList FIRST (read-after-unblock sequencing), reads its Notebook B sections from shared artifacts, bootstraps MCP tools, executes the full research pipeline, writes B-findings.md, marks task complete, sends DONE. </commentary> </example>