Mcp Notebooklm Orchestrator

What this skill does NOT do (enforced)

• ❌ Create new notebooks
• ❌ Upload sources
• ❌ Generate audio overviews
• ❌ Create video overviews, mind maps, or Studio content
• ❌ Write to notebooks

Any attempt to create artifacts must be redirected — tell the user to go to the NotebookLM web app. This skill is read-only on content; it may only edit metadata.

Query procedure

Step 1 — Check cached findings first (saves tokens)

mem_search(query: "notebooklm/{topic-guess}", project: "{project}")
  → if found, read the observation
  → if the finding is fresh (< 7 days old for living docs, < 30 days for frozen), use it
  → if stale or missing, proceed to step 2

Step 2 — List notebooks to find the right one

notebooklm_list_notebooks()
  → returns array of {id, title, description, tags}
  → pick the best match by title/tags/description

If none match the topic, fall through to Step 2 of research routing (local code + docs).

Step 3 — Query the notebook

notebooklm_query(
  notebook_id: "{id}",
  query: "{user question, verbatim or paraphrased}"
)
  → returns answer + source citations

Step 4 — Persist the finding

mem_save(
  title: "notebooklm/{notebook}/{topic}",
  topic_key: "notebooklm/{notebook}/{topic}",
  type: "external-research",
  project: "{project}",
  content: "Q: {question}\nA: {answer}\nSources: {citations}\nNotebook: {notebook-id}\nDate: {iso-date}"
)

Next session asking the same question hits the Engram cache instead of re-querying NotebookLM.

Odoo-specific instruction pattern

When the active overlay is Odoo (detected via .atl/overlays/odoo-*/), add a code-first constraint to every NotebookLM query:

Query prefix: "Answer with code-first examples from the Odoo source. Quote model
names, field names, and decorators verbatim. Cite the file path (addons/x/models/y.py)
when possible."

This prevents NotebookLM from returning marketing-style prose when the user wants code.

When NotebookLM is NOT the right tool

• The question is about THIS repo's actual code — use ripgrep, not NotebookLM.
• The question is about a framework's public API (React, Go stdlib, Django) — use Context7 directly.
• The notebook is empty or the notebooks list returned nothing — fall through to local code + docs.

Do NOT pad NotebookLM queries with filler in hopes of getting a hit. If the topic doesn't match, move on.

Caching contract

Every NotebookLM response gets persisted in Engram. The topic-key format is:

notebooklm/{notebook-slug}/{topic-slug}

Before calling NotebookLM, ALWAYS check Engram first with mem_search. The orchestrator pays for NotebookLM calls; Engram is free.

Staleness rules:

• Living docs (team conventions, onboarding): re-query every 7 days
• Frozen docs (version-locked Odoo upgrade notes): re-query every 30 days
• Research queries (one-off findings): never re-query — Engram is authoritative

Return envelope (in sub-agent result)

{
  "source": "notebooklm",
  "notebook_id": "...",
  "query": "...",
  "answer_summary": "2-3 sentences",
  "citation_count": N,
  "engram_key": "notebooklm/{nb}/{topic}",
  "cached_hit": true|false  // was it served from Engram?
}

The orchestrator uses cached_hit to compute the token-savings banner (see Phase E of V3.1 plan).

Failure modes

• NotebookLM MCP not available → return {source: "notebooklm", status: "unavailable"}, orchestrator falls through to local research.
• No matching notebook → return {source: "notebooklm", status: "no-match"}, orchestrator falls through.
• Rate limit → return {source: "notebooklm", status: "rate-limited", retry_after_s: N}, orchestrator waits or falls through.

See also

• _shared/research-routing.md — the 4-step priority order
• mcp-context7-skill/SKILL.md — tertiary research, defers to this skill
• ripgrep/SKILL.md — secondary (local) research