Research Pipeline — Token-Efficient Deep Research

Simple rule: Sources in hand → this skill. No sources → /deep-research. Legal document → /legal-assistant.

Architecture

Tier 1: NotebookLM (FREE)     → Ingest sources, initial analysis
Tier 2: Sub-agents (CHEAP)    → Validate claims, find gaps, check contradictions
Tier 3: Orchestrator (LEAN)   → Summaries only, final synthesis

Process

Phase 1: Research Prompt Architecture (Orchestrator — use /prompter)

Before sending anything to NotebookLM, architect a structured research brief. Vague prompts produce vague results.

1. Invoke /prompter to refine the user's research question into:

   • 3-5 specific sub-questions to investigate
   • Scope boundaries — what to include and exclude
   • Angles to cover — technical, market, competitive, regulatory, etc.
   • What a good answer looks like — expected output shape

2. Present the structured research brief to the user for approval before proceeding.

3. Once approved, use this brief (not the original vague question) as the input to NotebookLM deep research in Phase 2.

This step is what separates useful research from generic summaries. Claude is the research strategist; NotebookLM is the research executor.

Phase 2: Source Ingestion (NotebookLM — FREE)

This phase uses the /notebooklm skill's CLI commands. See that skill for full command reference.

1. CREATE + USE + VERIFY — Create a NotebookLM notebook and explicitly set it as active:

   # Create and capture UUID
   NOTEBOOK_ID=$(notebooklm create "Research: [Topic]" 2>/dev/null | grep -oE '[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}')
   # Explicitly set active notebook (create does NOT do this reliably)
   notebooklm use "$NOTEBOOK_ID"
   # Verify correct notebook is active
   notebooklm metadata --json 2>/dev/null | python3 -c "import sys,json; d=json.load(sys.stdin); print(f'Active: {d[\"title\"]}')"
   

   CRITICAL: create does NOT reliably set the active notebook. Without use, all subsequent commands silently target the PREVIOUSLY active notebook.

2. ALL subsequent NotebookLM commands MUST run in the SAME foreground session. Background processes (run_in_background: true) do NOT inherit the active notebook — they will silently target the wrong notebook or fail.

3. Add specific sources FIRST (web URLs, PDFs, local files):

   notebooklm source add "URL or file"   # Add known sources one at a time
   

   Add 3-6 specific high-quality sources before running deep research.

4. Run NotebookLM deep research (LONG operation — 2-10 minutes):

   notebooklm source add-research "topic keywords" --mode deep --import-all
   

   CRITICAL: Run in FOREGROUND with timeout 600000ms (10 min). Known issues:

   • Background mode produces empty output (0 bytes) or fails silently
   • Background processes lose the active notebook context, importing into the wrong notebook
   • IMPORT_RESEARCH RPC can fail with null result data — retry with research wait --import-all Always use --mode deep unless user requests fast mode.

5. VERIFY sources were added before proceeding:

   notebooklm source list
   

   If source count is 0, the deep research failed. Retry or fall back to /deep-research.

6. Ask NotebookLM key research questions (save output to temp file):

   notebooklm ask "What are the key findings and arguments?" > /tmp/nlm-findings.md
   notebooklm ask "What contradictions or debates exist?" >> /tmp/nlm-findings.md
   notebooklm ask "What evidence supports the main claims?" >> /tmp/nlm-findings.md
   

7. Read the temp file — this is NotebookLM's analysis (already summarized, small token footprint).

Fallback if NotebookLM fails:

If NotebookLM auth fails, sources can't be added, or deep research returns empty:

1. Auth failure (notebooklm login needed): Tell the user to run notebooklm login in a separate terminal (it requires interactive browser auth that Gemini CLI cannot do). Wait for them to confirm, then retry. If still failing, fall back to Step 2 below.
2. Fall back to sub-agent research: Skip Phase 2 entirely. Instead, spawn 3 research sub-agents (using the Agent tool) that each take 1-2 sub-questions from Phase 1's brief and research via WebSearch + WebFetch. Each returns findings to a temp file. Continue to Phase 3 (Gap Analysis) using sub-agent outputs instead of NotebookLM outputs.
3. Source add failures: Verify with notebooklm source list. If sources are missing, try adding them one at a time. URL sources are more reliable than file uploads.
4. Deep research returns 0 sources: The --import-all flag silently fails in background mode. Re-run in foreground with timeout: 600000. If still empty, use option 2 above.

Phase 3: Gap Analysis (Orchestrator — LEAN)

Read the NotebookLM analysis (already summarized — small token footprint). Identify:

• Claims that need independent verification
• Missing perspectives or counter-arguments
• Sources that need authority/recency checks
• Gaps in the analysis

Plan 2-4 targeted validation tasks for sub-agents.

Phase 4: Parallel Validation (Sub-agents — ISOLATED CONTEXT)

Spawn sub-agents using the Agent tool. Each gets a specific, narrow task.

See references/example-run.md for a complete end-to-end example showing all 6 phases with concrete BARMM content.

Required sub-agent output format — each agent MUST structure output as:

## Summary
[3-5 bullet findings — this is what the orchestrator reads]

## Evidence
[detailed sources, quotes, URLs — orchestrator skips this unless needed]

Agent 1: Claim Verifier

• Take the top 5-10 claims from NotebookLM's analysis
• Independently verify each via web search
• Return: verified/unverified status + supporting sources

Agent 2: Counter-Argument Finder

• Search for opposing viewpoints and contradictions
• Return: counter-arguments + sources that challenge the main findings

Agent 3: Source Authority Checker

• Evaluate the credibility and recency of key sources
• Return: source quality assessment + any outdated information flagged

Each sub-agent writes findings to a temp file. The orchestrator reads only the Summary section of each file — never the full Evidence output.

Phase 5: Synthesis (Orchestrator — LEAN)

1. Read sub-agent summaries (not full outputs)
2. Combine NotebookLM analysis + validation findings
3. Classify claims with confidence levels:
   • [S] Supported — verified by multiple independent sources
   • [P] Probable — supported but not independently verified
   • [?] Uncertain — conflicting evidence or single source
   • [U] Unsupported — no verification found or contradicted
4. Produce the final research brief

Phase 6: Output

Save to Obsidian vault:

1. Full report → ~/Vault/research/deep-research/yymmdd-topic-name.md

---