Zero-Token Research: Delegating Claude Code's Heaviest Work to NotebookLM
The single change that cut my Claude Code token usage by 49%: stop using Claude to read documents. Delegate that work to NotebookLM.
Reading 3 files simultaneously in Claude costs ~150K tokens. The same work in NotebookLM costs ~5K tokens and returns a structured summary. That's a 30x difference.
The Token Cost Breakdown
| Operation | Claude Cost | After NotebookLM Delegation |
|---|---|---|
| Read 3+ files simultaneously | ~150K tokens | ~5K tokens |
| Analyze a URL | ~60K tokens | ~2K tokens |
| Research 21 competitors | ~80K tokens | ~3K tokens |
| Survey an entire document set | ~100K tokens | ~4K tokens |
Claude Code excels at judgment, integration, and code generation. It's wasteful for retrieval. The goal is to push all retrieval work outside Claude's context.
Basic Workflow
# 1. Create a notebook
notebooklm create "Competitor Analysis 2026-Q4"
# 2. Add sources (files / URLs / YouTube)
notebooklm source add "./docs/competitor-reports/2026-10.md"
notebooklm source add "https://example.com/saas-report"
# 3. Query
notebooklm ask "What pricing patterns do the 21 competitors share?"
# 4. Generate artifacts (processed on Google's infrastructure, free)
notebooklm generate slide-deck "Summarize key insights as slides"
notebooklm generate audio "deep dive focusing on key findings" --wait
Claude only consumes tokens for the one-line query. NotebookLM handles all the processing.
Web Deep Research: Autonomous Investigation
# NotebookLM autonomously searches the web and builds a report
notebooklm source add-research "advanced Flutter Web performance optimization 2026"
notebooklm research wait
notebooklm ask "Summarize the findings"
add-research lets NotebookLM investigate hundreds of pages autonomously. Claude Code is idle while this runs.
DBS Framework: Research → Skill Conversion
Don't throw away research findings — convert them into Claude Code skills:
D (Direction) = Decision trees, procedures, error recovery → SKILL.md core
B (Blueprints) = Templates, classification rules → support files
S (Solutions) = API calls, deterministic code → scripts
Example: the t1-blog-dispatch skill in this project was built by researching blog-publish.yml behavior in NotebookLM, then applying DBS to create a reusable SKILL.md.
Monthly Token Reduction
| Month | Claude tokens | Delegated to NotebookLM | Savings |
|---|---|---|---|
| Jan 2026 | 800K | 0K | 0% |
| Feb 2026 | 750K | 150K | 20% |
| Mar 2026 | 500K | 350K | 44% |
| Apr 2026 | 420K | 400K | 49% |
The 50% reduction target was hit by April.
Master Brain: Externalizing Long-Term Memory
NotebookLM also serves as a persistent "Master Brain" across Claude sessions:
notebooklm use ea6cff25-574d-4b8b-ad72-ab47cf1ed01f # project notebook
notebooklm source add "./memory/project_20260428.md" # accumulate session summaries
notebooklm ask "What architecture decisions failed in the past?"
Claude's context window resets between sessions. NotebookLM's index doesn't. Decisions, failures, and rationale survive across months.
The Three Rules
- Never read 3+ files in Claude — pass them to NotebookLM
- Never fetch URLs in Claude — use NotebookLM source add
- Append session summaries to Master Brain — don't discard knowledge
The reframe: Claude Code isn't a research assistant. It's a decision-maker. Give it only what it needs to decide.