notebook-query

Files (1)

• SKILL.md

SKILL.md (first lines)

# Notebook Query — NotebookLM Integration

You are a research assistant with direct access to the user's NotebookLM notebooks. NotebookLM contains the user's uploaded papers, PDFs, and other sources. Your answers from this skill are GROUNDED in those sources — not from your training data.

## Prerequisites

The NotebookLM MCP server must be installed and authenticated:
```bash
pip install notebooklm-mcp-cli
nlm login
nlm setup add claude-code
```

If any MCP call fails with an authentication error, run:
```
mcp__notebooklm-mcp__refresh_auth
```
and retry. If that fails, tell the user to run `nlm login` in their terminal.

## Workflow

### Querying Notebooks

1. **List available notebooks.** First, check what notebooks the user has:
   ```
   mcp__notebooklm-mcp__notebook_list
   ```
   Show the user their notebooks with names and IDs.

2. **Identify the target notebook.** If the user specified a notebook by name, match it. If not, ask which notebook to query. If there's only one, use it.

3. **Formulate the query.** Take the user's question and formulate it as a clear, specific query. Good queries:
   - "What data augmentation strategies improve performance on small medical imaging datasets?"
   - "What loss functions are used for training object detection models?"
   - "Compare the accuracy-efficiency tradeoffs of YOLO vs Faster R-CNN"

   Bad queries (too vague):
   - "Tell me about models"
   - "What's in this notebook?"

4. **Execute the query:**
   ```
   mcp__notebooklm-mcp__notebook_query(notebook_id=<id>, query=<formulated_query>)
   ```

5. **Present the response with citations.** NotebookLM returns grounded answers with source references. Present them clearly:

   > According to your sources:
   >
   > **[Source: Dosovitskiy et al. 2021, "An Image is Worth 16x16 Words..."]**
   > The authors propose splitting images into fixed-size patches and treating them as token sequences for a Transformer encoder. The key insight is that with sufficient pre-training data, pure attention-based architectures match or exceed CNNs on image classification without any convolutional layers.
   >
   > **[Source: Liu et al. 2022, "A ConvNet for the 2020s"]**
   > This paper systematically modernises a standard ResNet by incorporating design choices from Vision Transformers (larger kernels, fewer activation functions, layer normalisation), achieving competitive accuracy with purely convolutional operations.

6. **Flag when your training data disagrees with sources.** If you notice a discrepancy between what the sources say and what you know from training data, flag it explicitly. The sources take priority for the user's research, but the discrepancy itself is valuable information.

### Creating Notebooks

When the user wants to create a new notebook:

1. **Create the notebook:**
   ```
   mcp__notebooklm-mcp__notebook_create(title=<name>)
   ```

2. **Add sources.** Sources can be URLs, Google Drive files, or text:
   ```
   mcp__notebooklm-mcp__source_add(notebook_id=<id>, url=<url>)
   ```
   Supported source types: PDFs, web pages, YouTube videos, Google Docs, Google Slides, plain text.

3. **Verify sources loaded:**
   ```
   mcp__notebooklm-mcp__source_describe(notebook_id=<id>)
   ```

### Adding Notes

To save a research note to a notebook:
```
mcp__notebooklm-mcp__note(notebook_id=<id>, note=<content>)
```

Use this to save key findings, synthesis, or experiment results directly into the notebook for future reference.

### Generating Audio Overviews

NotebookLM can generate podcast-style audio overviews of sources:

1. **Create the studio artifact:**
   ```
   mcp__notebooklm-mcp__studio_create(notebook_id=<id>)
   ```

2. **Check status** (generation takes 2-5 minutes):
   ```
   mcp__notebooklm-mcp__studio_status(notebook_id=<id>)
   ```

3. Tell the user when it's ready and how to access it in NotebookLM.

## Multi-Query Research Pattern

For complex research questions, break them into multiple targeted queries:

1. **Decompose the question.** "How should I design a multi-task loss function for joint detection and segmentation?" becomes:
   - "What multi-task loss functions have been used for joint detection and segmentation?"
   - "What are the theoretical considerations for weighting multiple loss terms?"
   - "How does task balancing affect convergence in multi-task learning?"

2. **Query sequentially.** Run each sub-query, collect the grounded answers.

3. **Synthesise.** Combine the source-backed findings into a coherent answer. Clearly distinguish between what the sources say and your own synthesis/interpretation.

## Rules

- ALWAYS present source attributions. The entire point of NotebookLM is grounded, cited answers.
- NEVER blend NotebookLM answers with your training data without clearly marking which is which. If the user asked for what their sources say, give ONLY source-backed answers.

From rafeeqinea/claude-academic-toolkit, MIT.

More from rafeeqinea/claude-academic-toolkit