# Deep Research Skill for Claude Code

> Inspired by [RhinoInsight: Improving Deep Research through Control Mechanisms for Model Behavior and Context](https://arxiv.org/abs/2511.18743)

A structured research workflow skill for Claude Code, supporting two-phase research: outline generation (extensible) and deep investigation. Human-in-the-loop design ensures precise control at every stage.

## Use Cases

- **Academic Research**: Paper surveys, benchmark reviews, literature analysis
- **Technical Research**: Technology comparison, framework evaluation, tool selection
- **Market Research**: Competitor analysis, industry trends, product comparison
- **Due Diligence**: Company research, investment analysis, risk assessment

## Commands

| Command | Description |
|---------|-------------|
| `/research` | Generate research outline with items and fields |
| `/research/add-items` | Add more items to existing outline |
| `/research/add-fields` | Add more fields to existing outline |
| `/research/deep` | Deep research each item with parallel agents |
| `/research/report` | Generate markdown report from JSON results |

## Installation

Copy the `skills/` folder to your Claude Code directory:

```bash
cp -r skills/* ~/.claude/skills/
```

## Workflow

### Phase 1: Generate Outline
```
/research <topic>
```
- Uses model knowledge + web search
- Asks for existing field definitions
- Creates `{topic}/` directory with separated files

### Phase 2: Deep Research
```
/research/deep
```
- Reads outline and fields automatically from current directory
- Launches parallel agents (5 per batch)
- Agents read fields.yaml independently (not passed in prompt)
- Outputs structured JSON per item
- Supports resume from checkpoint

### Optional: Expand Outline
```
/research/add-items    # Add more research targets
/research/add-fields   # Add more field definitions
```

## Output Format

### Directory Structure
```
{topic}/
  ├── outline.yaml    # items + execution config
  ├── fields.yaml     # field definitions
  └── results/        # deep research outputs
```

### outline.yaml
```yaml
topic: "your topic"
items:
  - name: "Item1"
    source: "source info"
execution:
  batch_size: 5          # parallel agents (default: 5)
  items_per_agent: 1     # items per agent (default: 1)
  output_dir: "./results" # output directory (default: ./results)
```

### fields.yaml
```yaml
basic_info:
  - name: "field_name"
    description: "field description"
    detail_level: "detailed|brief"
```

### Research Result (JSON)
Each item outputs a structured JSON file with all defined fields.

### Phase 3: Generate Report
```
/research/report
```
- Generates Python script to convert JSON to markdown
- Creates report with table of contents and anchor links
- Skips uncertain fields automatically

## References

- RhinoInsight: Improving Deep Research through Control Mechanisms for Model Behavior and Context

## License

MIT