Deep Research Skill

Conduct comprehensive research on any topic by combining real-time web search (via Rube MCP) with multi-model deep analysis and consensus synthesis (via PAL MCP). Produces structured research reports with sourced findings, cross-validated analysis, and confidence assessments.

Quick Start

# Quick overview
/sc:research "quantum computing developments" --depth shallow

# Balanced research (default)
/sc:research "Impact of AI regulation on open-source development"

# Exhaustive deep dive saved to file
/sc:research "carbon capture technologies" --depth deep --output reports/carbon.md

# More models for consensus
/sc:research "PostgreSQL vs CockroachDB for write-heavy workloads" --models 4

Behavioral Flow

1. Parse - Extract topic, depth, output path, model count
2. Decompose - Break topic into 3-7 sub-questions for comprehensive coverage
3. Discover - Find web search tools via Rube MCP
4. Search - Execute web research in parallel batches
5. Analyze - Deep analysis of findings via PAL ThinkDeep
6. Validate - Multi-model consensus via PAL Consensus
7. Report - Generate structured markdown report
8. Output - Save to file or display in console

Flags

Flag	Type	Default	Description
--depth	string	medium	Research depth: shallow, medium, deep
--output	string	-	Save report to file path
--models	int	3	Number of models for consensus (2-5)

Depth Levels

Depth	Sub-questions	Searches/question	Follow-ups
shallow	2-3	1	0
medium	3-5	1-2	1 per gap
deep	5-7	2-3	2-3 per gap

Phase 1: Parse and Plan

Extract topic from arguments. Decompose into sub-questions that provide comprehensive coverage when answered together.

Present research plan before proceeding:

Research Topic: <topic>
Depth: <level>
Sub-questions:
  1. <sub-question>
  2. <sub-question>
  ...
Estimated searches: ~N

Phase 2: Discover Search Tools

Use mcp__rube__RUBE_SEARCH_TOOLS to find web search and URL extraction tools:

RUBE_SEARCH_TOOLS:
  session: { generate_id: true }
  queries:
    - use_case: "search the web for information about a topic"
    - use_case: "scrape and extract content from a web page URL"

From the response:

1. Record session_id — reuse for all subsequent Rube calls
2. Check connection status for returned toolkits
3. If no active connection, call RUBE_MANAGE_CONNECTIONS and present auth link
4. Identify best tools for web search and URL content extraction
5. If tools return schemaRef, call RUBE_GET_TOOL_SCHEMAS for full schemas

Phase 3: Execute Web Research

Batch independent searches in parallel (up to 5 per call) via RUBE_MULTI_EXECUTE_TOOL:

Search query formulation:

• Rephrase sub-questions as effective search queries
• Use specific, factual language
• For controversial topics, search multiple perspectives explicitly
• Include date qualifiers if recency matters

After initial searches:

1. Parse and collect all results
2. Identify most relevant URLs
3. For medium/deep: extract full content from top 3-5 URLs
4. Identify information gaps

For medium/deep — follow-up searches:

• Generate refined queries targeting gaps
• Execute follow-up searches
• Extract additional URL content

Organize raw findings:

Sub-question 1: <question>
  Sources:
    - [Source Title](URL) - Key finding: <summary>
  Gaps: <what's still unclear>

Sub-question 2: <question>
  Sources:
    - ...

Phase 4: Deep Analysis (PAL ThinkDeep)

Use mcp__pal__thinkdeep for systematic analysis:

Step 1 — Analyze:

• Identify key themes and patterns across sources
• Flag contradictions between sources
• Assess source credibility and biases
• Identify well-supported vs. poorly-supported claims
• Note significant information gaps
• Synthesize preliminary narrative

Step 2 — Refine:

• Resolve contradictions with evidence-based reasoning
• Rank findings by confidence (high/medium/low)
• Produce structured outline for final report
• Identify 3-5 most important takeaways

Phase 5: Multi-Model Consensus (PAL Consensus)

5a. Discover Models

Call mcp__pal__listmodels. Select top N by score, preferring different providers for diversity.

5b. Run Consensus

Use mcp__pal__consensus with for/against/neutral stances:

Stance	Purpose
for	Evaluate findings charitably, look for strengths
against	Critically evaluate, look for weaknesses and gaps
neutral	Balanced evaluation, weigh strengths and weaknesses

5c. Incorporate Consensus

Synthesize:

• Areas of agreement — high confidence findings
• Areas of disagreement — flag for nuance
• Missing perspectives — gaps identified by any model
• Confidence adjustments — raise/lower based on feedback

Phase 6: Generate Report

# Deep Research Report: <Topic>

**Generated**: <date>
**Depth**: <shallow|medium|deep>
**Sources consulted**: <N>
**Models consulted**: <list>

---

## Executive Summary

<3-5 paragraph synthesis for general audience>

---

## Research Question

<Original topic and how it was decomposed>

---

## Key Findings

### Finding 1: <Title>
**Confidence**: High | Medium | Low
**Consensus**: Agreed | Mixed | Disputed

<Detailed finding with inline source citations>

**Sources**: [Source 1](url), [Source 2](url)

---

## Analysis

### Themes and Patterns
<Cross-cutting themes across sources>

### Contradictions and Debates
<Where sources/models disagreed>

### Information Gaps
<What remains unclear>

---

## Model Consensus

| Model | Stance | Confidence | Key Feedback |
|-------|--------|------------|--------------|
| <model_1> | For | X/10 | <summary> |
| <model_2> | Against | X/10 | <summary> |
| <model_3> | Neutral | X/10 | <summary> |

**Agreement Areas**: <where all models agreed>
**Divergent Views**: <where models differed>

---

## Sources

| # | Title | URL | Relevance |
|---|-------|-----|-----------|
| 1 | <title> | <url> | <contribution> |

---

## Methodology

1. **Web search** via Rube MCP (<N> searches)
2. **Deep analysis** via PAL ThinkDeep
3. **Multi-model consensus** via PAL Consensus (<N> models)

Phase 7: Output

• If --output <filepath>: Write report to file, confirm path
• Otherwise: Display full report in console

End with summary:

Research complete:
- Topic: <topic>
- Sources: <N> web sources
- Models: <N> reached consensus
- Confidence: <overall assessment>
- Key takeaway: <1-sentence summary>

MCP Integration

PAL MCP

Tool	Phase	Purpose
mcp__pal__thinkdeep	Analysis	Multi-stage hypothesis testing
mcp__pal__consensus	Validation	Multi-model cross-validation
mcp__pal__listmodels	Discovery	Available models for consensus
mcp__pal__challenge	Validation	Critical thinking on controversial claims

Rube MCP

Tool	Phase	Purpose
mcp__rube__RUBE_SEARCH_TOOLS	Discovery	Find search/scraping tools
mcp__rube__RUBE_GET_TOOL_SCHEMAS	Discovery	Load full schemas if needed
mcp__rube__RUBE_MULTI_EXECUTE_TOOL	Search	Parallel web searches
mcp__rube__RUBE_MANAGE_CONNECTIONS	Auth	Connect search integrations
mcp__rube__RUBE_REMOTE_BASH_TOOL	Processing	Handle large response data

Error Handling

Scenario	Action
No search tools found	Fall back to WebFetch for direct URLs
Connection not active	Call RUBE_MANAGE_CONNECTIONS, present auth link
Search returns no results	Reformulate with broader terms
Tool schema missing	Call RUBE_GET_TOOL_SCHEMAS first
PAL MCP unavailable	Skip consensus, report from web research only
ThinkDeep fails	Continue with raw findings
Rate limiting	Reduce parallel batch size to 2-3

Guardrails

• Present research plan before executing searches
• Preserve all source URLs for verifiability
• Search multiple perspectives for controversial topics
• --depth deep can make 15-25+ API calls — use judiciously
• If Rube unavailable but PAL available, degrade to analysis-only via WebFetch

Tool Coordination

• WebFetch - Fallback URL fetching when Rube unavailable
• Write - Save report to file
• Bash - File operations
• PAL MCP - Analysis, consensus, challenge
• Rube MCP - Web search, URL extraction