Deep Research Skill

Overview

This skill conducts thorough, multi-phase research using parallel subagents and iterative search methodology. It simulates ChatGPT Deep Research and Anthropic Deep Search by breaking complex topics into sub-questions, distributing work across 6-10 parallel research agents, and synthesizing findings into a structured report.

When to Use

Use this skill when the user requests:

• Deep research on a topic
• Comprehensive analysis
• Competitive intelligence
• Market research
• Thorough investigation (not quick facts)
• Multi-angle exploration of complex subjects

Research Methodology

Core Principles

1. Multi-pass queries — Never one-and-done; iterate based on findings
2. Source triangulation — Verify claims across 3-5 independent sources
3. Primary source hunting — Find original studies, docs, not just blog posts
4. Contradiction spotting — Flag where sources disagree; don't hide uncertainty
5. Synthesis over summary — Connect dots, identify patterns, surface insights

Parallel Agent Architecture

For deep research, spawn 6-10 subagents to explore different angles simultaneously:

Research Lead (you)
├── Agent 1: Background & definitions
├── Agent 2: Market/industry landscape
├── Agent 3: Key players/competitors
├── Agent 4: Technology/trends
├── Agent 5: Challenges/risks
├── Agent 6: Opportunities/future outlook
├── Agent 7: Case studies/examples
├── Agent 8: Data/statistics
└── Agent 9-10: Specialized deep-dives (as needed)

Search Tool Strategy

Use web_search with different modes per phase:

Use web_fetch to:

• Extract full article content from promising URLs
• Read primary sources, studies, documentation
• Get details that search snippets miss

Workflow

Phase 1: Scoping (5 min)

1. Clarify the topic — Ask user if the request is ambiguous
2. Identify sub-questions — Break the topic into 6-10 research angles
3. Define success — What does a good answer look like?

Example sub-question breakdown for "AI agent platforms":

• What are AI agent platforms and how do they work?
• What's the market size and growth trajectory?
• Who are the major players (established + startups)?
• What technologies power these platforms?
• What are the main use cases?
• What challenges/limitations exist?
• What's the competitive landscape?
• What trends are emerging?

Phase 2: Parallel Research (15-25 min)

Spawn subagents with sessions_spawn for each research angle:

# Example subagent spawn
sessions_spawn(
  task="Research [specific angle]. Use web_search with mode=deep-reasoning, 20-30 results. Fetch full content from 5-10 key sources. Return: key findings, statistics, quotes with sources, contradictions spotted.",
  runtime="subagent",
  mode="run"
)

Each subagent should:

• Use appropriate web_search mode for their angle
• Fetch 5-10 full articles with web_fetch
• Return structured findings with source citations
• Flag uncertainties or conflicting information

Phase 3: Synthesis (10-15 min)

As research lead, consolidate findings:

1. Aggregate results — Collect all subagent outputs
2. Identify patterns — What themes emerge across angles?
3. Spot contradictions — Where do sources disagree?
4. Fill gaps — Run targeted searches for missing pieces
5. Verify claims — Cross-check key statistics across sources

Phase 4: Report Writing (10 min)

Structure the final report as follows:

Output Format

# [Research Topic]

## Executive Brief

[150-250 words: The 3-5 most important takeaways. Lead with the answer. What should the reader know after finishing this report?]

---

## 1. Background & Context

[Foundational information, definitions, why this matters]

## 2. [Key Theme 1]

[Deep dive with supporting evidence]

## 3. [Key Theme 2]

[Deep dive with supporting evidence]

## 4. [Key Theme 3]

[Deep dive with supporting evidence]

## 5. Challenges & Risks

[What could go wrong, limitations, open questions]

## 6. Opportunities & Outlook

[Future trends, emerging developments, what to watch]

## Key Takeaways

- [Bulleted summary of 5-7 most important points]

---

## Sources

[Numbered list with full URLs, titles, and 1-line context for each source]

1. [Title](URL) — [Brief context: what this source contributed]
2. [Title](URL) — [Brief context]
...

Citation Guidelines

• In-text — Use numbered brackets: [1], [2-4], [5, 7]
• Sources section — Full URL, title, and 1-line context
• Minimum sources — 20-30 for deep research
• Quality over quantity — Prefer primary sources, industry reports, reputable publications

Tool Usage

web_search

# Broad exploration
web_search query="[topic]" type="deep-reasoning" count=30 freshness="year"

# Targeted lookup
web_search query="[specific fact]" type="fast" count=10

# Recent developments
web_search query="[topic]" type="neural" count=20 freshness="month"

web_fetch

# Extract full content
web_fetch url="https://example.com/article" extractMode="markdown" maxChars=5000

sessions_spawn (for parallel research)

# Spawn research subagent
sessions_spawn(
  task="Research [specific angle]. Search with mode=deep-reasoning, 25 results. Fetch 8-10 full articles. Return structured findings with citations.",
  runtime="subagent",
  mode="run"
)

Quality Checks

Before delivering the report, verify:

• Executive brief captures the 3-5 most important takeaways
• All major claims have 2+ source citations
• Contradictions/uncertainties are flagged, not hidden
• Sources section has 20-30 entries with full URLs
• Report answers the original question thoroughly
• No obvious gaps in coverage

Adaptation

For Quick Research (<10 min)

• Skip subagent spawning
• Run 3-5 targeted searches yourself
• Aim for 10-15 sources
• Condense report structure

For Ultra-Deep Research (60+ min)

• Spawn 10-15 subagents
• Include primary source documents, academic papers
• Add data tables, comparisons, timelines
• Include appendix with raw findings

Notes

• Context efficiency — Subagents run in isolated sessions; only their findings load into your context
• Parallelism — Spawn all subagents at once, then sessions_yield to wait for completion
• Iterative — If initial findings reveal new angles, spawn follow-up agents
• Time boxing — Set runTimeoutSeconds on subagents to prevent runaway research