deep-research

name	deep-research
description	Conduct deep research on a topic using multiple parallel Opus agents gathering from official sources, trusted publications, and expert blogs. Use this skill when the user says "deep research", "research deeply", or asks for comprehensive research on any topic. Produces a well-organized markdown report saved to research/.
allowed-tools	Task, Read, Write, Edit, Bash, Glob, Grep, WebSearch, WebFetch
model	opus
user-invocable	true
argument-hint	<research-topic>

Deep Research

Conduct deep, multi-source research on a topic using parallel research agents. Synthesize findings into a comprehensive, well-organized markdown document saved to research/.

Variables

RESEARCH_TOPIC: $ARGUMENTS

Instructions

Phase 1: Plan Research Angles

1. Parse the research topic

   • Extract the core subject from RESEARCH_TOPIC
   • Identify 4-6 distinct research angles/subtopics that together provide comprehensive coverage
   • Examples of angles: architecture/internals, getting started/usage, best practices, comparisons/alternatives, real-world case studies, advanced patterns, ecosystem/tooling, performance/benchmarks

2. Generate a sanitized filename

   • Convert RESEARCH_TOPIC to kebab-case
   • Remove special characters, limit to 60 chars
   • Example: "Kubernetes networking deep dive" → kubernetes-networking-deep-dive

3. Create output directory

   mkdir -p research
   

4. Check for existing file

   • If research/{filename}.md already exists, append a timestamp suffix: research/{filename}-{YYYYMMDD-HHMMSS}.md

Phase 2: Launch Parallel Research Agents

Launch 4-6 research agents in parallel using run_in_background: true. Each agent covers one research angle.

Each agent MUST be launched with subagent_type: "general-purpose" and model: "opus".

Agent prompt template (customize the ANGLE and FOCUS for each agent):

You are a deep research agent. Research the following topic thoroughly.

TOPIC: {RESEARCH_TOPIC}
ANGLE: {specific research angle}
FOCUS: {what specifically to investigate for this angle}

## Research Guidelines

1. **Use WebSearch** to find information from these source tiers (in priority order):
   - **Tier 1 (Official):** Official documentation, RFCs, specifications, official blogs
   - **Tier 2 (Trusted):** Reputable tech publications (InfoQ, ThoughtWorks, Martin Fowler, AWS/GCP/Azure blogs, Hacker News top posts, academic papers)
   - **Tier 3 (Expert):** Well-known technical expert blogs, conference talks, established open-source project docs

2. **Use WebFetch** to read the most relevant pages found via search. Read at least 3-5 high-quality sources.

3. **For each source, extract:**
   - Key facts, concepts, and technical details
   - Code examples or configuration snippets (if applicable)
   - The source URL for citation

4. **Quality filters — REJECT information that:**
   - Comes from content farms, SEO spam, or low-quality aggregators
   - Has no clear author or organizational backing
   - Contradicts official documentation without strong justification
   - Is outdated (prefer sources from the last 2 years unless the topic is historical)

5. **Return your findings as structured markdown with this format:**

## {Angle Title}

### Key Findings

- Finding 1 with technical detail
- Finding 2 with technical detail

### Details

{Detailed explanation with code examples if relevant}

### Sources

- [Source Title](URL) - Brief description of what was extracted
- [Source Title](URL) - Brief description of what was extracted

Be thorough. Provide specific technical details, not vague summaries. Include code examples, architecture diagrams (as text), configuration snippets, and concrete numbers where available.

Launch all agents in a SINGLE message with run_in_background: true.

Phase 3: Collect and Wait for Results

• Wait for ALL background research agents to complete
• Collect the output from each agent
• Note any agents that failed and what angle they covered

Phase 4: Synthesize and Write Report

1. Merge all research agent outputs into a single coherent document
2. Organize by topic, NOT by agent — restructure content logically
3. Apply the following document structure:

# {Research Topic Title}

> Research Date: {YYYY-MM-DD} | Sources: {count} | Research Agents: {count}

## Executive Summary

{3-5 paragraph overview of the most important findings across all research angles. This should stand alone as a useful summary for someone who doesn't read the full document.}

---

## Table of Contents

- [1. Section Name](#1-section-name)
- [2. Section Name](#2-section-name)
- ...

---

## 1. {First Major Section}

{Content synthesized from research agents, organized logically}

## 2. {Second Major Section}

{...}

## N. {Nth Major Section}

{...}

---

## Key Takeaways

- {Takeaway 1}
- {Takeaway 2}
- {Takeaway 3}

---

## Sources

| # | Source | Type | URL |
|---|--------|------|-----|
| 1 | {Title} | Official / Trusted / Expert | {URL} |
| 2 | {Title} | Official / Trusted / Expert | {URL} |

---

*Deep research conducted on {YYYY-MM-DD} using multi-agent parallel research*

4. Quality pass:

   • Remove duplicate information across sections
   • Ensure consistent terminology throughout
   • Verify no contradictions between sections (if found, note the discrepancy and which source is more authoritative)
   • Ensure all claims have a source citation
   • Fix formatting inconsistencies

5. Write the final document to: research/{sanitized-topic-name}.md

Phase 5: Report

Display a summary to the user:

## Deep Research Complete

**Topic:** {RESEARCH_TOPIC}
**File:** research/{filename}.md
**Agents:** {N} research agents completed
**Sources:** {N} unique sources cited
**Sections:** {N} major sections

### Executive Summary Preview

{First 2-3 sentences of the executive summary}

### Sections

1. {Section name} - {one-line description}
2. {Section name} - {one-line description}
...

Error Handling

• Agent failure: If a research agent fails, note the gap in coverage and mention it in the report. Do NOT block the entire research on one failed agent.
• No results for an angle: Include the section with a note that limited information was found.
• Topic too broad: If the topic is extremely broad, focus on the most actionable and technically deep angles rather than trying to cover everything superficially.

Success Criteria

• At least 4 research agents completed successfully
• Final document has 5+ major sections with substantive content
• All claims cite at least one source
• Sources table includes 10+ unique, high-quality sources
• Document is saved to research/ directory
• Executive summary provides a useful standalone overview
• No duplicate content across sections
• Consistent formatting and terminology throughout