Multi-agent parallel research orchestrator. Decomposes any research topic into 3-6 specialized angles using MECE principles, spawns Opus sub-agents to cover each angle simultaneously, runs gap analysis, then synthesizes findings into one comprehensive research document.
Anthropic’s own multi-agent research system outperforms single-agent by 90.2%. This skill applies those same proven patterns.
For detailed sub-agent prompt templates by research type: read agent-templates.md.
For the full research on multi-agent orchestration patterns: read orchestration-patterns.md.
A generic deep-research workflow. Inject the active requester/project context from the user brief, durable preferences/memory, repository guidance, and any supplied notes. Do not hard-code a creator profile, social handle, vault path, or niche unless the current task provides it.
Every research session follows this shape:
[Broad: Decompose question]
/ \
[Narrow: 3-5 parallel agents] ← Wave 1
\ /
[Evaluate: Gap analysis]
/ \
[Deep: 1-2 targeted agents] ← Wave 2 (if needed)
\ /
[Synthesize: Final report]
Start wide, go narrow in parallel, identify gaps, go deep on gaps, synthesize.
Before anything else, note today’s date. You can get it from the system prompt or by running date. Inject this date into every sub-agent prompt so they search for current information and properly date the final document. This is critical — sub-agents without date context will search for and cite outdated information.
Before spawning any agents:
Break the topic into Mutually Exclusive, Collectively Exhaustive angles. Each angle:
Common decomposition patterns:
| Research Type | Typical Angles |
|---|---|
| Content/Platform | General best practices, domain-specific strategy, real examples with metrics, psychology/copywriting, existing knowledge, social discourse |
| Technology | Current state/ecosystem, real shipped code, community sentiment, comparisons/alternatives, existing knowledge, implementation patterns |
| Business | Market data/benchmarks, domain-specific practices, strategy frameworks, existing knowledge, practitioner discourse |
Scale effort to complexity:
Spawn minimum 3, ideally 4-5 sub-agents in parallel using the available delegation tool. Use the strongest practical model for substantive research/synthesis; cheap models are acceptable only for smoke tests or decomposition.
Every sub-agent prompt MUST include these 6 elements:
WHO — “This research is for [requester/project]. Relevant context: [stable preferences, domain, project goals, constraints].”
WHY — The specific purpose, e.g. decision support, issue planning, spec drafting, artifact creation, or reusable knowledge. Agents that know WHY produce dramatically better results.
WHAT ANGLE — Specific scope AND explicit boundaries: “Cover X. Do NOT cover Y — another agent handles that.”
HOW — Which tools to use (see Tools Reference below)
SEARCH STRATEGY — “Start with SHORT, BROAD queries (2-4 words). Evaluate results. Then progressively narrow focus. Do NOT start with long, specific queries — they return poor results.”
SOURCE QUALITY — “Prefer: practitioner blogs, official docs, academic papers, primary sources. Avoid: SEO content farms, listicles, aggregator sites.”
You are researching [ANGLE] for [REQUESTER/PROJECT].
Relevant context: [stable preferences, domain, project goals, constraints, and intended audience].
TODAY'S DATE: [INSERT CURRENT DATE, e.g. 2026-02-10]
PURPOSE: [WHY this research matters — how the requester/project will use it]
YOUR ANGLE: [SPECIFIC SCOPE — what you cover]
BOUNDARIES: [What you do NOT cover — other agents handle those angles]
SEARCH STRATEGY:
- Start with short, broad queries (2-4 words)
- Evaluate what's available, then progressively narrow
- Cross-reference claims across multiple sources
- Aim for 2+ independent sources per key finding
SOURCE QUALITY: Prefer practitioner blogs, official docs, engineering posts, primary
sources. Avoid SEO content farms and listicles.
TOOLS TO USE: [SPECIFIC tools for this angle]
QUALITY BAR: Be exhaustive. Extract every actionable insight, specific number, concrete
example, framework, and contrarian take. Density matters — thin research is useless.
Include sources/URLs for everything. If you find only 3 bullet points, you failed.
OUTPUT FORMAT:
## Key Findings
[Numbered list with inline source citations]
## Evidence Quality
[Which findings are well-sourced vs. speculative]
## Contradictions Found
[Any conflicting information between sources]
## Notable Quotes
[Direct quotes from authoritative sources with attribution]
## Sources
[Full list with URLs]
| Angle Type | subagent_type | model | Tools |
|---|---|---|---|
| Web research | delegate_task |
strong | web_search, web_extract, browser when needed |
| Existing knowledge | delegate_task |
strong | session_search, search_files, read_file on relevant paths |
| Code examples | delegate_task |
strong | GitHub/code search tools available in the environment |
| Social discourse | delegate_task |
strong | xurl/xitter or configured social/search tools |
Social/X research: use the configured social-search skill/CLI if available; otherwise use web search and clearly label coverage limits.
After ALL sub-agents return (batch — do not process one-at-a-time to avoid anchoring bias):
Cross-reference all findings and produce ONE comprehensive document. Do NOT simply concatenate agent outputs — synthesize them into something greater than the sum of parts.
File location: use the user-specified destination. If absent, save a dated Markdown artifact in the current project/workspace notes or research folder; do not assume a personal vault path.
Topic folder organization: Group all research outputs into a topic subfolder within the chosen research/notes location. If a research session produces multiple files (master synthesis + companion documents from sub-agents), they ALL go in the same topic folder. Create the folder if it does not exist.
| Research Topic | Folder |
|---|---|
| Existing project topic | reuse the matching research/notes folder |
| New durable topic | Research/[Topic Name]/ or project equivalent |
| One-off research | Research/Other/ or a clearly named dated note |
Rules:
Research/[Folder]/ headersDocument structure:
# [Topic Name] [Year]
Research compiled for [requester/project context]. [N] parallel research tracks synthesized.
**Purpose**: [What this research will be used for]
**Date**: [Current date]
**Sources**: [N] web sources, [N] social posts, [N] existing-knowledge references, [N] code examples
---
## Executive Summary
[3-5 bullet points — the most important findings]
---
## Part N: [Angle Name]
### [Sub-topic]
[Dense, actionable content with specific numbers and examples]
---
## Domain-Specific Applications
[How findings apply to the requester/project domain specifically]
---
## Contradictions & Open Questions
[Where sources disagreed, what remains unresolved]
---
## Key Takeaways
[Numbered list of the most actionable insights]
---
## Sources
[All sources cited, organized by section]
Quality gate — do NOT save until ALL pass:
After saving, update the relevant index/README if a new file was created in a location not yet indexed.
| Tool | Use For | Notes |
|---|---|---|
| web_search / web_extract | Current web information and deep-dive URLs | Prefer primary/current sources |
| browser | Dynamic pages or pages web_extract cannot read | Use sparingly |
| GitHub/code search | Real code patterns | Search actual APIs/config, not vague keywords |
| xurl/xitter/social tools | X/Twitter or social discourse | Use only when configured |
| session_search/search_files/read_file | Existing knowledge | Always check first |