- name
- deep-research
- description
- Automated deep research that performs comprehensive multi-source investigation and produces detailed reports with citations. Use when user requests research, investigation, or in-depth analysis of any topic. Capabilities: generate structured research plans, and start sub agent to execute the plan.
Deep Research
Two-phase research workflow: planning then execution.
Overview
This skill provides a structured approach to deep research:
Phase 1: Planning (High Freedom)
- Discuss with user to clarify and refine research questions.
- Define what to investigate and what the report should cover.
- Set expectations for research depth and output.
- Create research plan document.
Phase 2: Execution (Low Freedom)
- Sub-agent reads the research plan
- Independently decides how to search for each sub-question
- Can dynamically add searches based on findings
- Analyzes content and generates report with citations
Phase 1: Generate Research Plan
The coordinator (main session) performs:
- Understand the research topic — Listen to user's request and understand what they want to investigate
- Collaborate with user — Discuss and clarify research questions together. Present 3-5 potential sub-questions or research angles for user to review
- Define scope together — Discuss what to include/exclude, confirm boundaries of the research
- Confirm report expectations — Ask user what sections they want, what depth, any specific focus areas
- Get user confirmation — Present the draft plan to user and wait for approval before proceeding
- Output: Research plan document — Only after user confirms, save to
plans/research-plan-{timestamp}.json
Key principle: The plan is a collaboration between coordinator and user. Never proceed to Phase 2 without explicit user confirmation of the research plan.
Research plan format (JSON):
{
"topic": "Original research topic",
"research_questions": [
"What are the latest breakthroughs in this field?",
"Who are the leading organizations or researchers?",
"What are the current limitations or challenges?",
"What are the practical applications?"
],
"scope": {
"include": ["recent developments", "key players", "technical details"],
"exclude": ["historical background before 2020", "unrelated applications"]
},
"report_requirements": {
"sections": ["executive_summary", "findings", "conclusion", "references"],
"depth": "comprehensive",
"min_sources": 8,
"focus_areas": ["technical analysis", "market landscape"]
}
}Research Plan Schema
Required fields:
- topic: Original research topic
- research_questions: Array of questions to investigate
- report_requirements: Object specifying output expectations
Optional fields:
- scope: Define boundaries of research (include/exclude)
- min_sources: Minimum sources to analyze (default: 8)
- max_sources: Maximum sources to analyze (default: 20)
- notes: Additional context or special instructions
Save plan to: plans/research-plan-{timestamp}.json
⚠️ WAIT FOR USER CONFIRMATION — Do not proceed to Phase 2 until user explicitly approves the research plan.
Key principle: The plan defines WHAT to research and WHAT the output should contain. It does NOT specify HOW to search (keywords, sources, rounds) - that is up to the research agent to determine dynamically.
Phase 2: Execute Research Plan
Launch sub agent with the research plan. Launch sub agent with session_spawn tool. Instruct subagent to use deep-research-executor to execute the plan EXPLICITLY.