Deep Research

Questions to Ask

Ask 3-4 questions at a time using AskUserQuestion (max per call). Run multiple rounds if needed. Tailor questions to the specific query — these are starting points, not a rigid checklist:

Round 1 — Scope & Intent:

• What is the primary goal of this research? (e.g., decision-making, learning, writing an article, building something)
• What do you already know about this topic? (helps avoid redundant research)
• Are there specific aspects you care most about? (narrow focus vs. broad survey)
• What time period matters? (last year, last 5 years, all time)

Round 2 — Depth & Boundaries:

• Are there specific subtopics you want included or excluded?
• Do you need academic/peer-reviewed sources, or are industry blogs and news sufficient?
• Is there a geographic or domain-specific focus? (e.g., US regulations, healthcare applications)
• Are you looking for consensus views, or also want contrarian/minority perspectives?

Round 3 — Context the User Didn't Think About (YOUR MOST IMPORTANT JOB): Based on the query and their answers, proactively suggest angles they may not have considered:

• "Have you thought about [related aspect]? It could affect your conclusions."
• "The query touches on [adjacent field] — should I include that?"
• "There's a common misconception about [topic] — want me to specifically investigate that?"
• "Recent developments in [area] might be relevant — should I cover that?"

Interview Rules

• Ask at LEAST 2 rounds of questions (6-8 questions minimum)
• NEVER skip this step, even if the query seems clear
• Use the user's answers to refine the query before decomposition
• If the user says "just go" or wants to skip, ask ONE final question: "Is there anything specific you do NOT want included?" then proceed
• Record all user answers — they become constraints for the researchers
• The user's expertise level affects how you write the final report (technical vs. accessible)

Output of This Step

After the interview, you should have:

• Refined query: The original query plus all context from the interview
• Scope boundaries: What's in and out of scope
• Source preferences: Credibility requirements, geographic focus, time period
• User expertise level: Affects report language (technical vs. accessible)
• Specific inclusions/exclusions: Topics the user explicitly wants or doesn't want

Step 3: Decompose Query (Do This Yourself — DO NOT Spawn an Agent)

Using the refined query and constraints from Step 2 (interview), decompose into N subtopics.

Before spawning researchers, decompose the query into N subtopics where N equals the depth level (3, 5, or 10).

For each subtopic, define:

• title: Clear, specific topic name
• keywords: 3-5 search terms for WebSearch
• angle: The specific aspect this researcher should focus on
• covered_topics: List of all other subtopics being covered

Decomposition Requirements

• Subtopics MUST be non-overlapping (no duplicate coverage)
• Together they MUST cover the full scope of the query
• Each subtopic should be specific enough for 2-3 web searches
• Include at least one subtopic for "recent developments / current state"
• Include at least one subtopic for "criticisms / limitations / challenges"

Note: Follow the decomposition structure defined in prompts/decompose.md for the output format and coverage validation criteria.

Example Decomposition

For query "AI safety in autonomous vehicles":

1. title: "Current AI Safety Regulations", keywords: ["autonomous vehicle regulations", "self-driving safety standards", "government AI vehicle policy"], angle: Regulatory landscape and compliance requirements
2. title: "Technical Safety Mechanisms", keywords: ["autonomous vehicle failsafes", "AI decision making safety", "sensor redundancy"], angle: Engineering approaches to safety
3. title: "Accident Data and Risk Analysis", keywords: ["self-driving car accidents", "autonomous vehicle crash statistics", "AI safety incidents"], angle: Empirical safety record
4. title: "Ethical Decision-Making Frameworks", keywords: ["trolley problem autonomous vehicles", "AI ethics self-driving", "moral algorithms"], angle: Ethical challenges in AI decision-making
5. title: "Recent Developments and Challenges", keywords: ["autonomous vehicle 2026", "self-driving safety breakthroughs", "latest AI vehicle technology"], angle: Current state and emerging issues

Step 4: Present Research Plan for Approval

Before spawning any agents, present the complete research plan to the user and wait for approval.

Plan Format

Display this to the user:

## Research Plan

**Query**: {refined query from interview}
**Depth**: {depth level} ({N} parallel researchers)
**Estimated token usage**: {quick: ~100K, medium: ~150K, deep: ~260K}

### Subtopics to Research

| # | Subtopic | Focus Angle | Key Search Terms |
|---|----------|-------------|------------------|
| 1 | {title}  | {angle}     | {keywords}       |
| 2 | ...      | ...         | ...              |

### Scope Constraints (from interview)
- Time period: {time period}
- Source types: {credibility requirements}
- Inclusions: {specific topics to include}
- Exclusions: {specific topics to exclude}

### Coverage Check
- ✅ Recent developments covered (subtopic #{N})
- ✅ Criticisms/limitations covered (subtopic #{N})
- ✅ Practical applications covered (subtopic #{N})
- {any additional coverage notes}

Getting Approval

Ask the user: "Does this research plan look good? You can:

1. Approve — I'll start the research
2. Modify — Tell me what to change (add/remove/adjust subtopics)
3. Change depth — Switch between quick (3), medium (5), or deep (10)"

Use AskUserQuestion with these options.

Do NOT proceed to spawning researchers until the user explicitly approves.

If the user requests modifications:

1. Adjust the plan accordingly
2. Re-display the updated plan
3. Ask for approval again

Step 5: Spawn All Researchers in Parallel (CRITICAL — One Message)

Spawn all N researcher agents in a SINGLE message. Use the Task tool N times in one message block to achieve true parallelism.

For each subtopic, make this Task call:

subagent_type: "general-purpose"