Deep Research

</essential_principles>

Available Domains

General (Default)

• Config: references/domains/general.md
• Agents: deep-research:core:research-worker, deep-research:core:gap-detector
• Tools: WebSearch, WebFetch

Geopolitical

• Triggers: conflict, war, sanctions, bilateral relations, media coverage, GDELT, CAMEO
• Config: references/domains/geopolitical.md
• Agents: Specialized analysts (conflict, sanctions, actors, sentiment, trends)
• Tools: GDELT MCP tools + WebSearch

Domain Detection

IF query contains (conflict, war, military, sanctions, embargo, bilateral,
   alliance, media coverage, sentiment, GDELT, CAMEO):
   → GEOPOLITICAL domain
ELSE:
   → GENERAL domain

Override: /deep-research --domain=geopolitical "query"

<available_workflows>

Workflow Variants

Load the appropriate workflow for the research mode:

</available_workflows>

The Research Loop

PHASE 0: DOMAIN CLASSIFICATION & REFINEMENT
├── Parse query and detect domain
├── Load domain config from references/domains/
├── Generate draft research brief
├── Present brief to user via AskUserQuestion (domain, dimensions, depth)
├── Refine based on feedback
└── Proceed when confirmed

PHASE 1: INITIALIZATION
├── Finalize research brief
├── Select domain-appropriate agents
└── Generate initial speculative draft

PHASE 2: DIFFUSION LOOP (max 3 iterations)
├── SPAWN WORKERS: 2-5 domain-specific workers in PARALLEL (single message)
├── SYNTHESIZE: Merge findings, resolve contradictions, add citations
├── GAP DETECTION: Launch domain-specific gap-detector
└── Decision: CONTINUE (spawn more workers for gaps) or COMPLETE

PHASE 3: FINAL REPORT
├── Polish using domain output template from config
├── Compile all sources with descriptions
└── Add executive summary

Interactive (default): Orchestrate directly, can ask questions mid-research. Background: Spawn deep-research:core:research-orchestrator for autonomous operation.

<execution_guide>

Phase 0: Domain Classification

Analyze the query for domain signals. Present a draft research brief with detected domain, proposed dimensions, key questions, and specialist agents (if geopolitical). Validate with user via AskUserQuestion for domain, dimensions, and depth level.

Phase 1: Spawn Workers

Load worker prompt templates from the appropriate domain config file:

• General: See references/domains/general.md for worker prompt template
• Geopolitical: See references/domains/geopolitical.md for specialist agent prompts

Critical: Spawn ALL workers in a SINGLE message with multiple Task tool calls for true parallel execution.

Phase 2: Synthesis and Gap Detection

After workers return, merge findings into a structured draft with citations. Then launch the domain-specific gap-detector:

• General: deep-research:core:gap-detector
• Geopolitical: deep-research:geo:gap-detector-geo (uses weighted coverage dimensions from domain config)

If gaps found, spawn targeted workers for those specific gaps. Maximum 3 iterations.

Phase 3: Final Report

Use domain-specific output template from the domain config file. Save to research_output/[topic-slug]_[date].md (or research_output/geo/ for geopolitical).

Required sections: Metadata header, Executive summary, Key findings with sources, Detailed analysis, Methodology note, Complete source list. Additional domain-specific sections are defined in each domain config.

</execution_guide>

<mid_research_steering>

Mid-Research Steering

</mid_research_steering>

<success_criteria>

Research is complete when:

• All core questions from the brief are addressed
• Coverage score >= 70% (from gap detector)
• Key findings are well-sourced (2+ citations each)
• Contradictions are noted and explained
• Maximum 3 iterations reached

Research is NOT complete just because the draft reads well or workers stopped returning new information. Always use the domain-appropriate gap-detector for objective evaluation.

</success_criteria>