Deep Research — Multi-Agent Adversarial Ensemble

A Claude Code skill for conducting deep, rigorous research using a structured multi-agent ensemble with built-in adversarial review.

What it does

Runs a 3-cycle iterative research process with specialized agent roles working in parallel:

SCOUTS (4-5 parallel) → broad search across streams
CRITIC → finds contradictions and weak evidence across ALL streams
STATISTICIAN → verifies methodology of cited studies
DEEP DIVERS (2-3 parallel) → closes gaps identified by Critic
SYNTHESIZER → integrates all findings into a coherent document
INTERACTION MAPPER → maps cross-interactions (nutrients × genetics × biomarkers)
DOMAIN_REVIEWER → domain-specific safety and validity check
FACT-CHECKER → verifies top-15 numerical claims (MANDATORY always)

Why an ensemble outperforms a single agent:

• Each SCOUT is isolated → no confirmation bias between streams
• CRITIC sees ALL streams simultaneously → catches cross-stream contradictions
• STATISTICIAN catches methodological garbage (p-hacking, underpowered studies)
• FACT-CHECKER catches hallucinated numbers at the last gate

Research Modes

Mode	Deliverables	When to use
personalized (default)	synthesis.md tailored to user profile	User's specific question
consensus	consensus_reference.md — universal, reusable	Building a knowledge base
consensus+interactions	consensus + interaction_map.md	When cross-effects matter
full	All of the above	Deep research with personalization

Four Levels of Output

Level 1: consensus_reference.md — "what science says" (population truth)
Level 2: interaction_map.md    — "when this changes" (conditional truth)
Level 3: synthesis.md          — "what to do for YOU" (personalized truth)
Level 4: unknowns_and_next.md  — "what we still don't know" (honest uncertainty)

Installation

This is a Claude Code custom slash command.

1. Copy research.md to ~/.claude/commands/research.md
2. In Claude Code, run /research your topic

Or for project-level install, copy to .claude/commands/research.md in your project root.

Usage

/research caffeine impact on sleep
/research creatine safety high 4h
/research omega-3 full high 8h
/research exercise outcomes consensus

Format: /research [topic] [priority?] [hours?] [mode?]

• priority: high / medium / low (auto-detected if omitted)
• hours: e.g., 4h, 8h (auto-detected if omitted)
• mode: personalized / consensus / consensus+interactions / full

Configuration

Set in your project's CLAUDE.md or pass as context:

OUTPUT_DIR: ./research_output/   # Where to save all research files
USER_CONTEXT: ""                 # Optional path to user profile for personalization
BILINGUAL: false                 # Set to true for bilingual output
SECONDARY_LANGUAGE: ""           # e.g., "ru", "es", "de", "zh"

Headless Launch

For long-running research sessions in the background:

claude --dangerously-skip-permissions -p "/research topic priority hours mode"
# Log: 2>&1 | tee /tmp/research_topic.log

Domains Supported

health · nutrition · ai · finance · psychology · business · general

Each domain has pre-configured research streams. For health/nutrition: STATISTICIAN and DOMAIN_REVIEWER launch automatically. For other domains: activate if needed.

Agent Reasoning Styles

Each SCOUT gets a unique reasoning style to reduce error correlation between agents (Hong & Page 2004):

SCOUT	Style
A	Analytical — systematic, taxonomies, meta-analyses first
B	Contrarian — seeks refutations, null results, minority views
C	Mechanistic — HOW and WHY, causal chains, pathways
D	Systems-thinking — feedback loops, second-order effects
E	Pragmatic — dose-response, NNT/NNH, implementation barriers

License

MIT