deep-research

name: deep-research description: "Multi-agent iterative research pipeline using Eric Jang methodology. 10-18 specialized agents across 3 cycles produce consensus references, interaction maps, personalized syntheses, and action plans with domain-specific stress-testing, iterative deepening, and anti-pattern guards." version: "4.2" license: MIT

Deep Research Skill

A structured multi-agent research pipeline that turns any topic into a comprehensive, fact-checked, bilingual (EN+RU) synthesis with actionable recommendations. Built on Eric Jang's iterative methodology from "As Rocks May Think".

How It Works

The skill orchestrates 10-19 specialized AI agents across 3 mandatory cycles:

Cycle 1: Broad Search    → 4-5 parallel SCOUTs explore the landscape
         Quality Gates   → CRITIC + METHODOLOGIST cross-check findings
         Reflection 1    → Identify gaps, generate competing hypotheses

Cycle 2: Deep Dives      → 2-3 targeted agents test hypotheses + stress-test questions
         Iterative Deep. → Auto-resolve CONTESTED claims (WEAK → follow-up DD)
         Reflection 2    → Convergence analysis, hypothesis verdicts

Cycle 3: Execute          → Python scripts (analysis, models, visualizations)
         Synthesize      → SYNTHESIZER creates integrated document
         Verify          → FACT-CHECKER + CITATION_VERIFIER + DOMAIN_REVIEWER
         Apply           → ACTION MAPPER updates user's protocols/goals

Research Modes

Agent Roles

v4.1 Features

Domain Stress-Test Questions

Each Deep Diver must answer ≥2 mandatory adversarial questions from their domain adapter. Examples:

• Health: "Under what conditions does this intervention become harmful for MY profile?"
• Macro: "What single event makes this consensus irrelevant within 18 months?"
• Company: "What does every successful player understand that customers never say out loud?"
• Science: "What replication failure would collapse this finding?"

Iterative Deepening on CONTESTED Claims

When Cycle 2 convergence finds CONTESTED claims (confidence <0.5), the pipeline auto-spawns targeted Deep Divers to find tiebreaker evidence. Max 2 rounds, 15% budget cap.

Anti-Pattern Guards

Each domain adapter includes a "Common Anti-Patterns" table — pre-flight checklist of domain-specific mistakes (e.g., "citing mouse studies as human evidence" for health, "extrapolating 3-year trend as permanent" for macro).

Company Domain Enhancements

• Raw language preservation: SCOUTs preserve verbatim customer/user quotes with tags
• Opportunity taxonomy: Action Mapper classifies opportunities as CONTRARIAN / TIMING_PLAY / SAFE_BET

v4.2 Features

Temporal Diff (UPDATE mode)

When re-researching a topic that already has a consensus reference, the pipeline auto-detects the previous research and offers UPDATE mode. In UPDATE mode:

• Full research runs identically to fresh mode (SCOUTs do NOT see previous consensus — no anchoring bias)
• TEMPORAL DIFF agent runs after SYNTHESIZER, comparing old and new consensus claim-by-claim
• Produces _temporal_diff.md with 5 categories: CONFIRMED / REVISED / CONTRADICTED / OBSOLETE / NEW
• Adds a ## Temporal Diff Summary to the new consensus_reference.md
• Archives previous version as _previous_consensus_[YYYY_MM].md
• Stability score: CONFIRMED / total — tells you how much the field moved

CLI flags:

• --update — force UPDATE mode (auto-detect previous consensus)
• --fresh — force FRESH mode (ignore previous consensus)
• No flag — auto-detect and ask user

Quick Start

1. Copy the template:

   cp .claude/commands/research/context_template.md .claude/commands/research/context.md
   

2. Fill in your data in context.md:

   • File paths to your profiles, lab results, protocols
   • Key biomarkers (if health/nutrition research)
   • Genetic variants (if relevant)

3. Run:

   /research creatine safety for athletes                      # → health domain (auto)
   /research global electricity demand consensus high 6h       # → macro domain (auto)
   /research AI agents in financial services consensus high 4h  # → company domain (auto)
   /research transformer architectures consensus high 4h       # → science domain (auto)
   

File Structure

.claude/commands/research/
├── SKILL.md              ← You are here (overview + quick start)
├── README.md             ← Brief quick start
├── context_template.md   ← Copy to context.md, fill in your data
├── context.md            ← YOUR private config (gitignored)
│
├── research.md           ← Main entry point (pipeline + architecture)
├── agents.md             ← Detailed agent role descriptions
├── prompts.md            ← Full prompt templates for each agent
├── cycle1.md             ← Cycle 1 instructions (SCOUTs + quality gates)
├── cycle2.md             ← Cycle 2 instructions (Deep Dives + convergence)
├── cycle3.md             ← Cycle 3 instructions (Synthesis + verification)
├── finalize.md           ← Finalization checklist + git/notification
│
├── domains/
│   ├── health.md        ← METHODOLOGIST + MEDICAL_REVIEWER + health consensus template
│   ├── macro.md         ← METHODOLOGIST + MACRO_REVIEWER + macro consensus template
│   ├── company.md       ← METHODOLOGIST + MARKET_REVIEWER + company consensus template
│   └── science.md       ← METHODOLOGIST + METHODOLOGY_REVIEWER + science template
│
└── examples/
    ├── example_output_tree.txt
    └── example_synthesis_excerpt.md

Output Structure

A completed research produces:

YYYY_MM_topic_name/
├── _PROGRESS_LOG.md           # Log + 2 reflections + hypotheses
├── synthesis.md               # Final synthesis EN (10 sections)
├── synthesis_ru.md            # Final synthesis RU (full translation)
├── consensus_reference.md     # Population-level truth (if consensus/full)
├── interaction_map.md         # Cross-domain interactions (if full)
├── unknowns_and_next.md       # Known unknowns + next experiments
├── _critic_review.md          # CRITIC output
├── _methods_review.md         # METHODOLOGIST output (domain-specific)
├── _domain_review.md          # Domain reviewer output (medical/macro/market/methodology)
├── _fact_check.md             # FACT-CHECKER output
├── _citation_audit.md         # CITATION_VERIFIER output
├── _temporal_diff.md           # TEMPORAL DIFF output (UPDATE mode only)
├── _previous_consensus_*.md   # Archived previous version (UPDATE mode only)
├── _action_map.md             # ACTION MAPPER output
├── stream_a_*.md ... stream_e_*.md   # Cycle 1 streams
├── deep_dive_a_*.md ...              # Cycle 2 deep dives
├── data/*.csv                        # Structured data (snake_case)
├── figures/*.png                     # Visualizations (300 DPI)
└── scripts/*.py                      # Analysis scripts

Configuration (context.md)

Dependencies

Required:

• Claude Code CLI (claude)

Optional (for Cycle 3 scripts):

• Python 3.10+ with matplotlib, pandas, numpy
• markdown library (for tools/view_research.py styled viewer)

Optional (for notifications):

• Telegram bot token (for tools/notify_research.py)
• tools/finalize_research.sh automates: git commit + push + notify + viewer

Known Caveats

1. WebSearch often unavailable in subagents — agents synthesize from training data and mark limitations
2. Context window limits — use /compact during long research; all intermediate results saved to disk
3. Agents may not write files — always verify with ls after each agent completes
4. CSV column names vary — Python scripts must include column normalization
5. PEP 668 on macOS — use venv for pip install: python3 -m venv scripts/.venv

Language

All instructions and prompts are in English. The synthesis output is always in English (synthesis.md). If you set preferred_language in your context.md, the pipeline also produces a full translation in your language (e.g., synthesis_ru.md).