Deep Research — Claude Code Plugin

Professional multi-agent deep research system for Claude Code. Investigates any topic across 100+ sources with recursive citation chasing, source credibility scoring, evidence triangulation, and knowledge synthesis.

What It Does

When you ask Claude to research something in depth, this skill deploys parallel specialized agents that:

1. Decompose your question into sub-questions using perspective-guided discovery (inspired by Stanford STORM)
2. Search across academic papers, industry blogs, official docs, community forums, news, and YouTube videos
3. Score every source on 6 credibility dimensions (authority, freshness, relevance, evidence quality, peer validation, independence)
4. Chase citations recursively — finding the sources behind each source (backward + forward citation chasing)
5. Triangulate evidence across 3+ independent sources before presenting claims as verified
6. Red-team the synthesis with a QA agent using Chain-of-Verification
7. Deliver a professional report with full bibliography, credibility scores, and confidence levels

Installation

claude plugin add Silence-view/deep-research

Or manually:

git clone https://github.com/Silence-view/deep-research.git ~/.claude/plugins/deep-research

Usage

Invoke via natural language:

"Deep research on multi-agent AI architectures in 2025"
"Fammi una ricerca approfondita sulle best practice per..."
"Comprehensive analysis of WebAssembly adoption trends"
"Survey the literature on source credibility scoring"

Or explicitly:

/deep-research

Research Modes

Mode	Sources	Agents	Duration	Best For
Quick	20-30	3	2-5 min	Focused questions
Standard	50-70	5	5-15 min	General topics
Deep	80-120	6-8	15-30 min	Complex topics
UltraDeep	120-200+	8-10	30-60 min	Literature reviews, critical decisions

Mode is inferred from your language. "Quick overview" -> Quick. "Everything about" -> Deep. "Comprehensive survey" -> UltraDeep.

Pipeline (10 Phases)

Phase 0:  Context Load (CLAUDE.md, ROSETTA.md, project context)
Phase 1:  Query Decomposition (STORM-inspired perspective discovery)
Phase 2:  Research Planning (agent assignment, query generation)
Phase 3:  Primary Research Wave (5-8 parallel agents)
Phase 3.5: Wave 2 Targeted Deep-Dive (gap filling)
Phase 4:  Source Credibility Assessment (6-dimension scoring)
Phase 5:  Recursive Citation Chasing (backward + forward, up to 2 levels)
Phase 6:  Evidence Triangulation (3+ independent sources per claim)
Phase 7:  Synthesis & Knowledge Construction
Phase 8:  Quality Assurance (red-team with Chain-of-Verification)
Phase 9:  Output & Integration (report + CLAUDE.md update)
Phase 10: Post-Research Debrief

Source Credibility Scoring

Every source is scored on 6 dimensions (0.0 to 1.0):

Dimension	Weight	What It Measures
Authority	25%	Institution/author reputation
Freshness	15%	Publication recency (with evergreen adjustment)
Relevance	25%	How directly it answers the question
Evidence Quality	20%	Empirical data vs. opinion
Peer Validation	10%	Peer review, citation count
Independence	5%	Freedom from conflicts of interest

Sources are tiered: Tier 1 (>= 0.75), Tier 2 (0.50-0.74), Tier 3 (0.25-0.49), Discarded (< 0.25).

Scientific Grounding

Built on research from 50+ sources including:

• STORM / Co-STORM (Stanford) — Perspective-guided multi-agent research
• OpenAI Deep Research — Triage-Clarify-Instruct-Research pipeline
• Perplexity Architecture — Three-layer source ranking
• Cochrane Standards — Forward and backward citation chasing
• WebTrust (Tsinghua 2025) — Automated credibility scoring (MAE 0.09)
• PRISM, ReAgent, MA-RAG — Multi-agent retrieval (2025)
• Anthropic Multi-Agent System — Opus + Sonnet architecture (+90.2% vs single-agent)
• Chain-of-Verification (Dhuliawala et al.) — Claim verification prompting

Plugin Structure

deep-research/
├── .claude-plugin/
│   ├── plugin.json
│   └── marketplace.json
├── skills/
│   └── deep-research/
│       ├── SKILL.md                    (main skill - 834 lines)
│       ├── references/
│       │   ├── research-methodology.md (scoring, chasing, agent templates)
│       │   └── output-templates.md     (report templates, evidence schema)
│       └── scripts/
│           └── validate_report.py      (automated report validation)
├── README.md
├── LICENSE
└── CHANGELOG.md

Validation Script

Validate generated reports:

python3 scripts/validate_report.py report.md --mode standard

Checks: executive summary, key findings, bibliography, source count, tier distribution, citations, URLs, methodology, limitations, report length, confidence tags.

Requirements

• Claude Code with Agent tool access
• WebSearch and WebFetch tools available
• Works best with Opus model (recommended) or Sonnet

License

MIT