Skip to content

zdenekmach/deep-research

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

10 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

deep-research

Version: 1.6.0

Multi-pass parallel research plugin for Claude Code. Systematically explores topics from multiple angles using parallel agents, gathers 25+ quality sources, and synthesizes findings into actionable insights with practical context.

What's Included

Research Commands

Command Purpose Sources Time
deep-research Multi-pass parallel research with SIFT evaluation 25-90 10-18 min
research Quick to moderate evidence-driven research 3-7 2-5 min

Quality & Review Commands

Command Purpose
critique Logical analysis of arguments — fallacies, evidence gaps, counter-arguments
verify Three-level content verification — facts, style, sources, implementation completeness
humanize Remove AI-generated text patterns while preserving content and expertise

Specialized Agents

Agent Model Purpose
deep-research-agent Opus Multi-pass research with SIFT framework, credibility scoring, 25+ sources
research-agent Sonnet Topic decomposition, synthesis, entity detection
fact-check-agent Haiku Verify claims, citations, links, statistics
critic-agent Sonnet 4D content critique (accuracy, conciseness, readability, actionability)

Installation

# Load directly
claude --plugin-dir ./deep-research

# Or clone and load
git clone https://github.com/zdenekmach/deep-research.git
claude --plugin-dir ./deep-research

Commands are available as /deep-research:<command>.

Key Features

Topic Decomposition

Before searching, the plugin decomposes your topic into 4-6 independent research streams. For business topics it uses a four-stream pattern (MARKET / TECHNOLOGY / COMPETITION / RISKS); for other domains it applies domain-specific decomposition.

Why this matters: single-pass search follows keywords from your prompt. If you ask about "longevity biotech — senolytics, rapamycin, epigenetic clocks, blood plasma," you get answers about exactly those four topics. GLP-1 agonists, NAD+ precursors, or cellular reprogramming won't surface because you didn't ask. Deep-research decomposes based on problem structure, not just the prompt — and catches what you didn't know to ask about.

A validation checkpoint runs before any search begins: Are all key dimensions covered? Is there an obvious missing perspective? Are streams specific enough to search but broad enough for context?

SIFT Framework & Credibility Scoring

Every source is evaluated using the SIFT methodology:

  • Stop — evaluate before using
  • Investigate — who is the author/organization?
  • Find — is there a better source?
  • Trace — where does the original claim come from?

Each source gets a credibility score:

Score Source Type Example
+3 Peer-reviewed Nature, Lancet, NEJM
+2 Institutional WHO, FDA, McKinsey
+1 Expert Named expert with credentials
0 General media News articles, industry press
-1 User-generated Reddit, Medium without credentials
-2 Anonymous/promotional Vendor whitepapers, anonymous posts

Key claims require sources with score >= +1. Market projections from research firms get 0 and are explicitly flagged.

Signal Map (Adaptive Research)

After initial broad search, each research stream is classified:

  • STRONG (8+ sources, high credibility) — deep dive with nuanced comparison, full detail file
  • MODERATE (4-7 sources) — standard deep dive
  • WEAK (<4 sources) — directional patterns only, no precise claims, no separate file

Research effort adapts to signal strength: more tokens and recursive exploration for strong streams, broad aggregation for weak ones.

2D Confidence Model

Recommendations use a two-dimensional model:

Sources Converge Sources Contradict
Strong signal Precise claim: "Do X because Y" Nuanced position: "X over Y, but consider Y when Z"
Weak signal Directional trend: "Evidence suggests X" Unknown: "Insufficient data. Verify Z"

No "you could do A or B" — always opinionated with reasoning, but precision matched to signal strength.

Practical Layer (v1.6.0)

Every research stream answers: What does this cost? Who offers it? How accessible is it today? What can someone actually do with this information? Includes concrete numbers (prices, timelines, providers) where available. If no practical data exists, it's explicitly flagged as a gap.

Adjacent Topics (v1.6.0)

Every stream identifies 2-3 related topics that the research touches but doesn't cover in depth. Shows the reader what's beyond the edges of this research and where they could continue.

Narrative Summary (v1.6.0)

The summary file reads as a cohesive narrative, not a structured data dump. Findings are woven across streams into a story — not siloed per section. Think "analyst report for a board member," not "database export."

Numbered Bibliography (v1.6.0)

Inline references [1][2][3] throughout the text. Sources grouped by credibility tier (High/Medium/Low) at the end of each file. Format: [N] Author/Org — "Title" — URL (Date) — Credibility: +X.

Comparison

Standard research Gemini Deep Research deep-research plugin
Sources 3-7 30-45 25-90
Passes 1-2, serial unknown (black-box) 5+ phases, parallel (3-4 agents)
Decomposition None Implicit (thematic sections) Explicit 4-6 streams with Signal Map
Source evaluation None None SIFT framework + credibility -2 to +3
Confidence model 1D (high/medium/low) None 2D matrix: signal strength x source convergence
Conflicts Flag Avoids Explicit resolution with reasoning
Practical Layer No Partial (prices) Yes (prices, providers, budget stacks)
Adjacent Topics No Partial (AI/tech) Yes (2-3 per stream)
Narrative flow No Strong Yes (v1.6.0)
Output Single file Single file Modular: summary + detail files
Time 2-5 min 5-10 min 10-18 min

Language Configuration

By default, subagent outputs are in English (prevents language mixing during synthesis). Final user-facing files follow your configured output language.

To configure output language, add to your project's CLAUDE.md:

## Output Language
- Final output: English
- Technical terms: standard English terminology

Or for other languages:

## Output Language
- Final output: Czech with full diacritics
- Technical terms: First mention = Czech + English in parentheses

Directory Structure

deep-research/
├── .claude-plugin/plugin.json    # Plugin manifest
├── commands/                     # User-invokable commands
│   ├── deep-research.md          # Multi-pass parallel research
│   ├── research.md               # Quick research
│   ├── critique.md               # Argument analysis
│   ├── verify.md                 # Content verification
│   └── humanize.md               # AI pattern removal
├── agents/                       # Specialized agent definitions
│   ├── deep-research-agent.md    # Opus-powered deep research
│   ├── research-agent.md         # Sonnet research agent
│   ├── fact-check-agent.md       # Haiku fact-checker
│   └── critic-agent.md           # Sonnet content critic
└── skills/                       # Agent skills and references
    ├── research/                 # Core research skill
    │   ├── SKILL.md              # Full research methodology
    │   ├── templates/            # Output templates (research + deep-research + market-report)
    │   ├── references/           # Anti-patterns, methodology
    │   └── domains/              # Domain-specific research templates
    └── subagents/                # Detailed agent methodologies
        ├── deep-research-agent.md
        ├── research-agent.md
        ├── fact-check-agent.md
        └── critic-agent.md

Domain Templates

The plugin includes domain-specific research templates that guide topic decomposition for common scenarios:

Template Use Case Key Output
market-entry New market evaluation GO/NO-GO with attractiveness score
technology-evaluation Technology adoption decisions Weighted decision matrix
regulatory-landscape Compliance & regulation mapping Gap analysis + remediation roadmap
wildlife-photography Example non-business template Species guide + photography settings

Create your own templates in skills/research/domains/ — the research agent automatically checks for matching domain templates during topic decomposition.

Output Templates

Template Flag Structure
research-output.md (default for /research) Executive Summary → Findings → Contradictions → Sources
deep-research-output.md (default for /deep-research) Narrative Summary + Signal Map + Practical Layer → Detail files per stream
market-report.md --format market-report 11-chapter market report (TAM/SAM/SOM, Porter's, PESTLE, SWOT...)

Output

Research outputs are saved to outputs/research/ by default. Each output includes:

  • YAML frontmatter (type, date, confidence, source count)
  • Narrative executive summary (v1.6.0 — woven across streams, not siloed)
  • Signal Map with stream classification
  • Detailed findings with inline references [N]
  • Practical Layer per stream (costs, providers, accessibility)
  • Adjacent Topics per stream (2-3 related areas)
  • Numbered bibliography grouped by credibility tier
  • Contradiction table with explicit resolutions

Requirements

  • Claude Code v1.0.33+
  • Internet access (for WebSearch and WebFetch)
  • Optional: SearXNG for enhanced search (218 engines)

Changelog

  • 1.6.0 (2026-03-31): Narrative Summary, Practical Layer, Adjacent Topics, Numbered Bibliography. Inspired by comparative analysis with Gemini Deep Research.
  • 1.5.0 (2026-02-09): Adaptive Resolution — Signal Map, 2D confidence model, adaptive token allocation.
  • 1.1.0 (2026-01-23): Recursive Depth Pattern — dynamic sub-agent spawning for complex topics.
  • 1.0.0 (2026-01-13): Initial release.

License

MIT — use freely, attribution appreciated.

About

Multi-pass parallel research plugin for Claude Code — 25+ sources, SIFT evaluation, fact-checking, content critique

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors