Deep Research for Claude Code

Deep research with source quality gates — every claim cited, every source classified, no hallucinated URLs.

Built for engineers and decision-makers, not academia. Optimized for speed and source transparency over exhaustive rigor. Get a trustworthy answer in minutes, not a 50-page paper.

Quick Start

git clone https://github.com/robertnowell/deep-research.git ~/.claude/skills/deep-research

Then in Claude Code:

/deep-research what are the tradeoffs between SQLite and PostgreSQL for new web apps?

What It Does

A 5-phase research pipeline that runs in your Claude Code session:

Phase	What happens	Agents
Analyze	Breaks topic into 3-5 research angles	Main context
Breadth	Parallel web research per angle	2-5 Sonnet agents
Depth	Fills gaps, verifies contradictions	1-3 Sonnet agents
Verify	Fact-checks single-source claims	Haiku agents (thorough only)
Synthesize	Confidence-labeled structured report	Main context

Scope Modes

Control depth vs. speed with a single keyword:

Scope	Breadth agents	Depth phase	Verification	Typical duration	Best for
quick	2	Skipped	Skipped	~2 min	Fast orientation on a topic
standard	3	1-2 agents	Skipped	~5 min	Most research questions
thorough	5	2-3 agents	Yes	~10 min	High-stakes decisions, publishing

/deep-research quick what is WebTransport?
/deep-research thorough should we migrate from Redis to Valkey?

Source Quality System

Every source is classified into four tiers. Every finding must cite at least one.

Tier	Label	Examples
Tier 1	Primary	Official docs, specs, peer-reviewed papers, author's own repo
Tier 2	Curated secondary	Major publications, MDN, Wikipedia (w/ citations), analyst reports
Tier 3	Community	Blog posts, Stack Overflow, forums, tutorials, newsletters
Tier 4	Unverifiable	Social media, AI-generated content, SEO listicles, content farms

Confidence Labels

Each key finding gets a confidence label based on source evidence:

• Established — 2+ independent Tier 1/2 sources agree
• Likely — 1 Tier 1/2 source, or 2+ Tier 3 sources corroborate
• Emerging — Single Tier 2/3 source, no contradiction found
• Contested — Tier 1/2 sources disagree with each other
• Speculative — Only Tier 3/4 sources, or single uncorroborated claim

Hard Quality Gates

Five enforced gates that must pass before the report is shown:

1. Citation requirement — Every finding must cite ≥1 source with a tier annotation
2. No fabricated URLs — Every URL must trace back to an agent's actual output
3. Question answered — Executive summary must state whether the question is fully answered, partially answered, or unanswerable
4. Source quality caveat — Auto-prepended if >50% of sources are Tier 3/4
5. Failed angles disclosed — No silently dropped research angles

How It Compares

	Raw WebSearch	199-bio	Imbad0202	This skill
Parallel agents	No	Yes	Yes (13)	Yes (2-5)
Source classification	None	Opaque 0-100 score	Academic-focused	4-tier transparent system
Quality gates	None	None enforced	PRISMA-style	5 hard gates
Confidence labels	None	None	None	5 levels
Scope scaling	N/A	4 modes	7 modes	3 scopes (quick/standard/thorough)
Codebase-aware	No	No	No	Yes
Failed angle disclosure	N/A	No	No	Required
Target audience	—	General	Academic	Engineers and decision-makers
License	N/A	MIT	CC-BY-NC	MIT

Codebase-Aware Research

If your research question relates to your local project, the skill automatically spawns an Explore agent to search your codebase alongside the web research — connecting external findings to your actual code, config, and architecture.

Example Output

Sample report: Deep research tools in Claude Code (standard scope)

See examples/sample-output.md for the full report demonstrating source tiers, confidence labels, research gaps, and quality gates in action.

Preview of key findings:

Finding	Confidence	Sources
Claude Code ships with capable research tools but has deliberate constraints	Established	Tier 1, Tier 2
The parallel subagent pattern is the foundation of every serious research approach	Established	Tier 1
Community research skills are the fastest path from zero to structured research	Likely	Tier 1, Tier 2
MCP search servers have serious subagent compatibility bugs	Established	Tier 1
MCP tool definitions consume 385 to 17,000+ context tokens per server	Likely	Tier 3
Anthropic frames MCP and Skills as complementary: "MCP = access, Skills = expertise"	Established	Tier 1

Sample report: Inventive uses of deep research (standard scope)

See examples/inventive-uses.md for a report on creative, unexpected applications — M&A due diligence, investigative journalism, OSINT, security research, and research-to-action pipelines.

Preview of key findings:

Finding	Confidence	Sources
The most inventive pattern is research → automated action, not research → report	Likely	Tier 2, Tier 3
M&A due diligence compresses 6-week timelines to 2 weeks	Established	Tier 1, Tier 2
Claude Code Security found 500+ undetected vulnerabilities via reasoning-based research	Established	Tier 1, Tier 2
Investigative journalism workflows replicate published investigations in under an hour	Likely	Tier 3
Scientific research: genome study compressed from months to 20 minutes	Established	Tier 1
Self-improvement systems cross-reference journal entries with Git commits	Emerging	Tier 2

File Structure

deep-research/
├── SKILL.md                      # Skill definition (the pipeline)
├── README.md                     # This file
├── LICENSE                       # MIT
├── references/
│   ├── source-evaluation.md      # Source tier definitions and red flags
│   └── synthesis-patterns.md     # Report templates and synthesis rules
└── examples/
    ├── sample-output.md          # Example: deep research tools comparison
    └── inventive-uses.md         # Example: creative and unexpected applications

License

MIT — see LICENSE.