Skills deep-research

📦

deep-research

Name: deep-research
Author: lumacoder

Medium Risk ⚙️ External commands📁 Filesystem access🌐 Network access

Conduct deep research with citation-backed verification

Also available from: 21pounder,21pounder,sickn33,199-biotechnologies,199-biotechnologies

Manual research takes hours and produces inconsistent results with unverified claims. This skill runs an 8-phase pipeline that collects, verifies, and synthesizes information from 10+ sources into professional reports.

Supports: Claude Codex Code(CC)

⚠️ 64 Poor

Download the skill ZIP

Upload in Claude

Go to Settings → Capabilities → Skills → Upload skill

Toggle on and start using

Test it

Using "deep-research". Research the state of quantum computing in 2025 - current capabilities, major players, and commercial applications

Expected outcome:

A structured research report with executive summary, 5 detailed findings with citations, source credibility analysis, synthesis of market trends, limitations and caveats, actionable recommendations, and a verified bibliography of 22 sources.

Using "deep-research". Compare PostgreSQL vs MongoDB for a high-traffic analytics platform

Expected outcome:

A comparison report covering performance benchmarks, scalability patterns, cost analysis, ecosystem maturity, team expertise requirements, and a final recommendation - each claim backed by 3+ verified sources with DOI or URL verification.

Security Audit

Medium Risk

v1 • 4/16/2026

This skill performs legitimate research orchestration with citation management and report validation. Static analysis flagged 435 patterns across 21 files, but the majority are false positives: the scanner treated markdown documentation (ARCHITECTURE_REVIEW.md, QUICK_START.md, SKILL.md, etc.) as executable code, flagging fenced code examples as shell execution. All Python scripts were reviewed and found to be benign - they handle research state management, citation tracking, source credibility evaluation, and report validation. Notable but acceptable risks: MD5 is used for non-cryptographic citation ID deduplication (citation_manager.py:69), and scripts make HTTP requests to public DOI/URL services for citation verification (verify_citations.py:130). These are documented limitations, not security concerns.

Files scanned

6,710

Lines analyzed

findings

Total audits

High Risk Issues (1)

scripts/citation_manager.py:69

MD5 used for identifier generation

MD5 hash algorithm is used to generate citation IDs from URLs in citation_manager.py. While MD5 is cryptographically broken, this usage is for non-security deduplication purposes only. An attacker could theoretically craft URL collisions, but this has no security impact in the research context.

Medium Risk Issues (2)

scripts/verify_citations.py:130 scripts/verify_citations.py:163-166

HTTP requests to external services

The citation verifier makes HTTP requests to doi.org and user-provided URLs for citation verification. These requests go to legitimate public services but could be a concern if the skill processes untrusted URLs without sanitization.

scripts/research_engine.py:128-129 scripts/research_engine.py:74-75

File writes to user home directory

Research engine writes state files and reports to Path.home()/.claude/research_output/. This creates files in the user home directory but uses a dedicated subdirectory, which is appropriate for this use case.

Low Risk Issues (3)

ARCHITECTURE_REVIEW.md:40 QUICK_START.md:13-15 SKILL.md:26-45

False positive: markdown code examples flagged as shell execution

Static scanner flagged 237 instances of shell-like syntax in markdown documentation files (ARCHITECTURE_REVIEW.md, QUICK_START.md, SKILL.md, AUTONOMY_VERIFICATION.md, etc.). These are fenced code blocks containing documentation examples, not executable code. The scanner cannot distinguish between markdown content and executable scripts.

ARCHITECTURE_REVIEW.md:4 SKILL.md:3

False positive: weak crypto flags in documentation

Static scanner flagged references to cryptographic concepts in markdown documentation files. These are instructional text, not actual cryptographic implementations. Only actual code usage (MD5 in citation_manager.py) is relevant.

SKILL.md:221 QUICK_START.md:29

False positive: hidden file references in documentation

Scanner flagged references to paths like ~/.claude in markdown files as hidden file access. These are documentation instructions telling the user where files will be saved, not actual filesystem access patterns.

Risk Factors

⚙️ External commands (5)

SKILL.md:26-45 SKILL.md:136-146 SKILL.md:451-453 scripts/research_engine.py:167-176 scripts/research_engine.py:257-275

📁 Filesystem access (5)

scripts/research_engine.py:74 scripts/citation_manager.py:152 scripts/validate_report.py:328 SKILL.md:221 SKILL.md:451

🌐 Network access (5)

scripts/verify_citations.py:130 scripts/citation_manager.py:162 scripts/source_evaluator.py:270 templates/report_template.md:322 tests/fixtures/valid_report.md:89

Audited by: claude

Quality Score

Architecture

100

Maintainability

Content

Community

Security

Spec Compliance

What You Can Build

Competitive landscape analysis

Research competitors, market positioning, and industry trends to inform strategic decisions with verified data and multiple source perspectives.

Academic literature review

Systematically collect, evaluate, and synthesize academic papers with DOI verification and credibility scoring for research projects.

Technology comparison report

Compare technology stacks, frameworks, or tools with evidence-backed analysis, verified benchmarks, and structured recommendations.

Try These Prompts

Quick topic overview

Research the current state of [topic] and provide a summary of the main trends, key players, and open questions. Use quick mode for a broad overview.

Standard comparison research

Compare [Option A] versus [Option B] for [use case]. Include performance benchmarks, adoption metrics, community support, and a recommendation based on verified sources.

Deep competitive analysis

Conduct a comprehensive competitive analysis of the [industry] market. Cover market leaders, emerging challengers, pricing strategies, feature comparisons, and growth trends with verified data from 15+ sources.

Ultra-deep academic research

Research [research question] using academic and industry sources. Verify all claims with 3+ independent sources, evaluate source credibility, and produce a full research report with bibliography. Use deep or ultradeep mode.

Best Practices

Use deep or ultradeep mode for research that influences important business or technical decisions
Run the built-in validation scripts after report generation to catch formatting issues and incomplete citations
Always review the triangulation phase output to understand which claims were verified versus flagged as unverified

Avoid

Using deep-research for simple fact lookups that a single search can answer
Accepting research output without reviewing the source credibility scores and verification status
Running multiple research sessions in parallel on the same topic without coordinating scope

Frequently Asked Questions

When should I use deep-research instead of a regular search?

Use deep-research when you need 10+ verified sources, cross-validated claims, or a structured report. For single facts or quick answers, a regular search is faster.

How does the skill verify citation authenticity?

It checks DOI resolution through doi.org, verifies URL accessibility, detects hallucination patterns in titles, and cross-references metadata for consistency.

What research modes are available?

Four modes: quick (3 phases, 2-5 minutes), standard (6 phases, 5-10 minutes), deep (8 phases, 10-20 minutes), and ultradeep (8+ phases, 20-45 minutes).

Can I resume an interrupted research session?

Yes. Research state is saved after each phase to the .claude/research_output directory. Use the --resume flag with the saved state file path.

Does this skill work offline?

No. It requires internet access for web search, source retrieval, and citation verification. The validation scripts can run offline on existing reports.

How does the skill handle source credibility?

It scores sources on domain authority, recency, expertise indicators, and bias detection. Academic and government domains receive higher trust scores than blogs or unknown sites.