Multi AI Research

Multi-AI Research & Analysis

Overview

Harnesses three AI systems (Claude via Task tool, Gemini CLI, Codex CLI) for comprehensive research and analysis with multi-perspective verification and iterative refinement.

Purpose: Produce analysis more thorough than any single AI could achieve through specialized roles, cross-validation, and systematic verification.

Key Innovation: Not just parallel execution - specialized research roles with cross-verification and iterative refinement until production-ready (quality ≥95/100, 100% citations, zero gaps).

The 3 AI Systems:

1. Claude Subagents (via Task tool) - Documentation, codebase analysis, synthesis
2. Gemini CLI - Web research, latest trends, community practices
3. Codex CLI - GitHub patterns, code examples, deep reasoning

Quality Guarantees:

• ✓ 100% coverage - All objectives addressed, zero gaps
• ✓ 100% citations - Every claim sourced (file:line or URL)
• ✓ Multi-perspective - 3 AI systems cross-validated
• ✓ ≥95/100 quality - Verified through 3-pass system
• ✓ Actionable - Specific recommendations with examples
• ✓ Resumable - External memory enables multi-session work

When to Use

Use this skill for:

Security Analysis

• Authentication/authorization assessment
• Vulnerability identification
• Best practice validation
• OWASP Top 10 coverage
• Penetration testing preparation

Architecture Analysis

• System design review
• Component mapping
• Integration pattern analysis
• Scalability assessment
• Technical debt evaluation

Code Quality Analysis

• Pattern detection
• Code smell identification
• Complexity metrics
• Refactoring opportunities
• Best practice adherence

Performance Analysis

• Bottleneck identification
• Algorithm complexity
• Resource usage patterns
• Optimization opportunities
• Benchmark analysis

Research Synthesis

• Multi-source research compilation
• Best practice identification
• Technology evaluation
• Pattern discovery
• Trend analysis

Comprehensive Reviews

• Pre-production audit
• System health check
• Compliance verification
• Documentation audit
• Knowledge transfer

Quick Start

Option 1: Automated Script

# Run complete analysis automatically
bash .claude/skills/multi-ai-research/scripts/analyze.sh "Security analysis of authentication system"

This will:

1. Create analysis plan
2. Launch parallel research (Claude + Gemini + Codex)
3. Perform deep analysis
4. Synthesize and verify
5. Iterate if needed
6. Generate final report

Option 2: Interactive Mode

Ask Claude Code to use this skill:

"Use multi-ai-research to analyze [objective]"

Claude will:

1. Create comprehensive analysis plan
2. Coordinate all three AI systems
3. Synthesize findings
4. Verify quality
5. Iterate until ≥95 quality
6. Deliver final report

The 5-Phase Pipeline

Phase 1: Planning & Strategy

Duration: 5-10 minutes Output: .analysis/ANALYSIS_PLAN.md

Claude creates comprehensive plan:

• Defines objectives and scope
• Plans file reading strategy (glob → grep → read)
• Assigns tasks to AI systems
• Sets verification criteria
• Defines success thresholds

Phase 2: Parallel Research

Duration: 10-20 minutes Output: .analysis/research/*.md

All three systems research simultaneously:

Claude Subagent:

• Official documentation analysis
• Codebase examination (progressive disclosure)
• Architecture mapping
• Pattern identification

Gemini CLI:

• Web research (latest 2024-2025)
• Community best practices
• Industry trends
• Common pitfalls

Codex CLI:

• GitHub pattern analysis
• Code examples from top repos
• Implementation references
• Testing strategies

Phase 3: Deep Analysis

Duration: 15-30 minutes Output: .analysis/analysis/code-patterns.md

Claude Analysis Agent with extended thinking:

• Progressive codebase analysis
• Pattern recognition across sources
• Architecture mapping
• Metrics calculation
• Risk assessment

Phase 4: Synthesis & Verification

Duration: 10-20 minutes Outputs:

• .analysis/SYNTHESIS_REPORT.md
• .analysis/verification/cross-check.md

Synthesis (Claude with extended thinking):

• Read all research findings
• Identify themes across sources
• Resolve contradictions
• Create unified narrative
• Full citations

Verification (Verification Subagent):

• 3-pass verification (completeness, accuracy, quality)
• Cross-source validation
• Citation checking
• Gap analysis
• Quality scoring

Phase 5: Iteration (if needed)

Duration: 10-30 minutes Output: .analysis/iterations/ITERATION_2.md

If quality <95 or gaps exist:

• Targeted research for gaps
• Quality improvements
• Re-verification
• Repeat until ≥95

Phase 6: Final Report

Duration: 5-10 minutes Output: .analysis/ANALYSIS_FINAL.md

Comprehensive final report:

• Executive summary
• Complete findings
• All sources synthesized
• Prioritized recommendations
• Implementation guidance
• Full citations

Total Time: 45-90 minutes for comprehensive analysis

Analysis Types

Security Analysis

What it checks:

• Authentication/authorization patterns
• Input validation
• Secret management
• Injection vulnerabilities (SQL, XSS, etc.)
• Dependency vulnerabilities
• Rate limiting
• Session security

Example:

Use multi-ai-research for "Security audit of authentication system"

Output:

• Critical/High/Medium/Low priority issues
• OWASP Top 10 coverage
• Code examples with file:line
• Specific remediation steps
• Industry best practices comparison

Architecture Analysis

What it examines:

• System components and boundaries
• Integration patterns
• Data flow
• Dependency relationships
• Scalability considerations
• Design patterns used

Example:

Use multi-ai-research for "Architecture analysis of microservices system"

Output:

• Component map with relationships
• Integration pattern analysis
• Scalability assessment
• Technical debt identification
• Refactoring recommendations

Code Quality Analysis

What it analyzes:

• Code patterns and organization
• Complexity metrics
• Code smells
• Best practice adherence
• Test coverage
• Documentation quality

Example:

Use multi-ai-research for "Code quality assessment for ./src"

Output:

• Quality score with breakdown
• Pattern analysis
• Refactoring priorities
• Specific code improvements
• Complexity hotspots

Performance Analysis

What it identifies:

• Algorithm complexity
• Bottlenecks
• Resource usage patterns
• Database query efficiency
• Network call patterns

Example:

Use multi-ai-research for "Performance bottleneck identification"

Output:

• Bottleneck analysis with file:line
• Optimization opportunities
• Before/after estimations
• Implementation guidance

Research Synthesis

What it compiles:

• Official documentation
• Web best practices
• GitHub patterns
• Industry standards
• Community insights

Example:

Use multi-ai-research for "Research GraphQL federation patterns 2024-2025"

Output:

• Multi-source synthesis
• Consensus findings (all sources agree)
• Multiple perspectives (sources differ)
• Code examples
• Implementation recommendations

How It Works

Progressive Disclosure

Never reads files blindly. Always uses 3-level approach:

Level 1: Metadata (glob) - ~50 tokens

glob "**/*.{ts,js,py}"     # Understand structure
glob "**/*.md"             # Find documentation
glob "**/package.json"     # Check dependencies

Level 2: Patterns (grep) - ~5k tokens

grep "export class|interface" --glob "**/*.ts"
grep "TODO|FIXME|BUG" --glob "**/*"
grep "password|secret|token" --glob "**/*.ts"

Level 3: Reading (read) - ~50k tokens

read "src/auth/login.ts"   # Only critical files
read "docs/architecture.md"

Result: 90%+ reduction in unnecessary file reads

External Memory Architecture

All state saved to files, not context:

.analysis/
├── ANALYSIS_PLAN.md           # Strategy and assignments
├── research/
│   ├── claude-docs.md         # Claude research
│   ├── gemini-web.md          # Gemini research
│   └── codex-github.md        # Codex research
├── analysis/
│   ├── code-patterns.md       # Pattern analysis
│   └── architecture-map.md    # System map
├── verification/
│   └── cross-check.md         # Verification results
├── iterations/
│   ├── ITERATION_1.md         # First pass
│   └── ITERATION_2.md         # Gap fills
└── ANALYSIS_FINAL.md          # Complete report

Benefits:

• Survives context window limits
• Enables multi-session analysis
• Resumable from any checkpoint
• No information loss

Cross-Validation Pattern

High Confidence (★★★★★): All 3 sources agree + code verification Medium Confidence (★★★☆☆): 2/3 sources agree Requires Investigation (★★☆☆☆): Sources conflict

Example:

## JWT Implementation (High Confidence ★★★★★)

**Claude**: "Uses JWT with HS256" (src/auth/jwt.ts:15)
**Gemini**: "HS256 is industry standard 2024" (URL)
**Codex**: "150+ repos use HS256 pattern" (GitHub)
**Code**: Verified at src/auth/jwt.ts:18-22

**Recommendation**: Implementation correct per standards

Quality Scoring

Comprehensive rubric (0-100):

• Comprehensiveness (/20): All aspects covered
• Accuracy (/20): All claims sourced and verified
• Specificity (/20): File:line precision, not vague
• Actionability (/20): Specific recommendations
• Consistency (/20): No contradictions

Quality Gates:

• ≥95: Production-ready
• 85-94: Needs minor refinement
• 75-84: Needs iteration
• <75: Requires rework

Iterative Refinement

Iteration 1 (Breadth): Broad coverage, identifies gaps Iteration 2 (Depth): Fill gaps, improve quality Iteration 3 (Polish): Final verification, perfection

Automatic iteration until:

• Quality ≥95
• Citation coverage = 100%
• Critical gaps = 0

AI System Roles

Claude Subagents (via Task tool)

Research Agent (Haiku):

• Progressive disclosure expert
• Documentation analysis
• Codebase examination
• Pattern detection

Analysis Agent (Sonnet):

• Extended thinking for synthesis
• Multi-source integration
• Pattern recognition
• Architectural insights

Verification Agent (Haiku):

• 3-pass verification
• Citation checking
• Gap analysis
• Quality scoring

Gemini CLI

Strengths:

• Native web search
• Latest trends (2024-2025)
• Community practices
• Multimodal analysis (if needed)

Use for:

• Best practice research
• Industry standards
• Latest vulnerabilities
• Framework comparisons

Codex CLI

Strengths:

• GitHub integration
• Code pattern search
• Deep reasoning (o3 model)
• Implementation examples

Use for:

• Code examples
• Design patterns
• Architecture reasoning
• Testing strategies

Configuration

Prerequisites

Required:

• Claude Code (with Task tool access)

Optional but Recommended:

• Gemini CLI: npm install -g @google/gemini-cli
• Codex CLI: npm install -g @openai/codex

Note: Skill works with Claude-only fallback if Gemini/Codex unavailable.

Gemini CLI Setup

# Install
npm install -g @google/gemini-cli

# Authenticate (OAuth - free)
gemini
# Follow browser authentication

# Test
gemini -p "test prompt"

Codex CLI Setup

# Install
npm install -g @openai/codex

# Authenticate (ChatGPT Plus/Pro account)
codex login
# Follow browser authentication

# Test
codex exec "test prompt"

Model Selection

Claude:

• Haiku: Research & verification (fast, efficient)
• Sonnet: Analysis & synthesis (balanced)
• Opus: Complex reasoning (if needed)

Gemini:

• gemini-2.5-flash: Quick research
• gemini-2.5-pro: Complex analysis

Codex:

• gpt-5.1-codex: Standard tasks
• o3: Deep architectural reasoning
• o4-mini: Quick operations

Examples

Example 1: Security Analysis

Objective: "Security audit of authentication system"

Phase 2 - Parallel Research:
├─ Claude: Analyzes src/auth/* for patterns
├─ Gemini: Researches "OAuth 2.0 security best practices 2024"
└─ Codex: Finds GitHub examples of secure auth

Phase 3 - Analysis:
└─ Claude: Identifies 3 critical, 5 high priority issues

Phase 4 - Synthesis:
└─ All agree: Missing rate limiting (CRITICAL)
   - Claude: No rate limit found in src/auth/login.ts
   - Gemini: OWASP recommends max 5 attempts/hour
   - Codex: 150+ repos use express-rate-limit
   - Recommendation: Implement with Redis backend

Final Report:
├─ Executive summary
├─ 8 issues (3 critical, 5 high) with fixes
├─ OWASP Top 10 coverage
├─ Specific code examples
└─ Priority implementation plan

Quality: 97/100 ✓

Example 2: Architecture Analysis

Objective: "Analyze microservices architecture"

Phase 2:
├─ Claude: Maps services via glob + grep
├─ Gemini: Researches microservices patterns 2024
└─ Codex: Finds service mesh examples

Phase 3:
└─ Claude: Identifies 7 services, 12 integration points

Phase 4:
└─ Synthesis: Service communication patterns
   - Consensus: REST for external, gRPC for internal
   - Trade-offs documented
   - Scaling strategies from Codex examples

Final Report:
├─ Component map (7 services, dependencies)
├─ Integration analysis (12 patterns)
├─ Scalability assessment
└─ Modernization recommendations

Quality: 96/100 ✓

Example 3: Research Synthesis

Objective: "Research state management patterns for React 2024"

Phase 2:
├─ Claude: Reviews React docs + examples
├─ Gemini: Web research "React state management 2024"
└─ Codex: Analyzes top 50 React repos

Phase 3:
└─ Pattern analysis: 5 major approaches identified

Phase 4:
└─ Synthesis by use case:
   - Small apps: Context (all sources agree)
   - Medium apps: Zustand (Gemini + Codex recommend)
   - Large apps: Redux Toolkit (battle-tested, Codex data)
   - Server state: TanStack Query (trending, Gemini research)

Final Report:
├─ Decision tree by project size
├─ Pros/cons with sources
├─ Migration strategies
└─ Code examples from Codex

Quality: 98/100 ✓

Best Practices

1. Be Specific with Objectives

❌ "Analyze the code"
✅ "Security analysis of authentication module for OWASP Top 10 compliance"

2. Trust the Verification

Multi-pass verification catches issues. If quality <95, iteration happens automatically.

3. Review External Memory

Check .analysis/ folder during execution to see progress.

4. Leverage Citations

Every claim has file:line or URL. Use for validation and deep dives.

5. Multi-Session Projects

Large projects can span sessions:

Session 1: Initial analysis → ITERATION_1.md
Session 2: Gap filling → ITERATION_2.md
Session 3: Final polish → ANALYSIS_FINAL.md

6. Check All Three Perspectives

High-value insights often come from comparing AI perspectives.

Troubleshooting

Low Quality Score (<95)

Cause: Gaps in coverage or missing citations Solution: Automatic iteration 2 fills gaps Check: .analysis/verification/cross-check.md for details

Missing Citations

Cause: Verification flags uncited claims Solution: Iteration adds missing attributions Prevention: All agents trained to cite sources

Gemini/Codex Unavailable

Fallback: Claude-only analysis with warning Impact: Reduced perspectives but still comprehensive Install: npm install -g @google/gemini-cli @openai/codex

Conflicting Information

Resolution: Synthesis phase investigates conflicts Method: Check ground truth (actual code/docs) Output: Documented reasoning for resolution

Related Skills

• anthropic-expert: Anthropic product expertise
• codex-cli: Codex integration patterns
• gemini-cli: Gemini integration patterns
• tri-ai-collaboration: General tri-AI workflows
• analysis: Code/skill/process analysis

Quick Reference

Command Line

# Full automated analysis
bash .claude/skills/multi-ai-research/scripts/analyze.sh "objective"

# Interactive with Claude Code
# Just ask: "Use multi-ai-research for [objective]"

File Locations

Quality Metrics

Analysis Time Estimates

multi-ai-research delivers production-ready analysis through systematic multi-AI collaboration, rigorous verification, and iterative refinement - ensuring nothing is missed and every claim is verified.