Dual AI Review

AI Review Skill (Gemini + Codex)

Overview

This skill enables multi-AI review workflows by leveraging both Google Gemini CLI and OpenAI Codex CLI for comprehensive analysis. Ideal for:

• Paper/thesis review and validation
• Code review from multiple perspectives
• Cross-checking AI outputs for accuracy
• Research verification and fact-checking

Prerequisites

Required Tools

• Gemini CLI: gemini --version (v0.21.1+ for Gemini 3)
• Codex CLI: codex --version

Configuration

• Gemini: Authenticated via gemini login
• Codex: Config at ~/.codex/config.toml

Enable Gemini 3 Models

# Run these commands in Gemini CLI interactive mode:
/settings  # Toggle "Preview Features" to true
/model     # Select "Auto (Gemini 3)"

Workflow Checklist

For every AI review task, follow this sequence:

1. Verify tool availability:

   gemini --version && codex --version
   

2. Ask user for review parameters via AskUserQuestion:

   • Review Mode: gemini-only, codex-only, dual-review (recommended), sequential
   • Review Type: paper, code, logic, general
   • Depth: quick, standard, thorough

3. Execute review based on mode (see Command Patterns below)

4. Synthesize results: Combine outputs, highlight agreements/disagreements

5. Report to user: Summary, confidence level, next steps

Command Patterns

Gemini CLI Commands

# Basic Review
gemini "Review the following content for [ASPECT]: [CONTENT]" 2>/dev/null

# Gemini 3 Flash - SWE-bench 78%, 3x faster, best value
gemini -m gemini-3-flash-preview "Deep analysis of: [CONTENT]" 2>/dev/null

# Gemini 3 Pro - Complex reasoning, highest quality
gemini -m gemini-3-pro-preview "Deep analysis of: [CONTENT]" 2>/dev/null

# File Review via stdin
cat file.py | gemini -m gemini-3-flash-preview "Review this code" 2>/dev/null

# YOLO Mode (Auto-approve)
gemini -y "Analyze and review: [CONTENT]" 2>/dev/null

# Resume Session
gemini -r latest "Continue the previous review" 2>/dev/null

Codex CLI Commands

# Code Review - use gpt-5.2-codex (best code ability, 400K context)
# NOTE: Codex does NOT support stdin pipe, include content directly in prompt
codex exec -m gpt-5.2-codex -c model_reasoning_effort="high" -s read-only \
  --skip-git-repo-check --full-auto "Review this code: [CONTENT]" 2>/dev/null

# Paper/Text Review - use gpt-5.1-codex-max (better text ability, supports million tokens)
codex exec -m gpt-5.1-codex-max -c model_reasoning_effort="xhigh" -s read-only \
  --skip-git-repo-check --full-auto "Comprehensive review: [CONTENT]" 2>/dev/null

# File Review - read file content first, then include in prompt
# Example: CODE=$(cat file.py) && codex exec -m gpt-5.2-codex ... "Review: $CODE"

# Built-in Code Review (for git repos)
codex exec review -m gpt-5.2-codex --skip-git-repo-check --full-auto 2>/dev/null

Review Modes

Mode 1: Dual Review (Recommended)

# Step 1: Gemini Review
gemini -m gemini-3-flash-preview "Review this [paper/code] for:
1. Logical consistency
2. Technical accuracy
3. Potential issues
Content: [CONTENT]" 2>/dev/null > /tmp/gemini_review.txt

# Step 2: Codex Review (use 5.2-codex for code, 5.1-codex-max for text)
codex exec -m gpt-5.2-codex -c model_reasoning_effort="high" -s read-only \
  --skip-git-repo-check --full-auto \
  "Review this [paper/code] for:
  1. Logical consistency
  2. Technical accuracy
  3. Potential issues
  Content: [CONTENT]" 2>/dev/null > /tmp/codex_review.txt

# Step 3: Compare and synthesize (Claude does this)

Mode 2: Sequential Review (Deep Analysis)

# Step 1: Gemini Initial Review
GEMINI_OUTPUT=$(gemini -m gemini-3-flash-preview "Initial review of: [CONTENT]" 2>/dev/null)

# Step 2: Codex Validates Gemini Review
codex exec -m gpt-5.1-codex-max -c model_reasoning_effort="high" -s read-only \
  --skip-git-repo-check --full-auto \
  "Validate and expand on this review:
  Original Content: [CONTENT]
  Initial Review: $GEMINI_OUTPUT" 2>/dev/null

Mode 3: Single AI Review

# Gemini Only
gemini -m gemini-3-flash-preview -y "Comprehensive review of: [CONTENT]" 2>/dev/null

# Codex Only (Code)
codex exec -m gpt-5.2-codex -c model_reasoning_effort="xhigh" -s read-only \
  --skip-git-repo-check --full-auto "Comprehensive review: [CONTENT]" 2>/dev/null

# Codex Only (Paper/Text)
codex exec -m gpt-5.1-codex-max -c model_reasoning_effort="xhigh" -s read-only \
  --skip-git-repo-check --full-auto "Comprehensive review: [CONTENT]" 2>/dev/null

Review Type Templates

Paper/Thesis Review

# Gemini - Focus on logic and clarity (use 3.0 Pro for best reasoning)
gemini -m gemini-3-pro-preview "As an academic reviewer, analyze this paper section:
1. Logical flow and argumentation
2. Clarity of expression
3. Methodology soundness
4. Citation appropriateness
5. Potential weaknesses

Content:
[PAPER_CONTENT]" 2>/dev/null

# Codex - Focus on technical accuracy (use 5.1-codex-max for text)
codex exec -m gpt-5.1-codex-max -c model_reasoning_effort="high" -s read-only \
  --skip-git-repo-check --full-auto \
  "As a technical reviewer, analyze:
  1. Mathematical/algorithmic correctness
  2. Code/implementation validity
  3. Experimental design
  4. Statistical analysis
  5. Reproducibility concerns

  Content: [PAPER_CONTENT]" 2>/dev/null

Code Review

# Gemini - Architecture and design (use 3.0 Flash for speed)
# Gemini supports stdin pipe
cat code.py | gemini -m gemini-3-flash-preview "Review this code for:
1. Architecture and design patterns
2. Code organization
3. Naming conventions
4. Documentation quality
5. Maintainability" 2>/dev/null

# Codex - Implementation and security (use 5.2-codex for code)
# Codex: include file content directly in prompt
codex exec -m gpt-5.2-codex -c model_reasoning_effort="high" \
  --skip-git-repo-check --full-auto \
  "Review this code for:
  1. Security vulnerabilities
  2. Performance issues
  3. Bug potential
  4. Edge cases
  5. Best practices compliance

  Code:
  [FILE_CONTENT]" 2>/dev/null

Logic Validation

# Cross-validate logical arguments
gemini -m gemini-3-pro-preview "Validate the logical consistency of this argument:
[ARGUMENT]
Check for: fallacies, assumptions, gaps in reasoning" 2>/dev/null

codex exec -m gpt-5.1-codex-max -c model_reasoning_effort="xhigh" -s read-only \
  --skip-git-repo-check --full-auto \
  "Analyze logical structure:
  [ARGUMENT]
  Identify: premises, conclusions, validity, soundness" 2>/dev/null

Depth Configuration

Quick Review

• Gemini: gemini-3-flash-preview (default), brief prompt
• Codex: gpt-5-mini, model_reasoning_effort="low"
• Use case: Fast sanity check

Standard Review

• Gemini: -m gemini-3-flash-preview
• Codex: gpt-5.2-codex (code) / gpt-5.1-codex-max (text), model_reasoning_effort="medium"
• Use case: Regular code/document review

Thorough Review

• Gemini: -m gemini-3-pro-preview with detailed prompts
• Codex: gpt-5.2-codex (code) / gpt-5.1-codex-max (text), model_reasoning_effort="xhigh"
• Use case: Critical papers, security audits, important decisions

Model Selection Guide

Gemini Models (需要启用 Preview features)

Codex Models

Model Selection Strategy

• Code tasks -> gpt-5.2-codex (stronger code ability)
• Text/Paper tasks -> gpt-5.1-codex-max (stronger text ability)
• Both needed -> Use 5.2-codex for code analysis, 5.1-codex-max for text analysis

Output Synthesis

After running both AIs, Claude should:

1. Compare Outputs:

   • List points where both AIs agree (high confidence)
   • Highlight disagreements (needs human judgment)
   • Note unique insights from each AI

2. Generate Summary:

   ## AI Review Summary
   
   ### Consensus Points (High Confidence)
   - [Points both AIs agree on]
   
   ### Divergent Views (Review Needed)
   - Gemini says: [X]
   - Codex says: [Y]
   - Recommendation: [Claude synthesis]
   
   ### Unique Insights
   - From Gemini: [...]
   - From Codex: [...]
   
   ### Action Items
   1. [Prioritized list of suggested changes]
   

3. Confidence Rating:

   • High: Both AIs agree
   • Medium: Partial agreement
   • Low: Significant disagreement (needs human review)

Error Handling

Gemini Errors

# If Gemini fails, check:
gemini --version  # Verify installation (need v0.21.1+)
gemini -h         # Check available options
/settings         # Make sure Preview features enabled

Codex Errors

# If Codex fails:
codex --version
cat ~/.codex/config.toml  # Check config
# Try without sandbox:
codex exec --yolo -m gpt-5.2-codex "test" 2>/dev/null

Fallback Strategy

If one AI fails:

1. Report the failure to user
2. Continue with available AI
3. Note reduced confidence in results

Best Practices

1. Always use stderr suppression (2>/dev/null) for clean output
2. For papers: Focus Gemini on logic/clarity, Codex (5.1-codex-max) on technical accuracy
3. For code: Focus Gemini on design, Codex (5.2-codex) on implementation
4. Save outputs to temp files for comparison when doing dual review
5. Use high reasoning effort for critical reviews
6. Cross-validate important findings with both AIs
7. Choose right Codex model: code -> 5.2-codex, text -> 5.1-codex-max
8. Codex stdin limitation: Codex does NOT support stdin pipe (|), include content directly in prompt

Example Workflows

Paper Section Review

User: "Review my methodology section using ai-review"

Claude:
1. Asks for review mode (dual-review recommended)
2. Reads the methodology content
3. Runs Gemini 3 Pro for logic/clarity check
4. Runs Codex gpt-5.1-codex-max for technical validation
5. Synthesizes results
6. Presents unified review with confidence levels

Code Security Audit

User: "Security review this file with ai-review"

Claude:
1. Asks for depth (thorough for security)
2. Runs Gemini 3 Flash + Codex gpt-5.2-codex with security-focused prompts
3. Compares vulnerability findings
4. Prioritizes by severity and agreement level
5. Provides remediation suggestions

Resume Capability

Both tools support session resume:

# Gemini
gemini -r latest "Continue with additional review points"

# Codex
codex exec resume --last "Continue the analysis"

Inform user: "You can resume this review session anytime with either tool."