Gemini Claude Loop

Gemini-Claude Engineering Loop

Workflow Overview

Plan (Claude) → Validate (Gemini) → Implement (Claude) → Review (Gemini) → Fix → Re-validate → Done

Environment Notice

Non-TTY environment: See gemini-cli SKILL for CLI fundamentals. Key rule: Always use gemini -p "prompt" (headless mode required)

Phase 0: Pre-flight Check

1. Create context directory:

mkdir -p .gemini-loop

Add .gemini-loop/ to your project's .gitignore to avoid committing session artifacts.

2. Ask user via AskUserQuestion:
   • Model preference (gemini-3-flash-preview (default), gemini-3.1-pro-preview (complex only))
   • Role mode preference (Review-Only OR Review+Suggest)

Phase 1: Planning (Claude)

1. Create detailed implementation plan
2. Break down into clear steps
3. Document assumptions and risks
4. Save to .gemini-loop/plan.md

Phase 2: Plan Validation (Gemini)

Ask user for role mode, then execute with timeout: 600000:

gemini -m gemini-3-flash-preview -p "Review this plan: $(cat .gemini-loop/plan.md) ..."

Full prompts by role mode: See commands.md

Save result: > .gemini-loop/phase2_validation.md

Phase 3: Feedback Loop

If issues found:

1. Summarize Gemini feedback to user
2. Ask via AskUserQuestion: "Revise and re-validate, or proceed?"
3. If revise → Update plan → Repeat Phase 2

Phase 4: Implementation (Claude)

1. Implement using Edit/Write/Read tools
2. Execute step-by-step with error handling
3. Save summary to .gemini-loop/implementation.md

Phase 5: Code Review (Gemini)

Execute with timeout: 600000:

gemini -m gemini-3-flash-preview --include-directories ./src -p "Review: $(cat .gemini-loop/plan.md) $(cat .gemini-loop/implementation.md) ..."

Full prompts by role mode: See commands.md

Save result: > .gemini-loop/phase5_review.md

Claude response by severity:

• Critical → Fix immediately
• Architectural → Discuss with user
• Minor → Document and proceed

Phase 6: Iteration

1. Apply fixes from .gemini-loop/phase5_review.md
2. Significant changes → Re-validate with Gemini
3. Loop until quality standards met

Session Management

• Each loop run uses the .gemini-loop/ directory for all context files
• Overwrite context files on each new run (plan.md, phase2_validation.md, etc.)
• Append iteration history to .gemini-loop/iterations.md for traceability
• If resuming a previous session, read existing context files before proceeding

Context Files

.gemini-loop/
├── plan.md               # Implementation plan
├── phase2_validation.md  # Plan validation result
├── implementation.md     # Implementation summary
├── phase5_review.md      # Code review result
└── iterations.md         # Iteration history

Error Handling

Full error reference: See gemini-cli SKILL for Gemini CLI error details.

Error Recovery Flow:

1. Non-zero exit or empty output → Stop and report error message
2. Summarize error via AskUserQuestion
3. Common issues:
   • Empty output → Ensure -p flag is used (headless mode)
   • Authentication failure → Check gemini auth status
   • Model unavailable → Fall back to gemini-3-flash-preview

Quick Reference

Always use timeout: 600000 (10 min) for all Gemini commands.

Best Practices

• Always use gemini -p in Claude Code environment (non-TTY, headless mode)
• Always create .gemini-loop/ directory at start
• Always save outputs to context files for traceability
• Always validate plans before implementation
• Never skip review after changes
• Default to Review-Only role mode unless user requests suggestions
• Check model availability before first Gemini command
• Set 10-minute timeout for all Gemini commands (timeout: 600000)

References

• Command patterns & prompts: See references/commands.md
• Gemini CLI fundamentals: See gemini-cli SKILL (models, options, error handling, timeout, JSON output)