talent-factory/gemini-analyst

Gemini Bridge: Usage Patterns

When to Delegate to Gemini

Token Budget Rule

Estimate token count before analysis:

• < 150K tokens → Claude handles directly

• 150K tokens → Delegate to Gemini

Quick estimation: ~750 tokens per ~600 words, ~500 tokens per ~400 lines of code.

Vision Rule

For image analysis tasks where Gemini's vision capabilities or extended context provide an advantage, prefer gemini_analyze_image. Claude Code can also read images natively via the Read tool, so use Gemini when the analysis requires Gemini-specific strengths (e.g., large diagrams, complex OCR, PDF parsing).

Validation Rule

For architecture decisions, security reviews, or critical design choices: Use gemini_compare_approaches to get an independent assessment. This implements the "propose with Claude, validate with Gemini" pattern.

Tool Reference

# Check bridge is working (do this once per session)
gemini_status()

# Short text prompts / second opinions
gemini_analyze_text(prompt, context=None, temperature=0.2)

# Large codebase analysis (up to 1M tokens with default model)
gemini_analyze_codebase(code_content, task, language=None)

# Image/screenshot/PDF analysis
gemini_analyze_image(image_path, question)

# Architecture/implementation comparison
gemini_compare_approaches(problem, approach_a, approach_b, criteria=None)

Output Attribution Pattern

Always label Gemini's output clearly:

**Analysis by Gemini 2.5 Pro**
[Gemini's response]

This is important for:

1. Transparency in multi-model workflows
2. Audit trails in educational settings
3. Debugging when models disagree

Model-Agnostic Design Principle

The bridge is designed so that the routing logic lives in agent configuration, not in application code. To swap Gemini for a different model:

1. Set the GEMINI_MODEL environment variable
2. All agents/commands continue working unchanged

This demonstrates to students: good architecture is model-agnostic.