Gemini 3.1 Pro
Google · gemini family · Official docs
Gemini 3.1 Pro is the current Gemini Pro target in Refrase. It should use the Gemini family adaptation rules: dedicated systemInstruction, explicit task framing, measurable constraints, and native structured output settings when available. Refrase does not claim measured lift for Gemini 3.1 Pro yet.
Specifications
Strengths
extractionanalysisgenerationlong-context-synthesis
Key capabilities
- ✓Current Gemini Pro-family model for reasoning, long-context work, and multimodal prompts (source: Google AI model docs)
- ✓1M token context window and 65,536 max output tokens (source: Google AI model docs)
- ✓Supports text, image, audio, video, and PDF inputs with text output (source: Google AI model docs)
How to prompt Gemini 3.1 Pro
Preferred instruction format
Gemini uses a dedicated systemInstruction field in API requests rather than embedding durable instructions only in user content.
Recommended practices
- Put durable role, behavior, and constraint instructions in systemInstruction.
- Use consistent XML-style tags or Markdown headings for task, context, constraints, examples, and output format.
- For long-context prompts, put critical task instructions before the context and the final question after the context.
- Use native structured output settings when available instead of relying only on natural-language JSON instructions.
Anti-patterns to avoid
- Do not mix inconsistent delimiter styles across examples.
- Do not rely on vague subjective constraints like 'brief' or 'detailed'; use measurable limits.
- Do not bury essential constraints in the middle of very long context.