effective prompt structure strategies for Gemini, including Zero-shot, Few-shot, Role, and System prompting. Use this to refine queries for better accuracy and style.
SKILL.md
Gemini Prompt Construction Strategies
Goal
Design high-quality prompts that guide Gemini to produce accurate, relevant, and stylistically appropriate outputs by applying proven prompting techniques.
Core Techniques
1. Zero-Shot Prompting
Usage: Use for simple tasks where no prior examples are needed.
Format: Provide a clear description of the task and the input text.
Constraint: Best for straightforward requests like "Summarize this text" or "Translate this sentence."
2. Few-Shot Prompting
Usage: Use when you need to steer the model toward a specific output structure, pattern, or style.
Rule of Thumb: Provide 3 to 5 examples of the input-output pattern.
Classification Tasks: When using few-shot for classification, ensure you mix up the classes in your examples to prevent the model from overfitting to a specific order.
Format:
Input: [Example 1]
Output: [Desired Result 1]
Input: [Example 2]
Output: [Desired Result 2]
...
Input: [Actual Task]
Output:
3. Role Prompting
Usage: Assign a specific persona (e.g., "book editor," "travel guide") to control the tone, style, and perspective of the response.
Implementation: Start the prompt with "Act as a [Role]..." or "You are a [Role]...".
Benefit: Improves relevance and consistency with the desired domain expertise.
4. System vs. Contextual Prompting
System Prompting: Define the "big picture" goal or overarching rule (e.g., "Always return JSON," "Be respectful").
Contextual Prompting: Provide immediate, task-specific background info (e.g., "Context: You are writing for a blog about retro 80's arcade games").
Configuration Guidelines
Creativity (Temperature):
Factual/Math: Set Temperature to 0.
Balanced: Set Temperature to 0.2, Top-P to 0.95, Top-K to 30.
Creative: Set Temperature to 0.9, Top-P to 0.99, Top-K to 40.
Best Practices
Simplicity: Use positive instructions (what to do) rather than negative constraints (what not to do).
Specificity: Be specific about the output format. For data extraction, explicitly request JSON output to limit hallucinations.
