Instruction file imported from amo-tech-ai/rocket-path-ai (.cursor/rules/gemeni/gemini-snippet.mdc). Copyright stays with the author.

Gemini Prompting Rules - Quick Reference

🤖 Gemini AI Prompting Rules

Critical Rules:

1. ALWAYS use few-shot examples (2-5 examples) - Zero-shot prompts are significantly less effective
2. ALWAYS specify output format - Use prefixes like "JSON:", "Answer:", or explicit format instructions
3. ALWAYS provide context - Don't assume the model knows domain-specific information
4. ALWAYS use consistent formatting - Examples must have identical structure

Prompt Structure:

[Context/Background]
[Clear Instructions]
[Constraints]
[2-5 Few-Shot Examples with consistent formatting]
[Input Prefix] [User Input]
[Output Prefix]

Model Parameters:

• Temperature: 0.2-0.4 for structured outputs, 0.5-0.7 for creative content
• Max Tokens: Set based on expected response length (100 tokens ≈ 60-80 words)
• Top-P: Default 0.95, lower for more focused responses

Common Patterns:

Structured Output:

You are [role]. Based on these examples, [task]:

Example 1:
Input: [input]
Output: [output]

Example 2:
Input: [input]
Output: [output]

Now process: [user input]
Output format: JSON with fields [fields]

Creative Content:

[Context about style/tone]
[2-3 examples showing desired style]
[Constraints]
Generate: [user request]
Temperature: 0.6

Avoid:

• ❌ Zero-shot prompts without examples
• ❌ Vague instructions without constraints
• ❌ Inconsistent formatting in examples
• ❌ Anti-patterns (showing what NOT to do)
• ❌ Assuming model has domain knowledge

Iteration Strategy:

1. If results don't match: Rephrase prompt
2. If still failing: Switch to analogous task
3. If format issues: Change content order
4. If fallback responses: Increase temperature

Full guide: See gemini-prompting.mdc for comprehensive documentation.