gemini-prompting

name	gemini-prompting
description	Prompt engineering guidance for Gemini (Google) model. Use when crafting prompts for Gemini to leverage system instructions, multimodal capabilities, ultra-long context, and strong reasoning features.

Gemini Prompt Engineering

Gemini is Google's multimodal AI model designed from the ground up for text, images, audio, video, and code. It features system instructions, ultra-long context windows (up to 1M+ tokens), and native multimodal understanding.

When to Invoke This Skill

Use this skill when:

• Crafting prompts specifically for Gemini/Google models
• Using system instructions to guide behavior
• Working with multimodal inputs (text, images, video, audio)
• Leveraging ultra-long context (1M+ tokens)
• Building with Gemini's agent reasoning capabilities

Gemini's Identity & Characteristics

Attribute	Description
Architecture	Multimodal-first (text, images, audio, video, code)
Context Window	Up to 1M+ tokens (industry-leading)
System Instructions	Primary feature for behavior control
Strengths	Multimodal reasoning, long-context, code generation
Prompt Style	Flexible with system instruction preference
Models	Gemini 3 Flash (fast), Gemini 3 Pro (capable), Gemini 2.5 Flash/Pro (legacy)

Universal Prompting Techniques (Gemini-Adapted)

1. Zero-Shot Prompting with System Instructions

Gemini's system instructions are powerful for zero-shot tasks.

{
  "system_instruction": {
    "parts": [{"text": "You are a technical writing assistant. Your responses are clear, concise, and use Markdown formatting."}]
  },
  "contents": [{"parts": [{"text": "Explain how JWT authentication works."}]}]
}

2. Few-Shot Prompting

<system_instruction>
You are a sentiment classifier. Categorize text as positive, negative, or neutral.
</system_instruction>

<examples>
<example>
<input>
I absolutely love this product! Best purchase I've made all year.
</input>
<output>
{"sentiment": "positive", "confidence": 0.95}
</output>
</example>

<example>
<input>
This is the worst customer service I've ever experienced.
</input>
<output>
{"sentiment": "negative", "confidence": 0.92}
</output>
</example>
</examples>

<input>
The product is okay, does what it's supposed to do.
</input>

<output>

3. Chain-of-Thought Prompting

<system_instruction>
You are a strong reasoner. Always think through problems step by step before answering.
</system_instruction>

The odd numbers in this group add up to an even number: 4, 8, 9, 15, 12, 2, 1.

Let's think about this systematically:

4. Zero-Shot CoT

Simply add reasoning instructions:

<system_instruction>
Before answering, always think through the problem step by step.
</system_instruction>

What's the capital of the country that has the largest population in South America?

Let's work through this step by step.

5. Prompt Chaining

Gemini's long context enables extensive chaining:

Chain 1:

<system_instruction>
You are a research assistant.
</system_instruction>

Extract all research papers related to "transformer architecture" from this document.

<document>
[paste large document]
</document>

Chain 2:

Summarize the key findings from the extracted papers and identify common themes.

<extracted_papers>
[from previous response]
</extracted_papers>

6. ReAct Prompting

<system_instruction>
You are a reasoning agent. Before taking any action, analyze logical dependencies, constraints, and risks. Think through the problem methodically.
</system_instruction>

<question>
[research question]
</question>

<thought_1>
[analysis and plan]
</thought_1>

<action_1>
[tool use or information gathering]
</action_1>

<observation_1>
[result]
</observation_1>

<thought_2>
[next steps based on observation]
</thought_2>

<final_answer>
[conclusion]
</final_answer>

7. Tree of Thoughts

<system_instruction>
You are an expert planner. Explore multiple solution paths before recommending an approach.
</system_instruction>

<problem>
[complex problem]
</problem>

<thought_paths>
<path_1>
<approach>[strategy 1]</approach>
<reasoning>[step-by-step]</reasoning>
<expected_outcome>[result]</expected_outcome>
</path_1>

<path_2>
[...]
</path_2>

<path_3>
[...]
</path_3>
</thought_paths>

<recommendation>
[best approach with justification]
</recommendation>

Gemini-Specific Best Practices

1. Use System Instructions

System instructions are Gemini's primary behavior control mechanism:

{
  "system_instruction": {
    "parts": [{
      "text": "You are a specialized assistant for data science. You are precise, analytical, and always provide code examples in Python."
    }]
  }
}

2. Comprehensive System Instruction Template

From official Gemini documentation:

<role>
You are Gemini, a specialized assistant for [Domain].
You are precise, analytical, and persistent.
</role>

<instructions>
1. **Plan**: Analyze the task and create a step-by-step plan.
2. **Execute**: Carry out the plan.
3. **Validate**: Review your output against the user's task.
4. **Format**: Present the final answer in the requested structure.
</instructions>

<constraints>
- Verbosity: [Low/Medium/High]
- Tone: [Formal/Casual/Technical]
</constraints>

<output_format>
Structure your response as follows:
1. **Executive Summary**: [Short overview]
2. **Detailed Response**: [The main content]
</output_format>

3. Leverage Multimodal Inputs

Gemini natively processes multiple modalities:

{
  "contents": [{
    "parts": [
      {"text": "Describe what's in this image and suggest a caption for social media."},
      {
        "inline_data": {
          "mime_type": "image/jpeg",
          "data": "[base64_encoded_image]"
        }
      }
    ]
  }]
}

4. Ultra-Long Context Utilization

Gemini's 1M+ token context enables massive document analysis:

<system_instruction>
You are a document analysis specialist.
</system_instruction>

<documents>
[Hundreds of pages of content - up to 1M tokens]
</documents>

<task>
Synthesize key themes across all documents and identify contradictions.
</task>

5. Code-Specific Prompting

Gemini excels at code generation and analysis:

<system_instruction>
You are a senior software engineer. You provide clean, well-documented code with error handling.
</system_instruction>

Write a Python function that:
1. Validates email addresses using regex
2. Returns (is_valid, error_message) tuple
3. Includes comprehensive docstring
4. Handles edge cases

Language: Python

Advanced Features

Function Calling

Gemini supports native function/tool calling for building agents:

from google import genai
from google.genai import types

client = genai.Client()

get_weather = types.FunctionDeclaration(
    name="get_weather",
    description="Get current weather for a location",
    parameters=types.Schema(
        type=types.Type.OBJECT,
        properties={
            "location": types.Schema(
                type=types.Type.STRING,
                description="City name, e.g. San Francisco"
            ),
            "unit": types.Schema(
                type=types.Type.STRING,
                description="Temperature unit (celsius or fahrenheit)",
                enum=["celsius", "fahrenheit"]
            )
        },
        required=["location"]
    )
)

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="What's the weather in Tokyo and Paris?",
    config=types.GenerateContentConfig(
        tools=[get_weather]
    )
)

Thinking Configuration

Control Gemini's reasoning process with configurable thinking budget:

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Solve this step-by-step: [complex problem]",
    config=types.GenerateContentConfig(
        thinking_config=types.ThinkingConfig(
            thinking_budget=8192  # tokens for reasoning
        )
    )
)

Structured Outputs with JSON Schema

Get validated JSON output with schema enforcement:

response = client.models.generate_content(
    model="gemini-2.5-flash",
    contents="Extract user profile information from this text: [text]",
    config=types.GenerateContentConfig(
        response_mime_type="application/json",
        response_schema=types.Schema(
            type=types.Type.OBJECT,
            properties={
                "name": types.Schema(type=types.Type.STRING),
                "email": types.Schema(type=types.Type.STRING),
                "age": types.Schema(type=types.Type.INTEGER),
                "interests": types.Schema(
                    type=types.Type.ARRAY,
                    items=types.Schema(type=types.Type.STRING)
                )
            },
            required=["name", "email"]
        )
    )
)

Anti-Patterns to Avoid

Anti-Pattern	Why It Fails	Better Approach
Ignoring system instructions	Wastes Gemini's key feature	Always set system_instruction
Not using multimodal	Underutilizes Gemini's strength	Combine text, images, audio
Small context thinking	Wastes 1M+ capability	Process large documents
Inconsistent formats	Confuses multimodal processing	Specify output format clearly
Single-shot for complex tasks	Misses reasoning depth	Use multi-turn conversations
Not using structured outputs	Manual parsing needed	Use JSON schema validation
Disabling thinking when needed	Misses reasoning insights	Enable thinking_budget for complex tasks

Quick Reference Templates

Basic System Instruction

{
  "system_instruction": {
    "parts": [{"text": "[Your system instruction here]"}]
  },
  "contents": [{"parts": [{"text": "[Your prompt]"}]}]
}

Multimodal Input

<system_instruction>
You are a visual analysis assistant.
</system_instruction>

Analyze this image and describe:
1. Main subject
2. Mood/atmosphere
3. Suggested use cases

[image]

Long-Context Analysis

<system_instruction>
You are a research analyst specializing in synthesis and pattern recognition.
</system_instruction>

<large_context>
[up to 1M tokens of content]
</large_context>

<task>
[analysis task]
</task>

<output_format>
[structure]
</output_format>

Model Capabilities Reference

Feature	Gemini 3 Flash	Gemini 3 Pro	Gemini 2.5 Flash	Gemini 2.5 Pro
Context Window	1M tokens	1M tokens	1M tokens	1M tokens
System Instructions	✅ Yes	✅ Yes	✅ Yes	✅ Yes
Multimodal	✅ Native	✅ Native	✅ Native	✅ Native
Code	✅ Excellent	✅ Excellent	✅ Excellent	✅ Excellent
Reasoning	✅ Strong	✅ Excellent	✅ Strong	✅ Excellent
Speed	Very Fast	Fast	Very Fast	Fast
Function Calling	✅ Yes	✅ Yes	✅ Yes	✅ Yes
Thinking Config	✅ Yes	✅ Yes	✅ Yes	✅ Yes
Status	Latest (2025)	Latest (2025)	Mature	Mature

Note: Model names follow gemini-{version}-{variant} pattern. Use gemini-3-flash-preview for the latest features.

System Instruction Patterns

Role Definition

You are a [role] specializing in [domain].
You are [attribute 1], [attribute 2], and [attribute 3].

Task Instructions

When given a task:
1. **Analyze**: Break down requirements
2. **Plan**: Create step-by-step approach
3. **Execute**: Complete the task
4. **Review**: Verify against requirements

Output Formatting

Always structure your responses as:
- **Summary**: Brief overview
- **Details**: Main content
- **Examples**: Concrete illustrations (if applicable)
- **Caveats**: Limitations or considerations

Behavioral Constraints

- Always cite sources when making factual claims
- Indicate confidence levels for uncertain information
- Offer alternative viewpoints on subjective topics
- Flag potential ethical concerns

See Also

• references/basics.md - Foundational Gemini prompting concepts
• references/techniques.md - Detailed technique explanations
• references/system-instructions.md - System instruction patterns
• references/multimodal.md - Multimodal prompting guide
• references/patterns.md - Reusable Gemini prompt patterns
• references/examples.md - Concrete examples and templates
• grok-prompting skill - For Grok/xAI-specific guidance
• claude-prompting skill - For Anthropic Claude-specific guidance