DeepMind Workflows & Best Practices

📖 3 min read deepmindgeminiworkflowsbest-practicesprompt-engineering

Gemini-specific prompt engineering, multimodal prompting strategies, cost optimization (tier selection: Flash vs Pro vs Ultra), streaming patterns, tool use, and production deployment patterns.

Key Takeaways

Tier selection: Flash for cost/speed, Pro for most workloads, Ultra for maximum quality
Multimodal prompting: Gemini excels at cross-modal tasks — mix text, image, audio, and video in a single prompt
Streaming: available across all tiers. Use for interactive apps and better UX
Google-native: deep integration with Cloud Storage, BigQuery, and Workspace for data pipelines

Prompt Engineering for Gemini

System Instructions

model = genai.GenerativeModel(
    "gemini-3.5-pro",
    system_instruction="""You are a senior data scientist at Google.
Respond with:
1. Clear explanation in 2-3 sentences
2. Python code example when relevant
3. Common pitfalls to avoid"""
)

response = model.generate_content("How do I handle class imbalance in ML?")

Multimodal Prompting

Gemini’s key strength is native multimodal processing:

# Text + image + audio in one prompt
response = model.generate_content([
    "Analyze this product image, describe its features, and suggest an audio jingle:",
    product_image,
    "Here's a competitor's ad audio for reference:",
    competitor_audio
])

Prompt Structure for Best Results

Technique	Example
Role assignment	”You are a senior architect reviewing this design”
Output format	”Respond as a JSON object with keys: summary, risks, recommendations”
Step-by-step	”First analyze the problem, then propose 3 solutions, then compare them”
Multimodal context	Include relevant images, charts, or audio alongside text

Cost Optimization — Tier Selection

Tier	When to Use	Cost Factor
Gemini 3.5 Flash	Simple Q&A, classification, routing, high-throughput	Lowest
Gemini 3.5 Pro	Most production workloads, coding, analysis	Standard
Gemini 3.5 Ultra	Complex reasoning, R&D, maximum quality	Premium

# Route based on task complexity
def get_model(task_type):
    if task_type == "classification":
        return "gemini-3.5-flash"    # Fast + cheap
    elif task_type == "analysis":
        return "gemini-3.5-pro"      # Balanced
    elif task_type == "research":
        return "gemini-3.5-ultra"    # Maximum quality

AI Studio Free Tier

No credit card required
Generous free quota for experimentation
1M token context included
All Gemini models accessible
Use for prototyping and evaluation before scaling via API

Streaming Best Practices

# Stream for better UX
response = model.generate_content(
    "Explain the transformer architecture in detail",
    stream=True
)
for chunk in response:
    if chunk.text:
        print(chunk.text, end="", flush=True)

# Track usage
print(f"\n\nInput tokens: {response.usage_metadata.prompt_token_count}")
print(f"Output tokens: {response.usage_metadata.candidates_token_count}")

Tool Use Patterns

Function Calling

model = genai.GenerativeModel(
    "gemini-3.5-pro",
    tools=[{
        "function_declarations": [{
            "name": "search_knowledge_base",
            "description": "Search internal documentation",
            "parameters": {
                "type": "object",
                "properties": {
                    "query": {"type": "string"},
                    "max_results": {"type": "integer", "default": 5}
                }
            }
        }]
    }]
)

Code Execution

model = genai.GenerativeModel(
    "gemini-3.5-pro",
    tools="code_execution"
)
response = model.generate_content(
    "Write and run Python code to calculate the first 20 Fibonacci numbers"
)

Google Ecosystem Integration

Cloud Storage

# Process files directly from GCS
model = genai.GenerativeModel("gemini-3.5-pro")
gcs_file = genai.upload_file("gs://my-bucket/document.pdf")
response = model.generate_content(["Summarize this document:", gcs_file])

BigQuery

# Gemini for BigQuery — natural language SQL
# Available directly in BigQuery console
# "Show me monthly revenue by product for Q1 2026, with growth rates"

Workspace Integration

Gemini in Google Workspace provides AI assistance across Docs, Sheets, Gmail, and Slides. Enterprise-only, managed via Google Workspace admin console.

Production Deployment

# Vertex AI for production
from vertexai.generative_models import GenerativeModel

model = GenerativeModel("gemini-3.5-pro")
# Deploy with:
# - Auto-scaling based on QPS
# - Monitoring via Cloud Monitoring
# - Logging via Cloud Logging
# - IAM for access control
# - VPC for network security

Where Next

For broader prompt engineering techniques, see the Prompt Engineering Deep Dive.

For Claude-specific and GPT-specific workflows, see Claude Workflows and OpenAI Workflows.