Gemini To Nova

Gemini to Nova 2 Lite Migration

Overview

Migrate Python application code and prompts from Google Gemini 2.0/2.5/3.x to Amazon Nova 2 Lite. Transforms SDK calls (google-genai → boto3 Bedrock Runtime) and rewrites prompts to follow Nova 2 Lite formatting and constraints.

Usage

Use this skill when:

• Converting Gemini Python code to call Nova 2 Lite via Bedrock
• Rewriting prompts originally written for Gemini to Nova 2 Lite format
• Migrating function calling / tool use from Gemini to Nova
• Adapting multimodal Gemini code (images, video, documents) to Nova 2 Lite
• Converting Gemini structured output to Nova's approach
• Switching from Gemini reasoning/thinking to Nova reasoning mode

Core Concepts

Key Differences

1. SDK (Python): google-genai → boto3 Bedrock Runtime converse API
2. Prompt format: XML tags / free-form → ##Section Name## delimiters
3. Multimodal system prompt: Full instructions allowed (Gemini) → Persona-only (Nova)
4. Structured output: Native response_mime_type + response_schema (Gemini) → Tool-forcing or inline schema (Nova)
5. Stateful turns: previous_interaction_id (Gemini) → Pass full message history (Nova)
6. Media ordering: Any order (Gemini) → Media MUST precede text (Nova)

What Cannot Be Migrated Directly

• Managed agents / Antigravity → Build with tool use + orchestration
• Deep Research agent → No equivalent
• Image generation → Use Amazon Nova Canvas (separate model)
• Speech/TTS → Use Amazon Polly
• Code execution tool → Implement custom code executor via tool calling
• Interaction persistence (store=true) → Manage state externally

Migration Workflow

You MUST follow these steps in order. After each step, confirm findings with the user before proceeding to the next.

Step 1: Analyze the Gemini Code

First, identify the Gemini SDK and API style:

• Deprecated SDK (import google.generativeai as genai) — uses GenerativeModel, generate_content
• Current SDK (from google import genai) with generateContent — uses client.models.generate_content
• Current SDK with Interactions API — uses client.interactions.create

Then identify which features are used:

• Basic text generation
• System instructions
• Multi-turn conversation (chat or explicit history)
• Function calling / tools
• Structured output (JSON mode / response_schema)
• Multimodal (images, video, documents)
• Streaming
• Reasoning / thinking mode
• Agents

You MUST flag any features in the "cannot migrate" list above and inform the user of alternatives before proceeding.

If the source model is a Gemini "pro" variant (gemini-2.5-pro, gemini-3.1-pro-preview), you MUST ask the user whether they have evaluated Nova 2 Lite for their use case. Pro-tier Gemini models have higher capability ceilings — the user should confirm Nova 2 Lite meets their quality requirements before investing in the migration.

Step 2: Classify the Use Case

Thinking/reasoning support by Gemini model:

Migration rules for thinking:

• If the source model does NOT support thinking → omit additionalModelRequestFields entirely. If the prompt uses CoT ("think step by step"), keep that prompt text as-is.
• If the source model supports thinking but it's NOT enabled in the code → omit additionalModelRequestFields entirely.
• If thinking IS enabled → ask the user which Nova reasoning effort level to use:

Do NOT copy Gemini's budget_tokens value directly — the models have different reasoning efficiency and the values are not interchangeable. Present the three options and let the user decide.

Determine the Nova 2 Lite use case type for correct inference config:

Text/Agentic:

Multimodal:

Step 3: Migrate the Code

You MUST read references/feature-mapping.md for the complete mapping table. You SHOULD read references/code-examples.md for before/after patterns.

If the source code uses generateContent (either google-generativeai or google-genai SDK) or the Interactions API (client.interactions.create), you MUST also read references/generate-content-patterns.md for the specific parameter mappings, Interactions API migration patterns, and deprecated SDK handling.

Ask the user which region-prefixed model ID to use:

Default to us.amazon.nova-2-lite-v1:0 if the user doesn't specify.

SDK transformation (Python):

# Gemini
from google import genai
client = genai.Client()
interaction = client.interactions.create(model="gemini-3.5-flash", ...)

# Nova 2 Lite
import boto3
client = boto3.client("bedrock-runtime")
response = client.converse(modelId="us.amazon.nova-2-lite-v1:0", ...)

Only include additionalModelRequestFields when reasoning is enabled:

# Reasoning DISABLED (default) — omit additionalModelRequestFields entirely
response = client.converse(modelId="us.amazon.nova-2-lite-v1:0", ...)

# Reasoning ENABLED — include reasoningConfig
response = client.converse(
    modelId="us.amazon.nova-2-lite-v1:0",
    ...,
    additionalModelRequestFields={
        "reasoningConfig": {"type": "enabled", "maxReasoningEffort": "medium"}
    },
)

Step 4: Migrate the Prompt

You MUST apply these transformations in order:

1. Replace XML/markdown delimiters with ##Section Name##:

   • <context> → ##Context Information:##
   • <task> → ##Task Summary:##
   • <instructions> → ##Model Instructions:##
   • <examples> → ##Examples##

2. Use canonical Nova section names:

   • ## Task Summary: — defines the task
   • ## Context Information: — background
   • ## Model Instructions: — behavioral rules
   • ## Response style and format requirements: — output format
   • ## Examples — few-shot examples
   • ## Reference — RAG grounding content

3. For multimodal prompts:

   • Move ALL task instructions from system prompt to user prompt
   • Keep system prompt as persona + response style only
   • Ensure media content precedes text in the content array

4. Add suppression guardrail where appropriate:

   DO NOT mention anything inside ##Model Instructions## or ##Examples## in the response.
   

5. For long context (>10K tokens):

   BEGIN INPUT DOCUMENTS
   DOCUMENT 1 START
   {content}
   DOCUMENT 1 END
   END INPUT DOCUMENTS
   
   BEGIN QUESTION
   {query}
   END QUESTION
   
   BEGIN INSTRUCTIONS
   {instructions}
   END INSTRUCTIONS
   

Step 5: Migrate Structured Output

If the Gemini code uses structured output, apply this step. Otherwise skip to Step 6.

Gemini's response_mime_type + response_schema:

• Simple JSON (≤10 keys): Use inline schema in prompt + temperature=0
• Complex JSON (>10 keys): Use tool-forcing with schema in toolSpec.inputSchema

See references/code-examples.md Example 3 for both patterns.

Step 6: Migrate Tool Calling

If the Gemini code uses function calling / tools, apply this step. Otherwise skip to Step 7.

Key differences:

• Gemini function_declarations → Nova toolSpec
• Gemini parameters (OpenAPI) → Nova inputSchema.json (JSON Schema)
• Gemini tool_config.mode (AUTO/ANY/NONE) → Nova toolChoice (auto/any/tool)
• Keep tool descriptions to 20-50 words
• Keep parameter descriptions to ~10 words
• Reference tools by name in system prompt: Use the 'tool_name' tool for X

Step 7: Present the Result

Output format:

## Migrated Code

**Use case:** {type}
**Inference config:** temperature={T}, reasoning={enabled/disabled}
**Breaking changes:** {list any features that couldn't be migrated 1:1}

### Code
{complete migrated code}

### Prompt
**SYSTEM PROMPT:**
{system prompt — persona only for multimodal}

**USER PROMPT:**
{user prompt with ##Section## formatting}

### Implementation Notes
- {inference config rationale}
- {any multimodal ordering requirements}
- {features requiring alternative approach}

Step 8: Validate

You MUST check the migrated code against all criteria below. If any check fails, fix and re-validate before presenting to the user:

• additionalModelRequestFields is omitted when reasoning is disabled, or contains reasoningConfig when enabled
• Inference config matches the use case table
• Multimodal: system prompt contains only persona
• Multimodal: media precedes text in content array
• Prompt uses ##Section## delimiters, not XML
• Tool schemas use JSON Schema format (not OpenAPI)
• No Gemini-specific features remain (previous_interaction_id, store, etc.)

Quick Reference

Common Mistakes

Forgetting the multimodal system prompt restriction

Problem: Copying full Gemini system instructions into Nova multimodal calls — task adherence degrades silently. Fix: Move everything except persona to the user prompt.

Passing additionalModelRequestFields when reasoning is disabled

Problem: Including additionalModelRequestFields={"thinking": ...} or any reasoning config when reasoning is not needed — Nova rejects unknown keys. Fix: Omit additionalModelRequestFields entirely when reasoning is disabled. Only include it with {"reasoningConfig": {"type": "enabled", "maxReasoningEffort": "..."}} when reasoning is needed.

Using XML tags in Nova prompts

Problem: Keeping <context>, <task> etc. from the Gemini prompt — Nova ignores these as structural markers. Fix: Replace with ##Section Name## delimiters.

Wrong media ordering

Problem: Putting text before images/video in the content array. Fix: Media content blocks MUST come before the text block.

Using OpenAPI schema format for tools

Problem: Copying Gemini's OpenAPI-style parameter schemas directly. Fix: Convert to JSON Schema format in inputSchema.json.