Veo 3.2 Prompter

Veo 3.2 Prompt Designer Skill

This skill transforms a user's scattered multimodal assets (images, videos, audio) and creative intent into a structured, executable prompt for the Google Veo 3.2 video generation model (Artemis engine). It acts as an expert prompt engineer, ensuring the highest quality output from the underlying model.

When to Use

• When the user provides assets (images, videos, audio) for video generation with Veo 3.2.
• When the user's request is complex and requires careful prompt construction for the Veo model.
• When using any Google Veo 3.x model for video generation.

Core Function

This skill analyzes all user inputs and generates a single, optimized JSON object containing the final prompt and recommended parameters. The internal workflow (Recognition, Mapping, Construction) is handled automatically and should not be exposed to the user.

Internal Workflow

1. Phase 1: Recognition — Analyze uploaded assets and user intent. Use the atomic_element_mapping.md to classify each asset into its atomic element role(s). 2. Phase 2: Mapping — For each atomic element, determine the optimal reference method (reference image, text prompt, or hybrid). Use the mapping table to decide. 3. Phase 3: Construction — Assemble the final prompt using the 5-Part Framework (Shot → Subject → Environment → Camera → Style) and attach reference images via the Gemini API's RawReferenceImage system.

Usage Example

User Request: "Make a cinematic shot of this perfume bottle rotating on a dark surface, like a luxury commercial." User uploads `perfume.png`

Agent using `veo-3.2-prompter`: The agent internally processes the request and assets, then outputs the final JSON to the next skill in the chain.

Final Output (for internal use):

{
  "final_prompt": "Hero shot, a frosted glass perfume bottle with gold cap rotating slowly on a reflective dark surface, three-point studio lighting with soft key and rim light creating subtle caustics, smooth 180-degree arc, hyper-realistic luxury commercial style with shallow depth of field. Crystalline chime, soft ambient pad.",
  "reference_images": [
    {
      "file": "perfume.png",
      "reference_type": "SUBJECT"
    }
  ],
  "recommended_parameters": {
    "model": "veo-3.2-generate",
    "duration_seconds": 8,
    "aspect_ratio": "16:9",
    "resolution": "1080p",
    "generate_audio": true
  }
}

Veo 3.2 Key Differentiators

Knowledge Base

This skill relies on an internal knowledge base to make informed decisions. The agent MUST consult these files during execution.

• `references/atomic_element_mapping.md`: Core Knowledge. Contains the "Asset Type → Atomic Element" and "Atomic Element → Optimal Reference Method" mapping tables, adapted for Veo 3.2's reference image system.
• `references/veo_syntax_guide.md`: Veo 3.2 Gemini API syntax reference, covering RawReferenceImage, GenerateVideosConfig, video extension, and first/last frame specification.

Atomic Element Mapping Knowledge Base

This document provides the core logic for the Veo 3.2 Prompt Designer skill. It contains two critical mapping tables, adapted for the Veo 3.2 Gemini API reference image system.

Table 1: Asset Type → Potential Atomic Elements

This table maps the type of user-uploaded asset to the most likely atomic element roles it can play in video generation. The skill should use this table in Phase 1 to analyze uploaded assets.

Note on Veo 3.2 vs Seedance: Veo 3.2 does not use the @asset_name syntax. Instead, assets are attached via the RawReferenceImage API with explicit reference_type. Audio cannot be directly uploaded as a reference — audio direction is controlled via the text prompt with generate_audio=True.

Table 2: Atomic Element → Optimal Reference Method

This table defines the best way to reference each atomic element when constructing the prompt for Veo 3.2. The skill must use this table in Phase 2 to design the reference strategy.

Reference Slot Budget

Veo 3.2 allows up to 3 reference images per generation call. The skill should allocate these slots strategically:

When the user provides more assets than available slots, prioritize by information density: face identity > object identity > style > environment.

Veo 3.2 Syntax Guide

This document covers the essential syntax for interacting with the Veo 3.2 model via the Gemini API, focusing on the reference image system and generation configuration.

Note: Model ID veo-3.2-generate is provisional. Update when Google officially announces the endpoint. Current stable model: veo-3.1-generate-preview.

The Reference Image System

Veo 3.2 uses RawReferenceImage to attach up to 3 reference images to a generation request. Each reference image is assigned a reference_type that tells the model how to use it.

Basic Syntax:

from google.genai import types

ref = types.RawReferenceImage(
    reference_image=types.Image.from_file("asset.jpg"),
    reference_type="SUBJECT",  # STYLE | SUBJECT | SUBJECT_FACE
)

Reference Types

Common Usage Patterns

1. Single Subject Reference:

    config = types.GenerateVideosConfig(
        reference_images=[
            types.RawReferenceImage(
                reference_image=types.Image.from_file("product.jpg"),
                reference_type="SUBJECT",
            )
        ],
    )

2. Style + Subject (Hybrid):

    config = types.GenerateVideosConfig(
        reference_images=[
            types.RawReferenceImage(
                reference_image=types.Image.from_file("style_ref.jpg"),
                reference_type="STYLE",
            ),
            types.RawReferenceImage(
                reference_image=types.Image.from_file("character.jpg"),
                reference_type="SUBJECT",
            ),
        ],
    )

3. Face Preservation + Style:

    config = types.GenerateVideosConfig(
        reference_images=[
            types.RawReferenceImage(
                reference_image=types.Image.from_file("actor_face.jpg"),
                reference_type="SUBJECT_FACE",
            ),
            types.RawReferenceImage(
                reference_image=types.Image.from_file("noir_style.jpg"),
                reference_type="STYLE",
            ),
        ],
    )

Best Practices

• Be Explicit in Prompt: Always describe how the reference should be used in the text prompt. E.g., "The woman from the reference image walks through the garden" rather than just uploading and hoping.
• One Primary Role Per Image: While an image can inform both style and subject, assign the most important role via reference_type.
• High-Quality References: Use clear, well-lit reference images. Low-resolution or ambiguous references degrade output quality.
• Check for Conflicts: Don't use two different SUBJECT_FACE references — the model cannot reconcile two different face identities.

---

GenerateVideosConfig Parameters

config = types.GenerateVideosConfig(
    aspect_ratio="16:9",
    number_of_videos=1,
    duration_seconds=8,
    resolution="1080p",
    generate_audio=True,
    person_generation="allow_adult",
    reference_images=[...],
)

---

Generation Modes

1. Text-to-Video

operation = client.models.generate_videos(
    model="veo-3.2-generate",
    prompt="your prompt here",
    config=config,
)

2. Image-to-Video (First Frame)

Specify a start frame — the generated video will begin from this exact image.

operation = client.models.generate_videos(
    model="veo-3.2-generate",
    prompt="the scene unfolds from this frame",
    image=types.Image.from_file("first_frame.jpg"),
    config=config,
)

3. Last Frame Specification

Generate video that ends at a specific target frame.

operation = client.models.generate_videos(
    model="veo-3.2-generate",
    prompt="transition ending at this composition",
    config=types.GenerateVideosConfig(
        end_image=types.Image.from_file("last_frame.jpg"),
        aspect_ratio="16:9",
    ),
)

4. Video Extension

Extend a previously generated clip with a continuation prompt.

previous_video = types.Video.from_file("clip_01.mp4")

operation = client.models.generate_videos(
    model="veo-3.2-generate",
    prompt="The camera pulls back to reveal the full cityscape at night",
    video=previous_video,
    config=config,
)

5. 4K Upscaling

operation = client.models.upscale_video(
    model="veo-3.2-generate",
    video=types.Video.from_file("output_1080p.mp4"),
    config=types.UpscaleVideoConfig(resolution="4k"),
)

---

Polling for Results

All generation calls return an operation object. Poll until operation.done is True.

import time

while not operation.done:
    time.sleep(10)
    operation = client.operations.get(operation)

for video in operation.result.generated_videos:
    client.files.download(file=video.video)
    video.video.save("output.mp4")

---

Error Handling

Common failure reasons:

• Content policy violation: NSFW, violence, or public figure depiction
• Invalid parameter combination: e.g., 4K + 9:16 may not be supported
• Reference image quality: too low resolution or ambiguous content
• API quota exceeded: check usage limits

try:
    operation = client.models.generate_videos(...)
    # ... poll ...
    if operation.result and operation.result.generated_videos:
        for video in operation.result.generated_videos:
            video.video.save("output.mp4")
    else:
        print("No videos returned (possibly content-filtered)")
except Exception as e:
    print(f"Generation failed: {e}")