video-generation（视频生成）

Video Generation Skill

Generate videos using AI (Google Veo 3.1, OpenAI Sora).

Capabilities:

• 🎬 Text-to-Video: Create videos from text descriptions
• 🖼️ Image-to-Video: Animate images as the first frame
• 🔊 Audio Generation: Dialogue, sound effects, ambient sounds (Veo 3+)
• 🎭 Reference Images: Guide video content with up to 3 reference images (Veo 3.1)

Prerequisites

Default: Vertex AI (10 requests/minute) ⭐

Vertex AI is the default backend with 1400x higher rate limits:

# 1. Set your project
export GOOGLE_CLOUD_PROJECT=your-project-id

# 2. Authenticate (opens browser)
gcloud auth application-default login

# 3. Enable the API (one-time)
gcloud services enable aiplatform.googleapis.com

Add to your .env file:

GOOGLE_CLOUD_PROJECT=your-project-id
GOOGLE_CLOUD_LOCATION=us-central1

Fallback: AI Studio (10 requests/day)

Only use if you don't have a GCP project:

• GOOGLE_API_KEY - Get from https://aistudio.google.com/apikey

For Sora (OpenAI)

• OPENAI_API_KEY - For OpenAI Sora

Available Models

Google Veo Models

Both models include:

• 720p/1080p resolution
• 4, 6, or 8 second duration
• Native audio (dialogue, SFX, ambient)
• Image-to-video (animate images)
• Reference images (up to 3)
• Video extension
• Batch/parallel generation

OpenAI Sora

• Best for: Creative videos, cinematic quality, complex motion
• Resolutions: 480p, 720p, 1080p
• Durations: 5s, 10s, 15s, 20s
• Features: Text-to-video, image-to-video

Workflow

Step 1: Gather Requirements (REQUIRED)

⚠️ Use interactive questioning — ask ONE question at a time.

Question Flow

⚠️ Use the AskUserQuestion tool for each question below. Do not just print questions in your response — use the tool to create interactive prompts with the options shown.

Q1: Image

"I'll generate that video for you! First — do you have an image to animate? - Yes (provide path — I'll use it as the first frame) - No, generate from scratch"

*Wait for response.*

Q2: Audio

"What audio preference? - With audio (default) — Veo 3.1 generates dialogue, SFX, ambient - Silent video — no audio"

*Wait for response.*

Q3: Model

"Which model would you like? - veo-3.1 — Latest, highest quality with audio (default) - veo-3.1-fast — Faster processing with audio - veo-3 / veo-3-fast — Previous generation with audio - sora — OpenAI, up to 20 seconds, no audio"

*Wait for response.*

Q4: Duration

"What duration? - 4 seconds - 6 seconds - 8 seconds (default)"

*Wait for response.*

Q5: Format

"What aspect ratio and resolution? - 16:9 landscape, 720p - 16:9 landscape, 1080p - 9:16 portrait, 720p - 9:16 portrait, 1080p - Or specify"

*Wait for response.*

Quick Reference

Step 2: Craft the Prompt

Transform the user request into an effective video prompt:

1. Describe the scene: Set the visual context
2. Specify action: What moves, changes, happens
3. Include camera work: "slow pan", "tracking shot", "dolly shot"
4. Add audio cues (Veo 3+): Use quotes for dialogue, describe sounds
5. Set the mood: Lighting, atmosphere, time of day

Example with dialogue (Veo 3.1):

• User: "a person discovering treasure"
• Enhanced: "Close-up of a treasure hunter's face as torchlight flickers. He murmurs 'This must be it...' while brushing dust off an ancient chest. Sound of creaking hinges as he opens it, revealing golden light on his awestruck face. Cinematic, dramatic shadows."

Example without dialogue:

• User: "a dog running on a beach"
• Enhanced: "Cinematic slow-motion shot of a golden retriever running joyfully along a beach at sunset, waves lapping, warm golden hour lighting, shallow depth of field"

Step 3: Select the Model

Default: veo-3.1 (highest quality, with audio)

Step 4: Generate the Video

Execute the appropriate script from ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/:

For Google Veo 3.1 (default, with audio):

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --prompt "your enhanced prompt with 'dialogue in quotes'" \
  --model "veo-3.1" \
  --duration 8 \
  --aspect-ratio "16:9" \
  --resolution "720p"

For Google Veo 3.1 with image input:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --prompt "The cat slowly opens its eyes and yawns" \
  --image "/path/to/cat.jpg" \
  --model "veo-3.1" \
  --duration 8

For faster generation:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --prompt "your prompt" \
  --model "veo-3.1-fast"

For OpenAI Sora (longer videos):

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/sora.py \
  --prompt "your enhanced prompt" \
  --duration 20 \
  --resolution "1080p"

List available models:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py --list-models

Video Extension (For Long-Form Continuity)

The --extend flag creates TRUE visual continuity by continuing from where a previous Veo video ended. This is the best approach for long-form videos.

Basic extension:

# First, generate initial clip
python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --prompt "A person walks through a forest at sunrise" \
  --duration 8

# Extend it with new content (adds ~7 seconds)
python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --extend veo_veo-3.1_20260104_120000.mp4 \
  --prompt "Continue walking, discover a hidden stream"

Multiple extensions (for longer videos):

# Extend 5 times (adds ~35 seconds of continuation)
python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --extend initial_clip.mp4 \
  --prompt "Keep exploring the forest, encounter wildlife" \
  --extend-times 5

Extension vs Stitching:

Extension Limits:

• Input video must be Veo-generated (max 141 seconds)
• Each extension adds ~7 seconds
• Maximum 20 extensions total (~2.5 minutes)
• Output resolution is 720p

Batch Generation (Parallel)

Generate multiple videos simultaneously for faster multi-scene workflows. Instead of waiting 15+ minutes for 5 sequential videos, generate them all in parallel (~3 minutes total).

Create a scenes.json file:

[
  {"prompt": "Scene 1: Cinematic hero shot of wireless earbuds on dark surface", "duration": 6, "output": "scene1_hero.mp4"},
  {"prompt": "Scene 2: Sound waves visualization, person enjoying music", "duration": 8, "output": "scene2_sound.mp4"},
  {"prompt": "Scene 3: Close-up of earbud in ear, person exercising", "duration": 8, "output": "scene3_comfort.mp4"},
  {"prompt": "Scene 4: Lifestyle montage, various settings", "duration": 8, "output": "scene4_lifestyle.mp4"},
  {"prompt": "Scene 5: Product with logo on clean background", "duration": 4, "output": "scene5_cta.mp4"}
]

Generate all scenes in parallel:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --batch scenes.json

With custom worker count:

python3 ${CLAUDE_PLUGIN_ROOT}/skills/video-generation/scripts/veo.py \
  --batch scenes.json \
  --max-workers 3

Batch config options per video:

Speed comparison:

Step 5: Deliver the Result

1. Provide the generated video file/URL
2. Share the enhanced prompt used
3. Mention generation settings (duration, resolution)
4. Offer to:

- Generate variations - Try different style/duration - Use a different API - Extend the video

Error Handling

Missing API key: Inform the user which key is needed:

• OpenAI: https://platform.openai.com/api-keys
• Google: https://aistudio.google.com/apikey

Content policy violation: Rephrase the prompt appropriately.

Generation failed: Retry with simplified prompt or different API.

Quota exceeded: Suggest waiting or trying the other provider.

Prompt Engineering Tips

For Audio (Veo 3.1)

• Dialogue: Use quotes for speech: "Hello!" she said excitedly
• Sound effects: Describe explicitly: tires screeching, engine roaring
• Ambient: Describe the soundscape: birds chirping, distant traffic
• Example: A man whispers "Did you hear that?" as footsteps echo in the dark hallway

For Cinematic Quality

• Include camera directions: "slow dolly", "tracking shot", "crane shot"
• Specify lighting: "golden hour", "dramatic shadows", "soft diffused light"
• Add film references: "Blade Runner style", "Wes Anderson aesthetic"

For Realistic Motion

• Describe physics: "natural movement", "realistic physics"
• Include environmental details: "wind in hair", "leaves rustling"
• Specify speed: "slow motion", "real-time", "time-lapse"

For Image-to-Video

• Describe what should change/move from the starting image
• Be specific about the action: "the cat slowly opens its eyes"
• Include environmental motion: "leaves blow past"

Negative Prompts

• Describe what NOT to include: --negative-prompt "cartoon, low quality, blurry"
• Don't use "no" or "don't" - just describe the unwanted elements

API Comparison