veo-generator - SKILL.md Agent Skill

name: veo-generator description: Generate videos using Google Veo 3.1 on Vertex AI. Use when the user says "生成视频", "做个视频", "generate video", "帮我生成一段视频", "create a video", "视频生成", "text to video", "文生视频", or when you need to create video content.

Veo 3.1 Video Generator

Generate videos from text prompts using Veo 3.1 on Vertex AI, save to CC Pages, and send to Discord.

Usage

Call the generation script directly:

~/.claude/skills/veo-generator/scripts/veo-generate.sh "your prompt here"

Options

# Basic generation (1 video, 16:9, 8s, standard model, 720p)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "a cat playing with a laser pointer"

# Portrait video for mobile
~/.claude/skills/veo-generator/scripts/veo-generate.sh "smartphone app demo" --aspect 9:16

# Shorter clip
~/.claude/skills/veo-generator/scripts/veo-generate.sh "ocean waves at sunset" --duration 6

# Faster model (lower quality but quicker)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "cinematic drone shot over mountains" --model fast

# 1080p resolution
~/.claude/skills/veo-generator/scripts/veo-generate.sh "product showcase" --resolution 1080p

# Multiple videos (1-4)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "abstract art animation" --count 2

# Negative prompt (avoid certain content)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "peaceful park scene" --negative "people, crowds, text"

# Disable prompt rewriter (use exact prompt)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "minimal geometric shapes" --no-rewrite

# Custom output filename
~/.claude/skills/veo-generator/scripts/veo-generate.sh "logo animation" --output brand-intro

# Longer timeout for complex videos
~/.claude/skills/veo-generator/scripts/veo-generate.sh "detailed cityscape timelapse" --timeout 600

# Combine options
~/.claude/skills/veo-generator/scripts/veo-generate.sh "B200 GPU rack with blinking LEDs, cinematic" --aspect 16:9 --model standard --resolution 1080p --duration 8

Image-to-Video

Use --image to pass an input image as the first frame:

# Image-to-video: animate an existing image
~/.claude/skills/veo-generator/scripts/veo-generate.sh "the robot starts walking forward" --image /path/to/robot.png

# With Imagen pipeline: generate image first, then animate
IMG_URL=$(~/.claude/skills/imagen-generator/scripts/imagen-generate.sh "a futuristic robot standing in a hangar, cinematic" --aspect 16:9)
IMG_PATH="${CC_PAGES_WEB_ROOT}/assets/imagen/$(basename "$IMG_URL")"
VID_URL=$(~/.claude/skills/veo-generator/scripts/veo-generate.sh "the robot powers up and walks forward" --image "$IMG_PATH")

# Reference images for visual guidance (up to 3 asset images)
~/.claude/skills/veo-generator/scripts/veo-generate.sh "a product showcase video" --ref product-front.png --ref product-side.png

Output

Videos saved to $CC_PAGES_WEB_ROOT/assets/veo/ as MP4
Returns the public URL: $CC_PAGES_URL_PREFIX/assets/veo/{filename}.mp4
When called from Discord context, send the URL using send-to-discord.sh --plain

Workflow for Discord

# 1. Generate video
URL=$(~/.claude/skills/veo-generator/scripts/veo-generate.sh "your prompt" --aspect 16:9)

# 2. Send to Discord
~/.claude/scripts/send-to-discord.sh --plain "$URL"

Important: Async Operation

Unlike Imagen (synchronous), Veo uses long-running operations:

Submit request → get operation ID
Poll fetchPredictOperation every N seconds
When done: true, extract video bytes or GCS URI

The script handles all of this automatically. Default poll interval is 10s, timeout is 300s (5 min). For complex prompts, use --timeout 600.

Progress output goes to stderr, so URL=$(...veo-generate.sh...) captures only the final URL.

Models

veo-3.1-generate-001 (standard, default) — Best quality, GA
veo-3.1-fast-generate-001 (fast) — Faster generation, good quality, GA
veo-3.1-lite-generate-001 (lite) — Lightest, supports audio generation, Preview

Older Models (available but not default)

veo-3.0-generate-001 / veo-3.0-fast-generate-001 — Veo 3 GA
veo-2.0-generate-001 — Veo 2 (no 1080p support)

Parameters

Aspect Ratio: 16:9 (landscape, default), 9:16 (portrait)
Duration: 4, 6, or 8 seconds (default: 8)
Resolution: 720p (default), 1080p (Veo 3+ only)
Sample Count: 1-4 videos per request
Negative Prompt: describe content to avoid
Enhance Prompt: Gemini-powered prompt rewriting (on by default)
Person Generation: allow_adult (default), disallow
Image Input (--image): Local image path, used as first frame for image-to-video
Reference Images (--ref): Up to 3 asset reference images for visual guidance

Prompt Tips

Be specific: "A golden retriever running through autumn leaves in slow motion, cinematic lighting" > "dog running"
Specify camera movement: "dolly zoom", "tracking shot", "aerial drone shot", "timelapse"
Specify style: "cinematic", "documentary", "anime", "watercolor", "photorealistic"
Supports Chinese prompts natively
Use --negative to exclude unwanted elements

Prerequisites

google-genai Python SDK (pip install google-genai)
Application Default Credentials configured (gcloud auth application-default login)
Vertex AI API enabled on the GCP project
CC Pages (GCS-backed via $CC_PAGES_WEB_ROOT)

Files

~/.claude/skills/veo-generator/scripts/
├── veo-generate.sh                  # Entry point (exec wrapper)
└── veo-generate.py                  # Core logic (google-genai SDK)

$CC_PAGES_WEB_ROOT/assets/veo/      # Generated videos (web-accessible)