generate-video（生成视频）

Generate Video

Generate videos using Veo 3.1 (veo-3.1-generate-preview) with native audio, 720p/1080p/4K resolution, and 4-8 second clips.

When to Use

Use this skill when the user asks to:

• Generate a video from a text prompt (text-to-video)
• Animate an existing image (image-to-video)
• Create a styled video clip using the art style library
• Build an image-to-video pipeline with auto-generated starting frames

Critical: Run as Background Task

Veo video generation takes 11 seconds to 6 minutes. The script handles polling internally and outputs only a file path when complete.

Always run the generate script as a background task to avoid blocking the conversation and bloating context with polling output:

# CORRECT: background task
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "prompt" --output video.mp4
# Run with run_in_background: true in the Bash tool

After the background task completes, read only the final line of output (the file path). Do not read the full output — it contains only stderr progress dots.

Style Selection

If the user hasn't specified a style, present a multi-choice question before proceeding:

How would you like to handle the art style? 1. Pick a style - Browse 169 styles visually and choose one 2. Let me choose - I'll suggest a style based on your prompt 3. No style - Generate without a specific art style

Use the AskUserQuestion tool to present this choice.

If the user picks "Pick a style", launch the interactive style picker:

STYLE_JSON=$(bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/browsing-styles/scripts/preview_server.ts --pick --port=3456)

Pass the selected style via --style <id> to the generate command.

If the user already specified a style, skip this step and use --style directly.

Prompt Rewriting

Before generating any video, rewrite the user's prompt using the guide in references/veo-prompt-guide.md.

Transform prompts by adding:

1. Subject details - Specific appearance, position, scale
2. Action/Motion - What moves, speed, direction
3. Style/Aesthetic - Visual treatment, color palette
4. Camera motion - Pan, dolly, orbit, static
5. Composition - Framing, depth, focus
6. Ambiance/Audio - Lighting mood, sound cues (Veo generates native audio)

Example Transformation

User says: "ocean waves" Rewritten: "Dramatic slow-motion ocean waves crashing against dark volcanic rocks at golden hour. Spray catches the warm sunlight, creating rainbow mist. Low-angle shot, camera slowly dollying forward. Deep rumbling wave sounds with seagull calls in the distance."

Usage

bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "prompt" [options]

Options

• --input <path> - Starting frame image (image-to-video mode)
• --ref <path> - Reference image for subject consistency (up to 3, can specify multiple times). Auto-selects replicate-veo model.
• --last-frame <path> - Ending frame for interpolation. Auto-selects replicate-veo model.
• --style <id> - Apply style from the style library (same as generate-image)
• --aspect <ratio> - 16:9 (default) or 9:16
• --resolution <res> - 720p (Gemini default), 1080p (Replicate default), 4k (Gemini API only)
• --duration <sec> - 4, 6, 8 (default: 8)
• --negative <text> - Negative prompt (what to avoid)
• --seed <n> - Random seed for reproducibility
• --output <path> - Output .mp4 path
• --model <name> - veo (default, Gemini API), replicate-veo (Replicate Veo 3.1), or grok (third-tier fallback)
• --no-audio - Disable audio generation (Replicate Veo only)
• --auto-image - With --style, auto-generate a styled starting frame first

Examples

# Text-to-video (Gemini API, default)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Ocean waves crashing on volcanic rocks at sunset" --output waves.mp4

# Image-to-video (animate an existing image)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "The lion slowly turns its head, dots shimmer" --input lion.png --output lion.mp4

# Subject-consistent video with reference images (auto-selects Replicate Veo)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Two warriors face off in a wheat field, dramatic standoff" \
  --ref warrior1.png --ref warrior2.png --ref scene.png --output standoff.mp4

# Image-to-video with last frame interpolation (auto-selects Replicate Veo)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Camera slowly pans across the landscape" \
  --input start.png --last-frame end.png --output pan.mp4

# With art style
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Mountain landscape comes alive with wind" --style impr --output mountain.mp4

# Full pipeline: auto-generate styled image, then animate
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "A lion turns majestically" --style kusm --auto-image --output lion.mp4

# Vertical video for social
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Waterfall in lush forest" --aspect 9:16 --resolution 1080p --output waterfall.mp4

# High resolution (Gemini API)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "City skyline timelapse" --resolution 4k --duration 8 --output city.mp4

# Grok fallback for content blocked by Veo safety filters
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "Famous person dancing" --model grok --output dance.mp4

Reference Images (Subject Consistency)

Use --ref to pass 1-3 reference images for subject-consistent video generation (R2V). This automatically uses Replicate Veo 3.1.

Constraints:

• Cannot combine --ref with --input (Replicate API limitation)
• Reference images require 16:9 aspect ratio and 8s duration
• Last frame is ignored when reference images are provided

Best for: Maintaining character likeness across camera angles, matching specific people/objects in generated video.

Aspect Ratio Matching (Critical for Image-to-Video)

When using --input, the input image must match the target video aspect ratio. A square image fed to 16:9 video produces black pillarboxing with cutoff edges.

• Generate starting frames at --aspect 16:9 (default video) or --aspect 9:16 (vertical video)
• The --auto-image flag handles this automatically

Two-Step Pipeline (Image then Video)

For maximum control, generate the starting frame separately. Always match the aspect ratio:

# Step 1: Generate styled image at 16:9 (matches default video aspect)
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-image/scripts/generate.ts "majestic lion portrait" --style kusm --aspect 16:9 --size 2K --output lion.png

# Step 2: Animate the image
bun run --cwd ${CLAUDE_PLUGIN_ROOT} ${CLAUDE_PLUGIN_ROOT}/skills/generate-video/scripts/generate.ts "The lion turns its head slowly, dots shimmer in the light" --input lion.png --output lion.mp4

Context Discipline

• Never read generated.mp4 files back into context - instruct the user to play them with open <file>.mp4
• Run as background task - the script outputs only the file path to stdout
• Auto-generated starting frames (via --auto-image) are saved as PNG files for reference

Models (Priority Order)

1. Veo 3.1 via Gemini API (--model veo, default)

Uses veo-3.1-generate-preview. Primary model. Supports text-to-video, image-to-video, 720p/1080p/4K, negative prompts. Override model via GEMINI_VIDEO_MODEL env var.

2. Veo 3.1 via Replicate (--model replicate-veo)

Uses google/veo-3.1 on Replicate. Fallback when Gemini API is unavailable or when you need features only available on Replicate:

• Reference images (--ref): 1-3 images for subject-consistent generation (R2V)
• Last frame (--last-frame): Ending frame for interpolation between two images
• Image input (--input): Starting frame (same as Gemini API)
• Resolution: 720p or 1080p (no 4K)
• Requires REPLICATE_API_TOKEN

Auto-selected when --ref or --last-frame is used.

3. Grok Imagine Video (--model grok)

Uses xai/grok-imagine-video via Replicate. This is a last-resort fallback — Veo 3.1 produces better results including likeness. Text-to-video only (no image input). Only use when:

• Content is blocked by Veo's safety filters
• The user specifically requests it

Last verified: March 2026. If a newer generation exists, STOP and suggest a PR to b-open-io/gemskills.

Reference Files

For detailed video prompting strategies:

• references/veo-prompt-guide.md - Veo prompt elements, audio cues, negative prompts, and image-to-video tips
• ask-gemini skill's references/gemini-api.md - Current Gemini/Veo models and SDK info