Generate Clip — AI Video Generation

You are helping the user generate custom AI video clips for their Remotion project. This creates unique footage that doesn't exist anywhere — going beyond stock footage.

Workflow

1. Understand the Need

Ask the user:

• What should the clip show? — A scene, product, animation, nature, abstract?
• Duration — How long? (typically 3-10 seconds for AI clips)
• Style — Cinematic, drone shot, close-up, slow motion, time-lapse?
• Source — Text-to-video (from description) or image-to-video (animate a still)?

2. Craft the Video Prompt

Write a scene description, following Replicate's official prompt-videos skill for technique (it's model-agnostic, so it applies to Veo too). Cover the camera movement, the subject and its action, the setting, lighting and mood, and the speed.

Good prompt examples:

• "Slow overhead crane shot of a modern city at golden hour, cars moving on the streets below, warm cinematic lighting."
• "Close-up of coffee being poured into a white mug, steam rising, soft morning light, shallow depth of field."
• "Abstract fluid art, flowing metallic gold and deep blue paint, mesmerizing slow motion."
• "Drone flying through a forest canopy, sunlight filtering through the leaves, cinematic nature footage."

For a video, also:

• One continuous shot per clip; Remotion does the cutting.
• No burned-in text or subtitles, and negative space where titles go.
• Keep scenes consistent: paste the same style block into every clip's prompt and keep one model for the whole video. See skills/remotion-production/rules/video-generation.md.

3. Generate via remotion-media (default)

Use remotion-media generate_video_from_text:
- prompt: [crafted prompt]
- output_name: clip-city-aerial
- model: veo3_fast      # default; veo3 for finals
- aspect_ratio: 16:9    # or 9:16 for shorts

To animate a still instead (image-to-video):

Use remotion-media generate_video_from_image:
- prompt: [how the image should move]
- image_urls: [1 image URL to animate, or 2 to transition from first to last frame]
- output_name: clip-animated-still

• The clip is saved to public/<output_name>.mp4 in the directory your agent was launched from, so start your agent (Claude Code, Codex, Gemini CLI…) from the Remotion project root.
• image_urls must be public URLs KIE can fetch, not files in public/. Reuse the KIE-hosted URL that generate_image returns (deleted after 14 days) or a Replicate output URL (valid 1 hour).
• KIE's Veo clips run ~8 s at 720p, 16:9 or 9:16. Plan to trim.
• If a KIE video call fails, fall back to Replicate (step 4). KIE's Veo depends on a preview Google is shutting down, so errors or empty results may be permanent; don't retry in a loop.

4. Generate via Replicate (specific models, or when KIE fails)

It needs the local replicate-mcp server with REPLICATE_API_TOKEN. Follow Replicate's official find-models and run-models skills:

1. Search, then read the schema. search_models with body: "image to video" (or "text to video"), then get_models (model_owner, model_name) and list_models_examples. Durations, resolutions and image inputs differ per model, so use only inputs that exist in the schema.
2. Create the prediction. Official models: create_models_predictions. Community models: create_predictions with version: "owner/name:<version_id>". Don't use Prefer: "wait" for video.

# Image-to-video with Kling v3 (check the schema first)
create_models_predictions:
  model_owner: "kwaivgi"
  model_name: "kling-v3-video"
  input:
    start_image: "[public image URL]"
    prompt: "[motion description]"
    # duration, aspect ratio, mode: names and allowed values from the schema

3. Poll get_predictions with prediction_id every 10–20 s until status is succeeded, failed or canceled. Video takes minutes.
4. Download the URL in output right away (it expires after 1 hour):

mkdir -p public/footage
curl -L -o public/footage/clip-[descriptive-name].mp4 "<output URL>"

Starting points, as of Oct 2026 (search and read the schema before calling):

• Quick default: Veo 3.1 via remotion-media
• Image-to-video, cinematic: Kling v3 (kwaivgi/kling-v3-video): start_image, optional end_image
• References (images, video): Seedance 2 (bytedance/seedance-2.0)
• Open-weights, cheaper: Wan 2.7 (wan-video/wan-2.7-t2v, wan-video/wan-2.7-i2v)

Ask for a duration that covers the scene plus about 1 s of handle, then trim in Remotion. See skills/remotion-production/rules/replicate-models.md and rules/video-generation.md for the full flow, durations and provider choice.

5. Show Usage in Remotion

Place the clip with <Video> from @remotion/media (npx remotion add @remotion/media). Trimming, looping, volume and cutting several clips together follow the official remotion-markup skill (embedding-videos.md, video-editing.md). The path is what changes per provider: remotion-media clips are at public/<output_name>.mp4, Replicate downloads at public/footage/.

import { Video } from "@remotion/media";
import { staticFile, useVideoConfig } from "remotion";

const { fps } = useVideoConfig();

// remotion-media clip, trimmed to 3 s and starting 5 s into the scene
<Video
  src={staticFile("clip-city-aerial.mp4")}
  from={5 * fps}
  trimBefore={1 * fps}
  durationInFrames={3 * fps}
  muted
  objectFit="cover"
  style={{ width: "100%", height: "100%" }}
/>

6. Combine with Other Assets

AI-generated clips work great when layered:

• Use as backgrounds behind text overlays
• Intercut with stock footage from Pexels
• Add voiceover and music on top
• Apply subtle zoom/pan with interpolate() for extra dynamism

Tips

• Start short — 3-5 second clips are the sweet spot for AI video
• Be specific about camera motion — "slow dolly zoom", "overhead crane shot"
• Combine clips — Generate several short clips and sequence them in Remotion
• Image-to-video — Generate a still with /generate-image first, then animate it for more control
• Iterate — If the first generation isn't perfect, refine the prompt