replicate-video — Claude Code skill

A Claude Code skill that generates short videos with Replicate's bytedance/seedance-2.0 model. Supports text-to-video and reference-image conditioning (image-to-video, multi-reference style/subject control).

When invoked, Claude will:

1. Read SKILL.md to know the model defaults, prompting tips, and when to fall back to other tools.
2. Run scripts/replicate_video.py generate ... with the right flags.
3. Save the MP4 to ./output_videos/seedance_<ts>.mp4 and return a RESULT: { ... } JSON line with the local path + Replicate URL.

Install

Clone into your project's skills folder:

# Anthropic's convention: skills live in .claude/skills (or ~/.claude/skills for user-wide)
git clone https://github.com/Bomx/replicate-video-skill .claude/skills/replicate-video
pip install -r .claude/skills/replicate-video/requirements.txt

Add your Replicate token to the project .env:

REPLICATE_API_TOKEN=r8_...

Use it from Claude Code

After restarting your session, type:

/replicate-video make a 7s cinematic drone pull-back over a misty pine forest, 1080p

Or just ask in plain English — the skill auto-triggers on phrases like "make a video", "animate this image", "text-to-video", etc.

Use the CLI directly

# Text-to-video
python scripts/replicate_video.py generate \
  --prompt "Cinematic aerial drone pulling back over a misty pine forest at sunrise, 35mm" \
  --duration 7 --resolution 1080p --aspect-ratio 16:9

# Image-to-video with multiple references
python scripts/replicate_video.py generate \
  --prompt "Aerial drone descends and enters through the front door of this inspection center" \
  --reference-image ./storefront.jpg \
  --reference-image ./interior_1.jpg \
  --reference-image https://example.com/google_earth.png \
  --duration 7 --resolution 1080p --seed 24353

# Vertical (TikTok / Reels / Shorts)
python scripts/replicate_video.py generate \
  --prompt "..." --aspect-ratio 9:16 --duration 5

# With audio reference + generated audio
python scripts/replicate_video.py generate \
  --prompt "..." --generate-audio --reference-audio ./voiceover.mp3

The last stdout line is always machine-parseable:

RESULT: {"status":"succeeded","output_url":"https://replicate.delivery/.../tmp.mp4","local_path":"output_videos/seedance_1714857600.mp4","duration_s":7,"resolution":"1080p","aspect_ratio":"16:9","model":"bytedance/seedance-2.0","elapsed_s":110.0}

Defaults

Flag	Default	Notes
--aspect-ratio	16:9	also 9:16, 1:1, 4:3, 3:4, 21:9
--duration	7	seconds, typical range 3–10
--resolution	1080p	also 720p, 480p
--generate-audio	off	flip on for ambient/sfx
--seed	random	pin for reproducibility
--model	bytedance/seedance-2.0	pass owner/name:hash to pin a version

Cost & latency

Seedance 2.0 generations take ~1.5–4 minutes depending on resolution and length. Each run costs real money (per Replicate's pricing). The CLI polls until done — Ctrl+C cancels.

Prompting tips

• Lead with camera move: "aerial drone pulling back", "slow dolly in", "static wide", "handheld follow".
• Then subject + environment + lighting/lens: "35mm cinematic, golden hour, shallow depth of field".
• For interior reveals from exterior, name the transition: "camera approaches and enters through the front door; interior matches reference image 1".
• Tell the model what each reference is for: "reference 1 is the storefront, references 2–4 are interior, references 5+ are Google Earth overheads".
• Keep prompts under ~80 words. Seedance ignores very long prose.

License

MIT — see LICENSE.