How to use the Short-Form Video Clips agent skill (video-shortform)

What the skill makes the agent do

Once installed, the agent plans a single short shot, such as a product reveal, a motion teaser or an ambient loop. It describes the subject, camera, lighting and motion, then sends that prompt to a video model. The source names Seedance 2 as the default, with Kling 3 or 4, Veo 3 and Sora 2 also supported.

The resulting MP4 is saved into your project folder so you can play it in the file viewer. The skill is built around one shot per request: text-to-video models handle one clear idea well but are weaker at multi-cut stories.

• Plans one shot with subject, camera, lighting, motion and (if supported) sound
• Sends the prompt through a unified media command rather than calling provider APIs directly
• Saves one MP4 to the project folder
• Includes a special path for a local HTML-to-MP4 motion-design renderer (hyperframes-html), which gets a short motion-graphics brief instead of realistic camera directions

Install the video-shortform skill

Download the ZIP from the skill page on this site and unzip it. The skill folder contains SKILL.md and example.html. Put the unzipped folder in one of two places: ~/.claude/skills/ to make it available everywhere, or a project's .claude/skills/ folder to limit it to that project.

If you use the Claude apps instead of files, upload the ZIP under Settings → Capabilities → Skills.

• Option A: unzip into ~/.claude/skills/ for a personal install
• Option B: unzip into your project's .claude/skills/ folder for a project-only install
• Option C: upload the ZIP in the Claude apps under Settings → Capabilities → Skills

Check what you need before you start

The skill needs the Open Design runtime, which provides the OD_NODE_BIN and OD_BIN media contract that the skill calls. You also need a video generation provider configured in your workspace, such as Seedance, Kling, Veo or Sora. Those providers typically require their own API key or account, which you set up with them directly.

A local hyperframes-html renderer is optional. You only need it if you want motion-design briefs rather than realistic camera shots.

• Open Design runtime with OD_NODE_BIN and OD_BIN available
• A video provider configured in the workspace (usually needs its own API key or account)
• Optional: local hyperframes-html renderer for motion-design briefs

Trigger the skill with a natural request

The skill responds to trigger words including video, clip, shortform and reel (the source also lists Chinese terms 短视频 and 动效). Describe the shot you want in plain language and include the subject, the camera movement and the lighting if you have opinions about them.

The skill reads the project's video settings for model, length in seconds and aspect ratio. These are fixed by the project, so set them there rather than in your request. The skill clamps the prompt to what the chosen model supports; for example, Seedance 2 caps at 10 seconds and Veo 3 supports 8 seconds with audio.

• Good: "Make a 5-second clip of a ceramic mug rotating on a paper backdrop with warm side-light."
• Good: "Create a short looping ambient clip of dust motes drifting in a sunbeam."
• Keep one idea per request. Ask for a second clip for a second shot.

Examples you can adapt

The skill ships with an example prompt in its metadata. It describes a 5-second product reveal: a ceramic coffee mug rotating on a soft paper backdrop, warm side-light from camera-left, and dust motes drifting through the beam, in 16:9 with a slow camera drift. You can reuse that structure for your own products.

When the agent finishes, it replies with a short shot summary, the filename returned by the dispatcher, and one idea for a variation you can ask for.

Safety notes: scripts and network access

This skill runs scripts and uses the network. The media command sends your prompt to the video provider you have configured, so do not put confidential material in the prompt unless you are comfortable with that provider handling it.

The skill's rules say not to call provider APIs by hand and not to retry silently when a generation fails. If a provider rejects a prompt (for example for content policy reasons), the agent is meant to give you the suggested next step rather than paste raw provider error text.

• Runs scripts and uses the network: review the SKILL.md before installing
• Prompts are sent to your configured video provider
• Failed renders are not retried silently

Tips for better clips

Keep clips at 10 seconds or less and plan one shot per clip. Text-to-video models are inconsistent, so expect to re-render. Re-renders are slow and can be costly, so describe motion and camera clearly the first time.

For motion-graphics titles or interstitials, ask for a motion-design card rather than a realistic scene. The hyperframes-html path is built for that.

• Specify subject motion and camera motion separately
• Name the light direction and colour temperature
• Ask for a looping ambient clip when you want a seamless background