Skip to content
Get packs

Guide · Agent skill

How to use the Video Generation agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

The video-generation skill is an instruction pack for an AI agent that turns a written scene description into a short MP4 video clip. It is aimed at creators, developers building agent workflows, and marketers who want to prototype visual concepts. After reading this guide you will know how to install the skill, what it needs before it works, how to phrase a request so the agent triggers it, and how to read the results.

Video Generationbytedance/deer-flow · MIT

Quick start

  1. 1Download the ZIP. Download the video-generation skill ZIP from our page.
  2. 2Install the skill. Unzip it into ~/.claude/skills/ for personal use or .claude/skills/ in a project, or upload it under Settings → Capabilities → Skills.
  3. 3Set your API key. Export GEMINI_API_KEY for Gemini Veo, or MINIMAX_API_KEY if you want MiniMax, and make sure Python with requests is available.
  4. 4Describe the video. Ask the agent for a video, naming the subject, style, camera, aspect ratio and any dialogue or sound.
  5. 5Add a reference image (optional). Provide a .png, .jpg, .jpeg, .webp or .gif file to guide the first or key frame.
  6. 6Check the result. Wait for the script to succeed, then review the MP4 the agent shares and ask for changes if needed.

What the video-generation skill makes the agent do

Agent skills are instruction packs, made of a SKILL.md file plus optional scripts, that teach an AI agent how to perform one specific task. This skill covers generating, creating or imagining videos. When you ask for a video, the agent writes a structured JSON description of the scene, covering subject, camera, dialogue and audio. It can optionally use a reference image as the first or guiding frame, then runs a bundled Python script that sends the request to a video AI service.

Once generation succeeds, the agent shares the resulting MP4 file. If generation fails, the script exits with status 1 and writes errors to stderr, and the agent should not present a video.

  • Writes a JSON prompt file in /mnt/user-data/workspace/
  • Optionally uses a reference image as a guiding or first/last frame
  • Calls scripts/generate.py to produce an MP4

Install the video-generation skill

Download the ZIP for this skill from our page. Then choose one of two routes.

For a personal install, unzip the file into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder. Alternatively, upload the skill in the Claude apps under Settings → Capabilities → Skills.

  • Personal: unzip into ~/.claude/skills/
  • Project: unzip into .claude/skills/ inside your project
  • Claude apps: Settings → Capabilities → Skills, then upload

Set up what the skill needs

The skill needs Python with the requests library, network access to the provider you choose, and one API key. It uses Gemini Veo by default, which requires GEMINI_API_KEY. If only MINIMAX_API_KEY is set, it uses MiniMax video instead. Keys are read from the runtime environment, so do not paste them into prompt files or command-line arguments.

You can force a provider with VIDEO_GENERATION_PROVIDER set to gemini or minimax. MiniMax also reads MINIMAX_API_HOST (default https://api.minimaxi.com) and MINIMAX_VIDEO_MODEL (default MiniMax-Hailuo-2.3).

  • Python with the requests library installed
  • GEMINI_API_KEY for Gemini Veo (default)
  • MINIMAX_API_KEY for MiniMax video (used when only this key is set)
  • Network access to the chosen provider

Trigger the skill with a natural request

The agent loads the skill when you ask it to generate, create or imagine a video. Describe the scene concretely: who or what is in it, the style, mood and colour palette, the camera, lighting and aspect ratio, and any dialogue or sound. The agent identifies these details before it writes the prompt, so missing details will be filled in by the agent rather than by you.

A reference image is optional. If you provide one, the agent uses it as the guiding or first frame. The skill notes that a reference image can improve quality significantly, and that iterative refinement is normal.

  • Good: "Create a 16:9 video of a fox crossing a snowy field at dawn, slow dolly-in, soft wind sound."
  • Weaker: "Make a video."

Run the generation script directly

The agent normally runs the script itself, but you can see the exact call it uses. The script takes an absolute path to a JSON prompt file, optional reference images, an absolute output path, and an aspect ratio that defaults to 16:9. The skill's instructions say not to read the Python file and to call it with the parameters instead.

Reference images must have a recognised extension: .png, .jpg, .jpeg, .webp or .gif, matched case-insensitively. Other extensions are rejected before submission. The script does not check image contents, and MiniMax ignores --aspect-ratio because it uses resolution and duration instead.

  • --prompt-file: absolute path to the JSON prompt (required)
  • --reference-images: absolute paths to reference images (optional)
  • --output-file: absolute path to the output file (required)
  • --aspect-ratio: default 16:9 (optional, ignored by MiniMax)
python /mnt/skills/public/video-generation/scripts/generate.py \ --prompt-file /mnt/user-data/workspace/prompt-file.json \ --reference-images /path/to/ref1.jpg \ --output-file /mnt/user-data/outputs/generated-video.mp4 \ --aspect-ratio 16:9

A worked example: a film-style scene

The skill's example asks for a short clip of a train station farewell from The Chronicles of Narnia. The agent searches for the scene, writes a JSON prompt with a background, characters, camera settings, dialogue and audio levels, then generates a reference image with the image-generation skill. Finally it runs generate.py with that prompt, the reference image and a 16:9 output path.

Be aware that this example uses copyrighted film material. The skill gives no guardrails for this, so use original scene descriptions for your own work.

  • Prompts are written in English, regardless of the user's language
  • The JSON format keeps prompts structured and parsable

Read the output and iterate

After a successful run, the video is typically saved in /mnt/user-data/outputs/. The agent should share the video first, along with any reference image, and give a brief description of the result. If the result is not right, ask the agent to adjust the prompt and generate again.

Do not expect identical results each time. Output depends on the external provider, so repeatable or production-grade footage is not guaranteed.

  • Check that the command succeeded before expecting a video
  • A file left over from an earlier run is not a new result
  • Ask for adjustments to refine the clip

What people use it for

  • Short concept clips built from a written scene brief for a product or campaign
  • Prototype visual ideas before committing to a full shoot
  • Testing video output inside a developer's agent workflow
  • Turning a reference still into a short moving clip

Frequently asked questions

How do I install the video generation skill?

Download the ZIP from our page, then unzip it into ~/.claude/skills/ for personal use or into .claude/skills/ inside a project. You can also upload it in the Claude apps under Settings → Capabilities → Skills.

Is the video-generation skill free?

The skill is published under the MIT licence. The video provider is separate: the skill needs a GEMINI_API_KEY or MINIMAX_API_KEY, and this guide does not state provider pricing.

What do I need to run it?

You need Python with the requests library, network access to your chosen provider, and either GEMINI_API_KEY for Gemini Veo (the default) or MINIMAX_API_KEY for MiniMax video. Without a valid key the script fails with a credentials error and produces no video.

Why did my video generation fail?

The script exits with status 1 and writes the error to stderr. Common causes include missing provider credentials, or a reference image with an unsupported extension, which is rejected before submission. Check the error message and fix the cause before retrying.

Can I use Gemini and MiniMax together?

Yes, but the skill chooses based on environment variables. If GEMINI_API_KEY is set, Gemini Veo is used by default. You can force a provider with VIDEO_GENERATION_PROVIDER set to gemini or minimax.

Can I use it to make videos of real people or copyrighted scenes?

The skill gives no guardrails for generating videos of real people or recreating copyrighted scenes. Only do this where you hold the rights and permission, and check the rules that apply to you.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

Choose this skill if you want short AI video clips from a written scene description, or you are building an agent workflow that needs video output and you have a Gemini or MiniMax API key. Pick something else if you need repeatable, production-grade footage, or if you cannot set up an API key and network access.

Download the skill