Stop, don't scroll past! Let's watch two clips first—remember to turn on the sound 😐.
The first is a launch video I made with Opus 5.5 when DeepSeek-V4.1-Flash-1017 was released yesterday. After posting it on X, many people thought it was an official DeepSeek promo:
https://x.com/AlchainHust/status/2111396576123830528
https://x.com/deepseek_ai/status/2111375640717033777
The second is a demo I created to test how well DeepSeek's new model could showcase its own motion capabilities:
https://x.com/AlchainHust/status/2111416302149505270
22 seconds, with music and sound effects, took less than half an hour from start to finish. The cost was about $0.27 (with a 99.4% cache hit rate).
Of course, neither clip was made with a simple "one-line prompt" because there's a skill behind them—the open-source star of today: Huashu-motion.
As you can see in the title, you can understand this as a skill that helps you use coding agents to create commercial-grade product promo videos.
Once huashu-motion is installed, you just tell Claude Code or Codex, "Make a 30-second launch video for my product." It will first read your product details (or search online for you), offer three design directions like Huashu-Design does, write the entire script once you choose one, compose the music itself, and then open an editor in your browser. If something isn't right, press C, click on the screen, write what needs changing, and it will only modify that specific part.
Benchmarking Against Tech Giant Launch Videos
This project started with Claude Motion. Anthropic released it on October 8th, allowing Claude to turn reports, charts, and product explanations into short animations using code, with interactive editing features. However, I found that it is still only in beta for Team and Enterprise plans. When Claude Design came out in April, I reverse-engineered its core capabilities into a skill, which became huashu-design. So this time, I had Claude Code start working directly with one requirement:
Our skill needs to fully replicate Claude Motion's capabilities but be much stronger!
What we produced on the first day looked boring to me. We had editors, version control, over a hundred tests, and a 19-second 3D city sample, but... it was terrible. The camera movements lacked rhythm, were too steady, meaning mediocre... This wasn't just about adjusting the animation for this instance.
So later, I set a goal for the agent, painting a vision that our skill should help users create top-tier work. We wanted visual effects, motion, sound effects, and music that convey the level of million-dollar product launch ads!
Therefore, our testing and production target is to enable you to create commercial-grade promo videos like those from Apple, Google, and Anthropic.
With such high standards, we first needed to understand how those videos are actually made. So it downloaded 21 real launch and opening videos from YouTube (Apple, Nothing, GoPro, Google, Vercel, Linear, and Anthropic's own) and used ffmpeg to measure cut points, beats, loudness, energy curves, and colors for each one. It also analyzed 142 high-engagement launch and motion videos, organizing techniques by product type.
These measured insights eventually became the director rules for huashu-motion. The subsequent steps—reading products, finding colors, writing rhythm sheets, composing music, cold reviews—are essentially breaking down the methods from hundreds of videos and teaching them step-by-step to the agent. All our tests used real product launch videos: DeepSeek V4.1 Flash, Kitten Fill Light Pro, huashu-design 3.0, Nüwa, and several friends' products.
What Happens After One Sentence
"One-click output" means this: You input one sentence, and you get a finished video with music, ready to publish. Any changes you make afterward are just fine-tuning details. The simplest usage is to tell it to just do it; it will automatically choose the safest direction, explain why, and mention the other two directions briefly upon delivery. It will still complete all steps like rhythm sheets and checks on its own.
If you want to participate more, here are roughly nine steps it will follow:
- First, read your previous preferences and experiences accumulated from past videos;
- Read your product details—website, App Store page, README, logo, brand colors—and send you a message with default settings for how it plans to proceed. You can change any disagreement and decide whether to include voiceover;
- Provide three directions, each with a rendered keyframe and a card showing color palette, rhythm, and music. You pick one or mix them;
- Once selected, write a treatment.md (rhythm sheet). Work cannot begin until this is done;
- Create the hardest shot first at final resolution, fonts, and colors, then do the rest;
- Run checks automatically, capture keyframes to verify if text is cropped, overlapping UI, or empty shots;
- Compose music and add sound effects synced to every element landing on screen;
GIF
- Open an editor in the browser and wait for you to point and edit;
GIF
- When you approve, export it. It collects your feedback and edits from this session, delivering your promo video. If unsatisfied, you can still freely modify it.
Twenty Videos in One Day
At 2 AM the day before yesterday, I said goodbye to sleep and told it:
I'm going to sleep now. Finish all remaining promo videos. Make all decisions yourself!
When I woke up, the remaining five videos were done. During the day, I opened multiple windows in FanBox to continue making more, totaling about twenty videos.
GIF
This segment is from huashu-motion's own launch video, featuring 16 shots from seven different videos, cut to the beat.
Kitten Fill Light Pro: The first vertical 16-second version became the App Store preview video and has been submitted for review. Later, three new directions were created using the new rhythm method. Here is one of them, switching through eight lighting colors per beat:
GIF
Grok Bot (fan-made, unofficial): A 30-second video where tasks are assigned in the evening, agents work independently overnight, and four reports arrive on your phone at 7 AM:
Get This Skill
The skill is open-sourced on GitHub:
https://github.com/alchaincyf/huashu-motion
Installation:
1npx skills add alchaincyf/huashu-motion
Or simply tell your agent:
1Help me install the huashu-motion skill2(github.com/alchaincyf/huashu-motion), then run a demo after installation
Requires Node 22+, Chrome (or Chromium/Edge) on your computer, and FFmpeg for video export and audio mixing. Currently tested only on macOS. For real piano/string instrument sounds, have the agent run om samples build to download ~55MB of free samples (CC0 and CC-BY licensed). Without this, synthetic sounds are used. Run om doctor to check missing dependencies; the agent can handle installations.
After installation, you can say things like:
1Use huashu-motion to make a 30-second launch video for my product, website is xxx.com
1Use huashu-motion to turn {data chart/PPT} into a 30-second explainer, clarifying why our registration rate doubled
1Use huashu-motion to make a launch video with voiceover, using my recording: {video/audio file}
I mainly use it in Claude Code with Opus 5.5. As seen above, DeepSeek V4.1 Flash also works well, though taste differences between models are real. Codex has been tested; Cursor, OpenClaw, Gemini CLI, and other skill-supporting agents should work too, though I haven't tested each individually. No extra API keys required.
Here are some limitations: It cannot hear, so music quality relies on checking rhythm sheets, energy curves, and spectrograms—you need to listen yourself. It cannot generate live-action footage or real people; use your own materials for that, and it handles synthesis and packaging.
The repo includes 20 motion languages, 10 starter templates, 23 sound effects, 3 demo scores, and methodology docs for directing, rhythm, color, sound, and voiceover.
Feel free to try it in WorkBuddy, Doubao Work, Grok bot, Muse, etc., and share results or issues with me.
One more thing! Since this is a product promo video tool, let it introduce its own launch 👇
