Launch Video Clone
Makes product launch videos in HyperFrames that don't look AI-made, from scratch or cloned shot by shot from a reference.
I built this because out of the box, a model makes the generic "AI motion" look: sliding text, safe fonts, the same ease on everything. The launch videos that look designed got there by studying a real reference shot by shot and fixing the timing and easing over many passes.
This skill turns that into a loop an agent can run, with scripts that measure motion instead of guessing it. The model can't watch video, so the skill turns a reference into frame sheets, motion trails and easing curves it can read, then grades every render against it.
What it does
- Original mode: picks the right category for the product (UI showcase, abstract 3D diagram, photoreal 3D product, editorial collage, liquid-glass UI, kinetic type into UI) and borrows cut rhythm and easing numbers from 1 to 3 donor videos.
- Clone mode: rebuilds a reference video as closely as possible, with the same shots, timing, easing, transitions and type, optionally re-skinned for another brand.
- Hybrid mode: mixes real camera footage with motion graphics, such as a paper-collage 3D world, a product UI typing the speaker's real words, and footage transitions.
- Grades every render against the reference frame for frame until they match.
- Writes a LESSONS.md at the end of every video, so the skill improves each time.
How it works
- Original: write the script and a shot-by-shot breakdown, build it in HyperFrames, then self-review against the motion rules.
- Clone: scripts/burst.py turns the reference into per-shot frame sheets, motion trails, easing curves and beat maps.
- Build each shot in HyperFrames, render, and run scripts/compare.py against the reference.
- Fix the timing and easing that don't match, and repeat until they do.
What you need
- Claude Code
- The HyperFrames CLI (Node 18+)
- ffmpeg and ffprobe
- python3 with numpy, pillow, opencv-python, scipy and matplotlib
- Optional: yt-dlp for references, Blender for photoreal 3D, ElevenLabs and Gemini keys for sound
Install it
git clone https://github.com/per-simmons/launch-video-clone.git
cp -R launch-video-clone ~/.claude/skills/launch-video-cloneThat installs it for every project. For one project only, copy it into that project's .claude/skills/ folder instead.
The full instructions, and the skill itself, are on GitHub.
Try asking
- “Make a launch video for [product].”
- “Clone this launch video: [URL].”
- “Clone this launch video and re-skin it for my brand.”
Good to know
- Renders are large, because HyperFrames writes one PNG per frame as scratch. Render outside the repo or to an external drive.
- API keys for the sound tools are read from environment variables only. Nothing reads .env files for you.
