launch-video-clone

A Claude Code skill for making product launch videos in HyperFrames that don't look AI-made.

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 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.

Modes

• Original (no reference video): pick the category that fits the product (UI showcase, abstract 3D diagram, photoreal 3D product, editorial collage, liquid-glass UI, kinetic type into UI), take cut rhythm and easing numbers from 1-3 donor videos of that category, write the script and a shot-by-shot breakdown, build it in HyperFrames and self-review it against the motion rules.

• Clone (a reference video is given): rebuild the reference as closely as possible (same shots, timing, easing, transitions and type), optionally re-skinned for another brand. The model can't watch video, so scripts/burst.py turns the reference into per-shot frame sheets, motion trails, easing curves and beat maps it can read, and scripts/compare.py grades every render against the reference frame for frame until they match.

• Hybrid (real camera footage + motion graphics): four sub-skills in subskills/: paper-world (footage → paper-collage 3D world → footage), word-synced-ui (a product UI typing the speaker's real words), footage-transitions (iris, cloud and paper passes, match cuts, options for review) and vo-beat-timing (frame-numbered timing sheet, safe mixing, Whisper checks on the final mix). Each is a standalone SKILL.md with its own scripts, so it can also be installed on its own.

Every mode ends by writing a LESSONS.md so the skill improves with each video.

Layout

SKILL.md              the workflow (start here)
types/                one file per kind of motion: build approach, traps, lessons
references/           motion-feel rules, living-stills recipes, the corpus of public techniques,
                      and worked example code from past clones (examples/)
scripts/              analysis, tracking, comparison and sound tools
subskills/            hybrid-mode sub-skills (paper-world, word-synced-ui, footage-transitions,
                      vo-beat-timing), each with SKILL.md + scripts/

Install

Copy this folder to either location:

# for all your projects
cp -R launch-video-clone ~/.claude/skills/launch-video-clone
# or for one project
cp -R launch-video-clone <project>/.claude/skills/launch-video-clone

Then ask Claude Code for "a launch video for " or "clone this launch video: ". In the skill, $SKILL_DIR means the folder you copied it to; set S="$SKILL_DIR/scripts".

Requirements

• HyperFrames CLI (npx hyperframes ..., Node 18+)
• ffmpeg / ffprobe on PATH
• python3 with: numpy pillow opencv-python scipy matplotlib

  pip install numpy pillow opencv-python scipy matplotlib
  

• yt-dlp to download a reference video (optional)
• Blender for photoreal 3D product shots and the paper-world sub-skill (optional)
• For the hybrid sub-skills (optional): a Whisper with word timestamps (mlx_whisper on Apple silicon or openai-whisper), a segmentation model (SAM / SAM 2, or Apple Vision on macOS), an inpainting model (LaMa), and Playwright/Chrome for rendering HTML overlays to RGBA frames

Sound (optional):

• ELEVENLABS_API_KEY for scripts/eleven.py (music and sound effects)
• GEMINI_API_KEY for scripts/listen.py (a model that listens to the audio)
• A separate venv for scripts/soundcheck.py, in your project:

  python3 -m venv .venv-sound
  .venv-sound/bin/pip install numpy torch audiobox_aesthetics google-genai
  .venv-sound/bin/python "$S/soundcheck.py" check renders/v1.mp4 ...
  

API keys are read from environment variables only. Nothing reads .env files for you: export them in your shell (export ELEVENLABS_API_KEY=...).

Images (optional): an image generator that accepts reference images (gpt-image or similar), for photoreal plates. UI, type and logos are always built in code or taken from real files.

Renders are large (HyperFrames writes one PNG per frame as scratch), so render outside the repo, e.g. ./renders (gitignored here) or an external drive.

License

MIT. See LICENSE.