Drop in any winning video ad. Get back a buildable recipe for recreating its look with AI: the artistic style, a scene-by-scene storyboard, the characters to lock, the exact image and video models, and a cost estimate. Gemini watches the video. Claude writes the recipe. You bring the taste.
This is the functional version of this week’s issue. No fluff. Copy what you need.
DOWNLOAD IT HERE 👇
video-style-recipe.zip
DOWNLOAD IT HERE 👆
Videos Referenced
AQPEL2YORnSShqaC22HCigQG4V1JSWRfCHS8vKwLHVm8s5KVqsnFTGuUGoJt85xukGJw0xFXQaHE8c1gL59q8BI5LOG_jgVt.mp4
Video I yeeted
BER-V02-minecraft-heartcut.mp4
Video I created
What you need (one time)
- Claude Code. This ships as a Claude Code skill. claude.com/claude-code
- Python 3.10 or newer.
- A Gemini API key, free at aistudio.google.com/apikey. The only key required to analyze.
- To render later, a fal.ai account and an ElevenLabs key for voice and music.
- Install the pack into .claude/skills/video-style-recipe/, then run bash setup.sh from the pack folder and paste your key into .env as GEMINI_API_KEY.
The core loop
- Analyze the reference with Gemini. Run the pack’s analyze_video.py against your downloaded reference. It uploads the video to gemini-2.5-flash and writes a structured analysis JSON.
- Read the JSON. It holds the archetype, the artistic style (render mode, palette, lighting, and a fal_can_match flag), the format, the pacing, a scene list (timestamps, shot scale, camera move, on-screen text, plus an image and motion prompt seed per scene), the characters, the products, the audio, and the recommended models. Checkpoint: confirm the detected style, whether fal_can_match is true, and the scene count before going further.
- Pick the fal stack for the detected style. See below.
- Write the recipe. Fill the template below from the JSON and save it next to the source.
- Build, opt-in. Master still first, lock the character, anchor every other scene to it, animate, voice and music, assemble.