The video-shotcraft project provides a library of motion-design instructions for AI agents writing video code. It addresses the numerical decisions behind animation timing, transitions and camera movement that an agent would otherwise have to invent.

Unlike a typical video generator that turns a prompt into pixels, the system described here does not use a neural network to generate the video itself. Instead, deterministic React code defines the full video, and Remotion renders it frame by frame in a browser before stitching the frames into an mp4 file. AI’s role is to write that code correctly.

That task requires decisions such as how many frames to hold a logo after it lands, which deceleration curve to use for a camera push-in and how long before a cut to trigger a flash. The repository author says an agent asked only to “make it look nice” will invent those values, producing what the author calls a “slideshow.”

Shot cards specify motion-design decisions

Video-shotcraft uses shot cards as configuration protocols rather than simple descriptions of visual effects. One example, the brand-ink-open card, covers the opening used in the template video.

The cards also document known pitfalls through accounts of previous fixes. The brand-ink-open card says: “the pause should be given to the moment the logo lands, not to a regular text card — once it was mistakenly added to a regular card and quickly rolled back.” The material presents lessons drawn from production work rather than a generated reference.

The library contains 157 cards in 10 categories: 28 ui-entrance cards covering the appearance and construction of user-interface elements; 26 typography cards for text, titles and text cards; 19 transition cards for transitions and cuts; and 17 effects cards covering particles, light and materials.

It also includes 15 interaction cards for product interaction demonstrations, 13 data cards for numbers, charts and counters, 11 opening cards for opening shots, 11 rhythm cards for rhythm and hitting the beat, 10 camera cards for movement and space, and seven outro cards for final assembly and logos.

Implementations and previews

Each card has an implementation in the demos/ directory, which contains 221 TSX files. Rather than paraphrasing a card, the agent reads the full document and opens the corresponding source code, including its actual coefficients.

The project’s online gallery provides 214 motion previews with search and filtering. The repository itself does not contain those previews. They are stored in a GitHub release and downloaded separately by the gallery/fetch-media.sh script through gh release download, meaning a local copy of the gallery initially opens without media.

Source: Incrypted