DEV Community

Cover image for video-shotcraft hands your agent shot cards, not a video model
Reno Lu
Reno Lu

Posted on

video-shotcraft hands your agent shot cards, not a video model

video-shotcraft is not a video model. It gives a coding agent a catalogue of shot recipe cards, 161 motion previews, and a finished Remotion project, then lets the agent write the animation as code. The output is a render from a real codebase.

The distinction matters because of what you can do afterward. A clip from a video model usually arrives without the composition behind it. A Remotion project is source: you can change a duration, swap a screenshot, retime a cut, and render again. video-shotcraft is built for exactly that loop, aimed at web and desktop product promos, with the individual cards reusable in feature demos, brand films, and launch videos.

Recipe cards instead of prompt roulette

Each shot card carries purpose, energy, suggested duration, parameters, implementation notes, and known pitfalls, and the repository ships tuned TSX demos holding the actual easing and timing values. Count the cards yourself before quoting a number: the README says 104 and the repository description says 106. The 161 motion previews cover 161 styles and are searchable and filterable in a public gallery, where you can compare variants and copy the card names you want.

That structure is the antidote to the usual failure mode of asking an agent for motion design. Rather than "make it feel cinematic", you say the shot you want: use the deck-deal-flyin and row-embed cards to present this feature, or design a product close-up inspired by spotlight-hero-card. The card names are the vocabulary, and the parameters inside them are the part an agent would otherwise have to invent.

If you name no card at all, the skill introduces its built-in template first and asks whether to use it, which is a reasonable default for someone who does not yet know what they want.

The template is the fast path

The bundled Ink Press template is a validated 36.2-second promo at 1920 by 1080 and 30fps: 10 shots in a paper-ink-amber style, with 2.5D camera moves over real page captures, title cards, transitions, and a fully pinned cinematic SFX pass. The workflow the README suggests is to have the agent swap in your product's screenshots, copy, and branding, reproducing the same structure with your material in it.

The project also eats its own cooking. The 38-second gallery intro was produced with the skill, with storyboard, shot implementation, and sound design all done by an agent following the toolkit's methodology. That is the most useful kind of demo, because the artifact and the claim are the same thing.

Installation is one line handed to the agent, or npx skills add Vincentwei1021/video-shotcraft, or a manual clone symlinked into ~/.claude/skills or ~/.codex/skills. License is Apache-2.0.

The headless notes are the part I would keep

Buried under the showcase is a short section on rendering from a headless Linux box, and it reads like it was written after a bad afternoon. Three walls, each with a fix. On a low-core box, remotion still and remotion render fail with "Maximum for --concurrency is 2", and the setting documented as working is --concurrency=1. Recent Chrome and Chromium dropped old headless mode, so pointing Remotion at the system chromium fails to launch and you want a chrome-headless-shell binary instead. If remotion.media is unreachable, the automatic headless-shell download is rejected, so you pass --browser-executable with a local path.

None of that is glamorous, and all of it is what stands between a working local demo and frames that render on a build box: with those three flags, the README reports that renders from the bundled template work. A toolkit that documents its own failure modes alongside the flags that resolve them is telling you where it has actually been run.

What the README does not claim is worth noting too: there are no numbers on render times, no comparison against video models, and one template rather than a library of them, with more described as on the way. The bet is that structured shot vocabulary plus real code beats one-shot generation for product video, and the gallery is there so you can judge that yourself before installing anything.


GitHub: https://github.com/Vincentwei1021/video-shotcraft


Curated by Agent Palisade — practical AI for small and mid-sized businesses.

Top comments (0)