The 157 cards are a production vocabulary, not a video generator
Video Shotcraft's README describes 157 shot recipe cards and 214 motion previews. Each card gives an agent a purpose, timing, parameters, implementation notes, and known failure points, while the demo tree supplies the corresponding TypeScript motion. That pairing is the useful part. An agent does not have to invent every easing curve from a card name, and a developer can inspect or change the code that produces a frame.
The skill still expects judgment. Its instructions begin by choosing a template, autonomous, guided, or single-shot mode. A full job can include product inspection, screen capture, visual direction, storyboarding, beat analysis, sound effects, implementation, rendering, and an independent final review. The repository makes those steps explicit, but it cannot decide whether your pricing screen is the right story or whether a camera move makes a feature easier to understand.
The 36.2-second Ink Press template is the practical starting point
Ink Press is a 36.2-second, 1920 by 1080 template running at 30 fps across 10 shots. It includes title cards, captured-page camera moves, transitions, captions, and a pinned sound pass. For a web or desktop product with a finished interface, this is the shortest route to a coherent result because the sequence and timing already exist. Your work becomes replacing screenshots, copy, colors, type, and product-specific emphasis.
The free-creation route is more demanding. Video Shotcraft tells agents to extract typography, spacing, color, and interface density from the target product, then carry those decisions into every title and transition. Real pages should be captured rather than rebuilt as fake UI. Customer data, secrets, and live account details must be removed before capture. Those rules are sensible, yet they depend on the agent following a long instruction set and the operator catching any miss.
What happened when we ran it
Our sandbox installed commit 5ddbf52 in 4 seconds, adding 57 packages and using 29 MB on disk. The checkout contained 988 files, about 48,041 lines of source, and occupied 54.8 MB. Our test method used an unprivileged Node 22 container with 3 CPUs, 8 GB of RAM, and no secrets. Npm audit found 0 known vulnerabilities.
The root package has no build target, so there was no build step to run. Vitest completed in 2 seconds with 23 passed and 0 failed. The scan found 3 CI workflow files and no conventional tests directory, though the root test script is present and passed. These checks cover helper behavior, not a complete 36.2-second render, visual quality, browser capture, audio timing, or the JianYing exporter.
Headless rendering has 3 documented failure points
The README names 3 problems on a 2-core Linux host. Remotion may reject its default concurrency, recent full Chrome or Chromium may fail because old headless mode is gone, and a blocked Remotion CDN can prevent the preferred browser shell from downloading. The documented route is concurrency 1 plus a local chrome-headless-shell binary. Anyone planning CI should prove that path on the actual runner before promising render times.
Dependencies also vary by chosen shot. Some reusable components need Three.js, React Three Fiber, and Remotion's Three integration. CameraMotionBlur requires another Remotion package. JianYing export uses a Python virtual environment and pyJianYingDraft. The macOS draft path was tested against JianYing Pro 11.2; the Windows module follows an upstream-supported route but has not been verified on a Windows machine.
The browser workbench can reopen a 10-shot template
After delivery, the skill can open a browser workbench on port 5198 and import a project manifest. The 10-shot template can then appear as separate shot, transition, caption, and sound tracks. Exposed properties let a user change copy, color, size, position, timing, or speed before exporting through Remotion again. This is more useful than handing a marketer an MP4 plus a directory of unexplained React components.
There are limits to that editability. A project needs the expected workbench manifest and schemas, and the editor only exposes properties the author chose to wire. The main workbench guide linked from the English README is marked Chinese. The repository does include English-facing overview material, but an English-only team may have to read code or translate the deeper operating guide.
Six open issues and PRs show the catalog is still moving
GitHub showed 6 open issues and pull requests on October 8, 2026, after a repository push on October 5. The latest release tag, published September 1, holds showcase media rather than a numbered software release. One active pull request proposes cutting the library to 124 cards and 150 demos after a frame-by-frame selection pass. That pending change is a sign of active curation, and also a warning against hard-coding today's card counts into your own automation.
Another open request describes an 80.8-second film with 2,424 frames where only frames 600 through 699 needed correction. The current workflow still favors full-film rerenders after changes, while the request proposes a safe patch path with audio reuse and boundary checks. That is a concrete cost for longer films. Video Shotcraft is most convincing when you want its opinionated process and editable code. It is much less attractive if your main need is cheap iteration on long timelines.