One skill produces a 45 to 60 second editable film
Guizang Product Video Skill gives Claude Code or Codex a production workflow for software promos and release videos. The default brief is a 45 to 60 second Chinese film in horizontal format. The agent reads the repository or changelog, chooses claims a viewer can understand, mounts product components, animates them, builds an audio track, and exports an MP4 plus the project used to render it. The repository documentation is Chinese, and the README does not link an English translation.
Three visual modes control how much of the product's identity survives. repo keeps the existing design, default wraps scenes in the bundled CodePilot look, and hybrid keeps the interface while changing the surrounding composition. Product text must remain at least 22 px in a 1080p film. That rule matters because a technically faithful interface can still become unreadable after it is squeezed into a social video frame.
Seven stages keep the agent inside a reviewable process
The seven-stage workflow starts by fixing the version range, product surface, audience, channel, aspect ratio, language, and style. It then checks the environment, researches the code, mounts one real component, writes a direction file, builds each scene, mixes audio, and reviews the render. Scope must come from the user because a repository cannot decide whether the film should cover one release, three weeks of work, or the entire product.
A formal film needs DIRECTION.md, three distinct creative directions, and 3 to 5 visual choices tied to something specific in the product. The first 5 to 10 frames must already form a complete poster image because X and Threads may use the opening frame as a thumbnail. Delivery checks inspect that opening, the timeline, expected files, and audio evidence. The instructions still require a person to watch the frames and listen to the mix.
What happened when we ran it
Our sandbox installed commit c2ce349 in 12 seconds, pulled 35 packages, and occupied 37 MB on disk. The build succeeded in 6 seconds. Pytest finished in 7 seconds with 15 passed, 0 failed, and 3 skipped of 15. Pip-audit reported 0 known vulnerabilities. The container had 3 CPUs, 8 GB of RAM, Python 3.12 on Debian, no secrets, and no elevated privileges.
The checkout contained 90 files and about 2,036 lines of source, taking 1.4 MB before installation. It had a tests directory, no Dockerfile, and 0 CI workflow files. Those results cover installation, the supplied build, and the supplied tests. They do not grade a finished film, measure browser rendering speed, test a product integration, or prove that the audio workflow behaves on every FFmpeg version.
The browser starter needs four pieces of system software
The documented baseline is Node.js 22 or newer, Python 3.9 or newer, FFmpeg with FFprobe, and Playwright Chromium. Inside the video project, React supplies the UI, GSAP drives the main timeline, Three.js or canvas handles effect layers, and esbuild prepares the browser code. The headless export turns on SwiftShader for WebGL, so a machine without a usable GPU still has a software-rendered path.
React products get the shortest route because the skill includes component integration examples. Monorepo packages and unusual TypeScript options can be mapped in integration.config.mjs without changing the product repository. Server-dependent, native, or context-heavy components may need providers, fixtures, or an adapter. If direct mounting fails, the instructions allow a source-faithful reconstruction, but require the evidence record to say what was rebuilt and why.
An open FFmpeg 8.x report reached 112 GB
Open issue 5 describes a 60 second film with 28 sound cues on FFmpeg 8.1. One of 6 repeated runs kept writing silence, reached 112 GB in about 4 minutes, and stopped only when the disk filled. Open pull request 4 reports a related 48 second, 21-cue case on FFmpeg 8.0.1 that reached 23 GB. Both reports propose bounding the SFX stem more directly than the current apad filter chain.
This is a narrow version-specific report, not a failure from our Debian sandbox. It still changes how we would trial the tool. Put early FFmpeg 8.x runs on a volume with a quota, watch the output size, and test the exact sound plan more than once before leaving a render unattended. A product-video skill that can fill a workstation disk needs that warning near the install command, even when its Python tests pass.
AGPL code and BSL fallback assets need separate review
The main workflow, scripts, starter, and original audio assets use AGPL-3.0. The CodePilot-derived files under assets/fallback/ retain Business Source License 1.1 terms, so the default visual pack does not inherit the main license. The README says generated video is not automatically an AGPL work merely because this skill rendered it. Product components, logos, fonts, music, and other incorporated assets keep their own terms.
The project has 689 stars but no release tag
GitHub showed 689 stars, 60 forks, and 4 combined open issues and pull requests on October 6, 2026. The open list contained 2 issues and 2 pull requests. The repository was created on September 18 and last pushed on October 1, when commit c2ce349 added first-frame thumbnail checks. There were no GitHub releases or tags. That is active early work, with a clean small test run and very little release history to judge.
Guizang is best for a developer who wants an agent to make the film while leaving every frame and audio decision open to inspection. Recordly is easier when a polished screen capture is enough. Remotion gives a team a lower-level React foundation. For this skill, the deciding costs are Chinese-only guidance, the local media stack, license review, and a supervised FFmpeg trial.