Image: GitHub
Why it mattersA founder can ask Claude for a 20 second launch film. The skill writes the code. A built-in script checks the output before you send it anywhere.
A founder wants a 20 second launch film by Friday. The hire for it is five figures and a week. Video-director is a free Claude Code skill that writes the code for the film. The code renders every frame to an MP4. A script then checks the result before the file goes out.
The project is by win4r. It sits at 194 stars with 15 forks on GitHub, under the MIT licence. The skill is written for Claude Opus 5.5 in Claude Code. It is a set of instructions and quality checks, not a model and not a hosted service.
What it can make
The readme lists the kinds of film the skill is built for. The list has launch films, promos and explainers. It has data stories, kinetic typography and character animation. It also has illustrated shorts, loops and short videos for social feeds. Two example films come with the project. The first is a hand-drawn city. It is 30 seconds long, 720 by 720 pixels, at 60 frames a second. The second is a 3D character animation. It is 16 seconds long, 1920 by 1080 pixels, at 24 frames a second, with sound.
The video code can be plain HTML and CSS. It can also use Canvas, SVG, three.js, GSAP or Remotion. The skill picks the stack from the plan it writes first. The output is always a frame-by-frame render to MP4.
The gates and the checks
The skill works through a plan then a build. The SKILL.md file lays out thirteen principles and a per-film set of controls: the clock, continuity, the register of the voice, and the text policy. One rule from that file: the hook and any call to action must be on screen by 3 seconds. They are shown with the sound off, at measured size and contrast.
Two Python scripts hold the quality checks. The check_timeline.py script reads the plan. The qa_video.py script measures the finished render. The measurements cover text contrast, frame problems, and audio loudness, among other things.
The eval numbers it reports
The repository carries three test briefs with graders. The author ran these before and after the skill, and reports two rounds.
In round one the skill passed 28 of 28 checks. The baseline Claude run passed 25 of 28. Blind graders preferred the skill's output two times to one. The author says the skill run cost about twice the time and twice the tokens of the baseline.
In round two the current skill passed 30 of 32 checks. An earlier version of the skill passed 26 of 32. Blind graders chose the current version three times to zero. The author says the two runs cost about the same.
These are the author's own numbers, from the author's own graders, on three briefs the author chose. A buyer should treat them as the author's claim about the author's work. A real test is to run the three briefs against your own current workflow. Read the measurements. Then keep the version that wins.
What it needs
The skill needs ffmpeg, ffprobe and Python 3. The Python side needs the numpy and Pillow libraries. The browser side needs Node, Playwright and Chromium, or Remotion. One optional piece is faster-whisper, which checks that the narration matches the script. All of these are free and open source. No API key is needed beyond the Claude Code account.
Source
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.