Video Director

A Claude Code skill that directs short videos from brief to finished render. Product launch films, explainers, brand films, designed demos, vertical social shorts, event and data videos, and live-action story ads, from about 6 seconds to 3 minutes, mixing motion graphics with photos, footage and generated clips.

You direct; Claude designs and builds. Nothing is animated until you have approved it as still frames, and every visual decision is a pick from 2-3 real options, never a reaction to one take.

Reviewing a storyboard: flipping through a scene's frames and pinning a note to one

Every example on this page is from a real project: a 52-second ad for Best Light, a Mac webcam app. All the people in it are AI-generated.

What's inside

Rules measured from real films, not guessed. Text sizes, positions, reading times and motion rules come from measuring 59 films across seven types. The visual ideas come from studying 34 films and ads. The lessons from building real films were checked against 56 more before they became rules.
8 video types, each with its own defaults, films worth studying, and a "default versus better" table: product launch, explainer, brand film, demo or tutorial, social short, event, recap or data, showcase, and live-action story.
57 animation styles, grouped by lead medium, each with an example film to look up.
A zoomable storyboard page you can pin notes to, a motion spec with a named move, timing and easing for every change, and an animated test before the full build.
Sound made by film type. Live-action films take their sound and spoken lines from the video model, checked with speech-to-text. Motion-design films get a few real recorded effects. Every film's music is fitted to the cut on its bar lines (scripts/fit_music.py). Every sound file is listed with its source and licence, and scripts/check_sources.py checks the list. real-sfx by Clipwing is included for placing real recorded effects from your own sound library.
Render checks in Python: frames on both sides of every cut, choppy motion, text size and reading time by role, sound levels, continuity between shots, delivery specs, and an optional motion watch by a video model.

How it works

1. Brief and script. It asks what the film is for, who it's for, what tools you have, and whether to study reference films or build from scratch.

2. Direction options. Before any storyboard, you pick a direction from 2-3 labeled still frames of the same moment.

Two direction options for the same moment, labeled A and B

3. The storyboard page. One HTML page you open in a browser. It shows the whole film in order, with each tile showing how its scene ends.

The storyboard overview: eight scene tiles with times and lines

It shows every change between scenes, side by side, with the move that connects them: the camera, what carries over, the timing and the easing.

The last frame of one scene beside the first frame of the next, with the cut described

Then it shows each scene as a card. Flip through its frames, zoom in, pin numbered notes to any spot, and copy or download every note to hand back to Claude.

A scene card with frame tabs, an Add pin button and the frame

4. Motion spec and sound plan. Every change on screen gets a named move, timing and easing before anything is animated. The sound is planned at the same time: where each sound comes from, and Suno prompts for the music.

5. Animated test, then the full film, with its sound, built with HyperFrames.

6. Render checks and polish rounds until you sign off.

Frames from the finished ad

Install

In Claude Code:

/plugin marketplace add SallyRS/video-director
/plugin install video-director@video-director

Or copy skills/video-director/ into ~/.claude/skills/.

Needs

Python 3 and ffmpeg for the render checks.
HyperFrames for the build.
Optional: yt-dlp to study reference films, a clip tool such as Higgsfield for generated footage, Codex for stills, and a Gemini API key for the motion watch (scripts/watch_cut.py). The skill works without them and says what it can't check.
Optional, for sound: Suno for music (a paid plan for anything commercial), whisper.cpp to check what generated clips say, aubio or librosa to find a track's bar lines, and your own sound library for the included real-sfx (it needs ffmpeg with the drawtext and rubberband filters, on macOS Homebrew's ffmpeg-full, and numpy for some scripts).

The only sound files included are real-sfx's interface pack (uisfx, CC0). Sounds from the free libraries aren't included: their licences don't allow sharing the files on their own. references/sound-stage.md lists where to get sounds and what each licence allows.

Credits

Built on and informed by other people's open work:

HyperFrames by HeyGen: the HTML-to-video engine every film is built with.
motion-video-kit by echris6 (MIT): its measured rules from professional launch films shaped how this skill carries objects across cuts and treats reference films as something to measure, not imitate.
saas-motion-kit by tugrawork-creator (MIT): its variety rules for shot length, easing and transitions in HyperFrames films.
motionmaxxing by Tejashmakwana (Apache-2.0): the finding that a gallery of example films makes every brand look the same, which is why this skill pairs a default with a better decision instead.
onetake by feitangyuan: its numbers on rhythm and what carries across a cut. Ideas only; no text or code is used.
real-sfx by Alexander Korbanov at Clipwing (MIT): included in skills/video-director/real-sfx/ for placing real recorded effects, copied unchanged from Clipwing/skills at commit edbefcb. That repository's README gives the scripts as MIT and has no licence file, so the folder carries the standard MIT text in LICENSE. Its interface sounds are uisfx by Yuki Capital (code MIT, sounds CC0), with their own licence files.
Jesse Eisenbart's 20-style showcase: the starting point for the animation style list.

The films named inside the skill are cited as examples to study. No footage, layouts or brand looks from them are included.

License

MIT