AI Video Storyboard Workflow: Approve Stills Before Anything Moves, the Video Director Method

Most bad AI videos were bad before the first frame moved. The idea was wrong, or the layout was wrong, or the opening shot made a promise the rest never kept. Then someone spent a day of generations and a stack of credits animating it, and only saw the problem once it was playing.

Live-action film learned this a century ago, which is why storyboards exist. AI video mostly skipped the lesson, because generating a clip feels as cheap as drawing a frame. It is not. A clip costs money, time and, worst of all, attachment: once something moves, you start defending it.

A new free skill for Claude Code, Video Director, is built entirely around that problem. Its core rule, from the README: "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."

That second half is the part worth stealing, whether or not you ever install it.

What Video Director is

Video Director appeared on GitHub on November 6 under the MIT license. It is a skill, a set of written instructions plus helper scripts, that tells Claude Code how to direct short videos from about 6 seconds to 3 minutes: product launch films, explainers, brand films, tutorials, vertical social shorts, recaps and live-action story ads.

You stay the director. Claude designs and builds. The final film is assembled in HyperFrames, HeyGen's free, open-source engine (Apache 2.0) that turns web pages into video. Generated footage, still images and music come from outside tools you choose.

The README says its rules on text size, position, reading time and motion come from measuring 59 films across seven types. I cannot check the measurements, but the rules are written down where you can read and argue with them, and that is more than most "AI video" products give you.

The routine, stage by stage

The skill's own instructions set out about ten stages and stop for your approval at nearly every one. Here is the shape of it in plain language, leaving out an optional stage for studying reference films.

The brief. It asks direct questions and refuses to guess: what type of film, what shape of screen (16:9, 9:16, square, 4:5), narration or not, whether the film will be made of motion graphics, footage, screen recordings or a mix, whether people appear, and what must be real, like your logo, your font or your product's actual screens.

Style. For films with motion graphics, it shows a page of animation styles with a picture of each, from cut paper and clay stop-motion to kinetic type and pixel art, and suggests three or four.

Script. One job per beat. The skill treats your words as yours: it keeps your wording and order.

Direction options. This is the heart of it. Claude draws the opening one or two scenes two or three different ways, as real still frames, and puts them on one page side by side. The instructions insist the options differ in their central idea, "not three colors of one layout." Each option states its whole plan: the idea in one sentence with the product in it, the one lead medium, one drawing style, how it moves between ideas, and the end line.

Look decisions. One decision at a time, each with two or three options on the same frames: how the product is shown, then color, then the key scene's layout, then where text sits.

Storyboard. Every scene gets a start frame and an end frame, plus a middle frame when the change cannot be pictured from the two ends. It all lives on one web page you open in a browser. You can zoom in, pin numbered notes to any spot, and copy all your notes back to Claude, which has to answer every point in a numbered list before redrawing.

Motion spec and sound plan. Every change on screen gets a named move, a timing and an easing (how the motion speeds up and slows down) before anything is animated. Sound is planned at the same moment: which moments make a sound, where it comes from, and prompts for music.

Animated test. Only the two to four hardest scenes get built first.

Full build, then render checks. Python scripts look at frames on both sides of every cut, choppy motion, text size and reading time, sound levels and continuity.

Polish rounds, one numbered render per round of notes.

Why choosing beats reacting

Here is the mechanism that makes this work, and it is about people, not software.

When you see one version of something, your brain asks, "Is this good?" That is a weak question. You compare the frame against a fuzzy picture in your head, and the frame usually wins because it exists. When you see three versions of the same moment, the question changes to "Which of these is closest to what I mean?" Now you are comparing real things against each other, and the differences teach you what you actually want.

Video Director's instructions even plan for the awkward answers. If you say "1 looks nicer," Claude must say which option it took you to mean. If you like two but "it feels too X," it keeps both, fixes X on both, and shows them again. And my favorite rule in the whole file: if you dislike every option, it must "name what all the options shared before making new ones." That one sentence catches the most common failure in creative feedback, which is rejecting three versions of the same hidden assumption.

The second mechanism is cost. A still frame from an image model costs a fraction of a clip, renders in seconds and is easy to throw away. Video Director pushes every argument you would have had about a clip back to the stage where losing it costs almost nothing.

Put this into practice

You can run the full skill or borrow the routine. Both are worth doing.

To run the skill:

1. In Claude Code, type /plugin marketplace add SallyRS/video-director, then /plugin install video-director@video-director. Or copy the skills/video-director/ folder into ~/.claude/skills/.
2. Install Python 3, ffmpeg and HyperFrames, which the README lists as needed.
3. Start small. Ask for a 15-second vertical social short about something you know well.
4. At the direction stage, resist picking fast. Write one sentence on why each option is wrong before you choose the right one.
5. On the storyboard page, pin notes to exact spots instead of writing general comments. "This headline sits too low" pinned to the headline beats "the layout feels off."

To borrow the routine without the skill, in any AI video tool:

1. Write the brief as decisions, not vibes: screen shape, narration, who appears, what must be real.
2. Generate two or three stills of your opening shot that differ in idea, not color.
3. Choose one and write down why, with exact values. Video Director keeps a dated direction-notes.md file for exactly this, so later sessions read decisions instead of scrolling chat.
4. Make start and end stills for every scene before you animate any of it.
5. Animate the hardest scene first. If it fails, you have saved the budget for everything else.

Where it breaks

It is not one install. The skill itself is free, but the full chain needs Python, ffmpeg and HyperFrames, and on macOS the sound step wants a fuller ffmpeg build from Homebrew. Generated footage needs a separate clip tool; the README names Higgsfield. Stills can come from Codex. An optional motion check uses a Gemini key.

Commercial music costs money. Music is planned with Suno, and the README says plainly you need "a paid plan for anything commercial." Free sound libraries are not bundled because their licenses do not allow sharing the files alone, so you supply your own sounds.

It is built for designed films. The skill says it is not for recording a real product screen as a demo, captioning existing footage, or a lone logo sting. If your work is mostly cutting real footage, a traditional editor is still your tool.

Missing tools mean missing checks. Without the optional pieces, the README says the skill "says what it can't check." That is honest, and it also means some of its quality checks only exist once you install everything.

More stops means more of your time. Every stage waits for you. For a one-off meme that is overkill. For a client ad, that is the point.

The rules are someone's taste. They come from films the author studied. They are a strong default, not law, and you should overrule them when your film needs it.

The habit underneath

Strip away the software and Video Director teaches one habit: never judge a single take. Make the alternatives cheap, make them real, and choose.

That habit works in any tool you already pay for. Next time you start an AI video, stop before the first generation. Make three stills of the opening moment that disagree with each other about what the film is. Pick one, and write down why. You will spend ten minutes, and you may save the afternoon you would have spent animating the version you were about to throw away.

---

• Sources: Video Director on GitHub · Video Director SKILL.md · HyperFrames