A video made with code is a function of time

Not generated video

Most people hear about AI video and picture a model hallucinating pixels. That is one way to make video, and it is not what happens here. When Claude Code builds a motion-graphics video, it writes code. Nothing is generated frame by frame by a model, so text stays sharp and brand colors stay exact.

The tradeoff is real. You get motion graphics: explainers, product walkthroughs, animated charts, titles. You don't get live-action footage or a person talking.

The mechanism

Claude writes a function. Give it a time t, and it returns one picture. To make a video, you ask that function for 30 to 60 pictures per second and play them in order. A headless browser, meaning a browser that runs without a window, draws each frame. A free tool called ffmpeg stitches the frames into an MP4. Sound works the same way, as another function of t.

The important property is determinism. The same t always gives the same picture. So any frame can be redrawn or inspected on its own, without rendering the whole video.

Why that makes it checkable

Claude can't watch a video. It can look at a single frame as an image. Because frames are reproducible, it can render a handful at the moments that matter, say the second a title appears or the moment two elements cross, notice a clipped word or an overlap, change the function and look again.

That loop is what separates this from asking for a video and hoping. The model is checking its own work against pictures, one frame at a time.

Storyboard first

The single habit that saves the most time is asking for a storyboard before any rendering. Five to seven still frames, one per beat, with a line of description each.

The reason is cost. Changing a still takes seconds. Changing a rendered video means regenerating it, and a long render can take many minutes. The worst case is the loop where you render, dislike something, render again, and lose an afternoon. A storyboard moves your decisions to the cheap end of the process, before any code is committed.

Where it breaks

It is still code, so complex scenes take time to render and can have bugs that only show at specific moments. Spot-checking frames catches most of them, not all. Quality also depends on the brief: a vague request produces a safe, generic look, and a style reference image helps a lot.

What to do tomorrow

If you have Claude Code on your computer, open an empty folder, run it there, and paste this. Claude Code will offer to install ffmpeg if it is missing:

Build me a 20-second motion-graphics video explaining [your topic], in [full screen or vertical for social]. Before you write any code, give me a storyboard of 5 to 7 still frames, one per beat, with one line each, and wait for my approval. After I approve, build the video in code, check individual frames at the key moments, fix anything wrong, and export an MP4.

Approve the stills first. That one step is most of the skill.