Skip to content
Engineering
Jul 9, 2026 · 4 min

A Video Is a Storyboard First

How I generate a branded video: write the shot list, generate each shot in one house style, then bring it to life. Same system as the images. With an example.

Video
Generative AI
Build in Public

A 30-second video is not 30 seconds of footage. It's about six decisions, in order.


Everyone wants to "generate a video." So they type a paragraph into a model, get five seconds of pretty drift, and wonder why it doesn't feel like an ad.

Because the hard part of a video was never the pixels. It's the shot list. What happens, in what order, and why.

So I do the same thing I do for images: I write the system first, then generate against it.

Write the storyboard, not the prompt

One line per shot. Intent and framing, nothing more.

Shot 1: the command. A finger presses one key.
Shot 2: the horizon ignites. The brand motif.
Shot 3: the company assembles from light.
Shot 4: the system, one green core in a dark room.
Shot 5: the mark resolves from pixels.

That's the whole film, before a single frame exists. Five decisions I can argue about, reorder, or throw out for the price of editing a text file.

Generate each shot in one house style

Each shot is a spec, exactly like my image manifest: a concept plus the shared Pixl signature. Near-black, one green light, film grain. Generate the still, and it's the keyframe for that shot. Because every shot inherits the same signature, the five frames read as stills from one film instead of five random clips.

Then the keyframe goes into an image-to-video model, which gives it motion. Then I cut the shots together with sound.

Direction, not prompting

This is the part people skip. A video model, left alone, drifts: faces melt, text turns to soup, the camera wanders off to shoot something you didn't ask for.

The storyboard is the leash. Each shot is scoped so small that there's not much room to wander. I'm directing keyframes, not wishing at a paragraph.

The honest state

One good clip is easy now. A coherent 30 seconds, with continuity between shots, consistent branding, and audio that lands on the cut, is still real work. The still frames are reliable. The motion is the part that still needs a human saying "no, again."

No faked demo here. What's below is what the system actually produced.

The example

Here's a five-shot Pixl storyboard, built exactly this way: one line per shot, each frame generated in the house style. I ran the motion pass too. It isn't good enough to show yet, so you get the storyboard, not a bad clip.

The storyboard is the video. The rest is rendering.

Storyboard, 5 shots
Shot 01: The command
01The command
Shot 02: The horizon ignites
02The horizon ignites
Shot 03: The company assembles
03The company assembles
Shot 04: The system
04The system
Shot 05: The mark resolves
05The mark resolves

Five keyframes from one manifest, one house style. The motion pass isn't good enough to ship yet, so I'm showing the storyboard, not a bad clip. That's the point: the storyboard is the film. No stock, no fake footage.

Want to discuss multi-agent patterns?

We love talking about orchestration, AI workflows, and engineering challenges.