Roomba and a bullhorn.
Why the hell not?

Making it work on paper first

I started in Photoshop, putting the two objects together before trying to generate anything. How big should the bullhorn be? Where would the cable go? Does it even need a cable? I wanted those details worked out so I had a reference for each shot.

Then I wrote the pitch, made a shot list and storyboarded it in Figma. The scene was an ordinary home office with hardwood floors and someone trying to work while the Roomba bumped around behind them.

Even if the concept is ridiculous, the execution can't be random. The storyboard gives the joke structure.

Generating it

The challenge was getting AI to follow the scene I had already designed. That meant being specific about the objects, their proportions, the room and the camera angle.

Even with my Photoshop composite as a reference, Gemini kept changing those details. The bullhorn’s scale was off, the objects changed or the room looked wrong. I ended up generating the room separately, then compositing the Roomba and bullhorn into it myself in Photoshop. That gave me more control over the stills I would use for the video.

I used Grok to animate those stills, which introduced another round of problems. Bullhorns blended into each other, and getting a usable shot took repeated attempts.

The first usable clip took two to four tries. Once the prompt and camera style worked, the rest was plug and play, and that's what kept every shot consistent.

The edit and the sound

With the clips ready, it was time to edit. Each was about ten seconds long, but I only needed the parts that helped tell the story. In After Effects, I played with the timing and switched between shots to keep things moving. The video needed to be punchy and visually entertaining, even with the sound off. Once I was happy with the sequence, I added animated type and an end card.

Then came the sound. I pulled sound effects from Artlist for all the little details: the Roomba bumping into furniture, the keyboard clicks, the hum of the motor. I picked a silly music track from its library to set the mood, then used its text-to-speech tool with a robot filter to get that flat, monotone voice shouting “strategic direction.”

Finally, I brought it into Premiere Pro and adjusted the levels so the sounds worked together. The voice needed to come through, the little noises needed room, and the music couldn’t overpower everything. The Roomba already looked annoying, but the sound made it feel like something you’d hate having in your office.

What I'd tell another designer

1. Make the idea work before you prompt. Sketch it, storyboard it. AI isn’t doing the work for you. The idea, direction, and final decisions are still yours.

2. Whatever you leave out, AI fills in. AI is just a dumb tool so be thorough. Pretend you’re describing the scene to someone over the phone who can’t see what you’re picturing. Be specific.

3. Expect to fix things by hand. Getting close isn’t the same as getting it right. I still had to composite images, cut around glitches and adjust the sound. Plan for that work.

This took Photoshop, Figma, Gemini, Grok, After Effects, Artlist and Premiere Pro. AI helped me make the ridiculous version I had in mind, but I still had to know what to keep, what to fix and when the joke landed. Human-in-the-loop is where the design experience and craftsmanship came in. Otherwise, I’d just have a folder full of weird Roomba videos.