--- title: "One Desk Photo, Three AI Clips: Building a Cinematic Agent Command Center With Replit's Video Generator" description: "How one rear-view desk photo became a 24-second cinematic sequence of AI agent terminals using Replit's video generator: why the second clip changed direction instead of polishing the first, why a named palette and a 'do not include' list beat mood words, how chaining clips from the last frame keeps the scene continuous, the generator's limits, and three copy-pasteable prompts." date: "2026-09-23" type: "build" tags: [] legacy: true ---

Overview

I started with one ordinary photo: me at my desk, shot from behind. I ended with a 24-second sequence where the desk turns into a command center, a ring of code terminals comes online one at a time like independent AI agents, and the camera orbits the whole thing until every screen fires at once.

It took three clips from Replit's video generator. None of them is a single impressive prompt. What made it work was changing direction once, and then refusing to start over.

The sequence: a first concept clip animated from the original photo, a refined clip with a stricter palette and more screens, and a continuation clip that starts from the last frame of the refined one so the scene never resets.

The generator: image-to-video or text-to-video, clips of up to 8 seconds, 9:16 or 16:9, 720p or 1080p. Anything longer is built by chaining clips together.

Clip 1: the first concept

The first pass animated the original photo into an 8-second vertical 1080p scene:

It looked good, and it looked like every other AI "hacker at a desk" clip. The orange was carrying the whole mood, and with everything glowing there was no single thing to look at. There also weren't enough screens. The idea I wanted was one person running many agents, and a couple of floating panels don't say that.

Clip 2: less colour, more screens

So the second clip changed direction instead of just tweaking the first:

Why the "no" list mattered most

The line doing the most work is the negative one. "Futuristic tech interface" is a strong pull for these models. Left alone, they fill every screen with pie charts, radar sweeps, globe wireframes and cyan HUD brackets, because that's what the genre looks like in their training data. Asking for "code terminals" doesn't stop that. Naming the things you don't want does.

The palette works the same way. "Minimal colour" gets interpreted. "Black, charcoal, amber, white-gray only" gets followed. When you name four colours, a fifth one on screen is obviously wrong, and the model seems to treat it that way too.

Staggering the activation is what sells the agent idea. Twelve screens lighting up at once is a light switch. Twelve screens lighting up one after another reads as twelve separate things starting work.

Clip 3: continue, don't restart

Eight seconds isn't enough to get from "a desk" to "a command center that does something". The obvious next step is to go back to the original photo and prompt a longer story. That's the mistake. Every clip generated from the original photo starts from an empty desk, so stitching them together gives you a scene that keeps resetting.

Instead I took the final frame of clip 2 and used it as the starting image for clip 3. The new clip starts exactly where the old one ended: same screens, same lighting, same camera position. On a cut it plays as one continuous shot.

Pulling the last frame out of an MP4 is one ffmpeg command:

ffmpeg -sseof -1 -i clip-2.mp4 -update 1 -q:v 1 clip-2-last.png

-sseof -1 seeks to one second before the end, and -update 1 keeps overwriting the same output file, so the file you're left with is the very last frame. Use PNG. A JPEG of an already-compressed video frame loses more detail, and that softness gets carried into the next clip.

With the scene continuous, the continuation clip could move the story forward:

That last beat is the payoff the first two clips were building toward: many independent agents, one moment where they all act together. It only lands because the viewer watched those screens come online one by one in the previous shot.

Carry the camera move over too. Clip 2 ends mid-orbit, so clip 3 has to continue the orbit, not start a new one. If the direction or speed changes at the join, the cut shows even when the frames match.

The prompts

Here are the three directions written as reusable prompts. Swap in your own photo: anything shot from behind or over the shoulder at a desk works best, because the model doesn't have to invent a face.

Clip 1: first concept (image-to-video, original photo)

Animate this photo into an 8-second vertical 9:16 video at 1080p. Rear view of the person sitting at the desk, facing the monitor. The camera moves slowly from left to right behind them. Semi-transparent terminal-style interfaces float in the air around the monitor, drifting slightly with parallax as the camera moves. Cinematic orange-and-black tech grade: deep blacks, warm orange highlights, soft glow from the screens on the person's shoulders and the desk surface. The person stays mostly still, with small natural movement only.

Clip 2: agent workstation (image-to-video, original photo)

Animate this photo into an 8-second vertical 9:16 video at 1080p. Rear view of the person at the desk. The camera orbits clearly around behind them, a visible arc and not a subtle pan. Over the shot the desk gradually transforms into a futuristic command center. Multiple terminal windows appear around the person and activate ONE BY ONE, each like an independent AI agent starting its own task. Each screen shows only scrolling code and command-line text. Palette: deep black, charcoal, amber and white-gray ONLY. No other colours. Do NOT include: charts, graphs, dashboards, pie charts, radar sweeps, globes, maps, HUD brackets, neon cyan or blue, colourful UI elements. Code and terminal text only.

Clip 3: continuation (image-to-video, last frame of clip 2)

This image is the final frame of the previous shot. Continue directly from it: same person, same screens, same lighting, same palette. 8 seconds, vertical 9:16, 1080p. The camera continues the same orbit around the person in the same direction and at the same speed, rising slightly as it goes. The code screens begin responding to one another, as if the agents are passing work between them. The terminals reorganise into an arc around the person. The center monitor becomes the master terminal, brighter and more prominent than the rest. In the final beat every screen synchronises: they all execute together in one shared moment, a single pulse of amber across every terminal at once. Palette stays deep black, charcoal, amber and white-gray only. No charts, graphs or colourful HUD elements.

What the generator can do

This sequence only used one corner of it. The video generator works from either a text prompt or an existing image, and covers:

Constraint Current limit
Clip length Up to 8 seconds per generation
Aspect ratio 9:16 or 16:9
Resolution 720p or 1080p
Longer sequences Chain clips: extract the last frame, use it as the next start image
Sound Generated separately (music, sound effects, voiceover) and layered on afterwards

The 8-second limit sounds tight, but it forces a useful structure: each clip gets one job. Clip 1 sets the scene, clip 2 turns the screens on, clip 3 pays it off. A 24-second prompt trying to do all three would almost certainly have rushed at least one of them.

Sound is the obvious next layer here. Replit can generate music, sound effects and voiceover separately. A low drone under clip 2, a soft keyboard-click or relay sound as each terminal comes on, and one hit on the synchronised moment in clip 3 would do more for this sequence than any visual change.

Where it goes wrong

Recap

If you want more in this shape, there is ASCII Studio, a Replit build that turns images, 3D models and a webcam into ASCII, a Replit prompt for a scroll-scrubbed cinematic site, and a premium property site with scroll choreography.