How we made an AI TV commercial in 24 hours (every prompt included)

One day, one character, six worlds, one finished spot. Watch the film, then copy every prompt that built it.

We gave ourselves 24 hours to make a complete TV commercial with AI. Not a mood reel, not a test. A finished spot: four cinematic scenes, one consistent character, a voiceover, an original score, and a final edit.

This is the film:

And this post is everything behind it. If you’ve just watched the video and came for the prompts, jump straight to them. If you’re new here, keep reading; the whole build is below, including the takes we rejected and why.

Every clip was generated in Magnific Spaces.

The 24-hour AI commercial workflow

The concept: every idea you’ve ever had already exists somewhere, and the only question was always how to make it real. So the commercial drops one character into six impossible worlds, from an ordinary bedroom into a dream that keeps escalating.

The constraint was the clock. No reshoots on location, no per-subject training, no post house. That forced a prompt structure that repeats across the whole film:

• Audio comes first. The opening line of each prompt sets the sound design before a single visual. No music, natural sound only. It anchors the mood before the model draws a frame.
• Timecoded beats. Each shot is broken into [00:00-00:04] style segments. You direct the edit inside the prompt, not after.
• Character references. @img1, @img2 and friends keep the same protagonist consistent across a bedroom, a desert, a dragon and a void.
• Stability instructions. “Background structure is static” tells the model what should not move. Naming the stillness is as important as naming the motion.
• One lighting source, one grade. Almost every scene runs on a single hard practical light and a teal-and-amber grade, 21:9 anamorphic, heavy film grain. That’s what makes six wildly different worlds feel like one film.

Now, scene by scene, in the order they appear in the spot.

Prompt 1: the sleeping zoom (vertical camera push-in)

The opening shot of the commercial. A bird’s-eye view of an ordinary bedroom, a slow push straight down into the character’s forehead, and a fade to black. Everything after this happens inside his head.

The prompt:

No music, natural sound only — starting with a silent ambient room hum, then a slow heartbeat starts and grows progressively louder and more intense as the camera pushes in.

[00:00-00:03]: Overhead, directly vertical bird's-eye view looking straight down from the ceiling, starting 20 feet higher up. An ordinary bedroom — a regular bed with rumpled sheets and a duvet, a nightstand, a lamp, clothes and everyday clutter, plain painted walls. Below, a young man lies flat on his back on the bed under the covers, staring straight up at the ceiling. The character's appearance is based on @img1 and @img2 as the character references.
[00:03-00:06]: He slowly closes his eyes tightly, squeezing them super tightly with extreme frustration visible across his face.
[00:06-00:08]: The camera smoothly and continuously zooms directly down in a straight vertical line, plunging straight into his forehead as the scene fades completely to black.

Background structure is static — the bed, bedroom walls, and floor do not move or deform. Only the camera tracks downward.

Cinematic POV, 21:9 anamorphic, masterful dramatic lighting. Single practical light source — a bedside lamp or window moonlight — casting soft natural shadows. Heavy film grain. Mood: claustrophobic, introspective, sinking inward.

Why it works: the camera movement is described as one continuous physical action (“plunging straight into his forehead”), and the prompt explicitly locks the room in place. Only the camera moves. That’s the whole shot.

Generate the video on Magnific

Prompt 2: the steampunk workshop (timecoded lighting change)

The first world of the dream. An extreme close-up in near-total darkness. Brass goggles, a dead vacuum tube, a single tap, and then light.

The prompt:

No music, natural sound only — faint click of a switch, followed by a sudden electrical hum.
Cinematic extreme close-up. 21:9, macro lens, very shallow depth of field, soft bokeh.

Timecoded beats:
[00:00-00:04]: The scene starts in absolute pitch darkness. A man's face in a steampunk workshop is barely silhouetted as he slides his brass steampunk goggles down over his eyes. He peers intensely at a dark, unlit glass vacuum tube bulb with cold wire filaments. Extreme texture of glass, brass gear machinery, and dust in near-blackness. At second 04, he reaches out and gently taps the unlit glass bulb with his finger.
[00:04-00:08]: The moment he taps the bulb at second 04, the filament inside the vacuum tube bulb suddenly clicks and surges with a warm orange light, fully turning on. The brilliant amber glow instantly rakes across one side of the man's face, catching his squinting eye and skin pores for the remainder of the clip.

Background structure is coherent and stable. Ambient elements (rising heat shimmer, dust particles drifting in the newly cast beam of light) have natural movement.
Cinematic steampunk anamorphic film, warm saturated teal-and-amber grade with heavy film grain.

Why it works: the entire clip is built around a single event at an exact second. “At second 04, he taps the bulb” appears in the first beat, and the second beat opens with the consequence. Cause and effect, timestamped.

Prompt 3: Lençóis Maranhenses (character consistency across cuts)

White dunes, jade lagoons, golden hour. A wide establishing shot cut against a close-up of the same character, same wardrobe, same light direction.

The prompt:

No music, natural sound of wind blowing over dunes and gentle water ripples.
A two-part cinematic sequence set in the vast sand dunes of Lençóis Maranhenses. The character's appearance is based on @img2 and @img1 as the character subject references. Throughout the entire sequence, he is wearing a simple plain brown t-shirt, no jacket.
Timecoded beats:

[00:00-00:04]: Extreme wide-angle shot of the man standing in the middle of immense white sand dunes, looking out at deep jade-green lagoons under a golden hour sun raking from frame-left, casting long deep shadows across the white sand. Camera is static, capturing his small silhouette against the landscape.

[00:04-00:08]: Cut to a close-up shot of the man facing the camera directly, looking straight into the POV with concern. He is dressed only in his brown shirt, no jacket. Behind him, the white sand and jade-green lagoons are softly blurred with a beautiful shallow depth of field. The golden hour light rakes from frame-left, carving deep shadow across the opposite side of his face. Camera is static.
Background structure is static — sand dunes, lagoons, and sky do not move or deform. Ambient elements (desert wind gently rippling the jade-green water and blowing fine sand grains) have natural subtle motion.
Cinematic anamorphic film, 21:9 aspect ratio. Heavy film grain and anamorphic lens character — subtle flares, bokeh streaks, shallow depth where it counts. Lighting practical and natural, single dominant hard source — golden hour light raking from frame-left — carving deep shadow on the opposite side, no artificial fill. Color grade teal-and-amber: deep desaturated shadows, warm highlight roll-off, rich living skin tones. Surreal, restrained, suspended. Masterful, deliberate cinematography.

Why it works: wardrobe continuity is spelled out twice (“brown t-shirt, no jacket” in the setup and in the second beat), and the light direction is locked to frame-left in both shots. Character consistency is a prompting decision, not luck.

Prompt 4: the dragon (directing scale and performance)

The dream at full escalation. A dragon over a hybrid megacity, half medieval European castles, half Chinese imperial palaces, all of it impossibly huge. And one very nervous rider.

The prompt:

No music, natural sound only — the roar of wind, the deep beat of enormous wings, and the man's subtle, adrenaline-fueled chuckle carrying over the air.

A two-part cinematic sequence. The character's appearance is based on @img1 and @img3 as the character subject references.

Timecoded beats:
[00:00-00:04]: Extreme wide aerial shot. A massive dragon soars through a vast sky, flying over a colossal megacity of gargantuan proportions giving a feeling of absolute gigantism. Below, half of the towering structures are massive European medieval castles and the other half are towering, spectacular Chinese-themed imperial palaces with huge sweeping curved roofs, all of them impossibly gigantic and poking up high into the cloud layer. The man rides high on the dragon's back, a tiny speck against the monumental hybrid architectural empire. He has a subtle, nervous laugh while holding on for dear life onto the dragon's horns.
[00:04-00:08]: Cut to a tight close-up shot — the man stays very close to the dragon, body pressed low, hands white-knuckled as he clings desperately to the dragon's horns. He has a subtle, adrenaline-fueled nervous laugh, wide-eyed with apprehension while his hair and brown t-shirt whip violently in the slipstream. Shallow depth of field blurs the colossal medieval and Chinese palace spires and clouds rushing past behind him.

Background structure is static and coherent — the dragon's body, wings, and the hybrid megacity do not deform or glitch. Ambient elements (wind-driven clouds, whipping hair and fabric, the slow beat of wings) have natural forceful motion.

Cinematic anamorphic film, 21:9 aspect ratio. Film grain, teal-and-amber color grade — deep desaturated shadows, warm highlight roll-off, rich skin tones. Lighting practical and natural: hard directional golden light breaking through clouds, deep shadow on the opposite side. Subtle lens flares. Texture everything.

Why it works: scale is directed with emotion, not numbers. “A tiny speck against the monumental hybrid architectural empire” tells the model how the shot should feel, and the performance note (“subtle, adrenaline-fueled nervous laugh”) repeats across both beats so the character stays in character.

The ElevenLabs voiceover prompt

A TV commercial needs a voice. Ours was generated with ElevenLabs v3 using the Dorian Malice voice, straight from the Space.

The script:

Every idea you've ever had. has always been real...
Somewhere.
It was always how do you make it real?
For most of human history, the tools just didn't exist.
Not anymore.
Now the idea in your head... and how you make it real... are the same distance apart.
Magnific. What you imagine. Made real.

Notice the punctuation. The odd periods and ellipses aren’t typos, they’re pacing. In ElevenLabs, punctuation is direction.

Generate Audio on Magnific

The AI music prompt (ElevenLabs Music v2)

The score came from ElevenLabs Music v2. Same principle as the video prompts: describe the arc second by second and name what you don’t want.

The prompt:

Inspiring cinematic orchestral masterwork capturing the moment of revelation as humanity tames fire. Begins with a deep, soulful solo cello painting a vast landscape of raw possibility. A lone piano note strikes at 10 seconds like a spark of genius, as warm strings rise in a hopeful, majestic swell. At second 19, a breathtaking, triumphant crescendo surges upward, exploding at second 20 into a soaring, golden climax of strings that evokes the life-giving warmth, illumination, and wonder of the first flame. Resolves in quiet, awe-struck reverence. No percussion, no vocals, no choir.

“A piano note strikes at 10 seconds.” “Exploding at second 20.” You can timecode music the same way you timecode shots. And the last line matters as much as the first: no percussion, no vocals, no choir.

What a 24-hour deadline rejects

Not everything made it into the commercial. Here’s what we cut and why, because the rejects teach more than the picks.

Lençóis Maranhenses, mid-day version (rejected)

Same location, same character, but shot under a bright mid-day sun with tropical blues and clean highlights. Technically fine. Tonally wrong. The bright commercial look broke the film’s single-source, golden-hour language, so we reshot it at golden hour with the light raking from frame-left. Compare the two prompts and you’ll see the only real change is the lighting paragraph. That one paragraph changes everything.

The underwater cathedral (rejected)

A sunken cathedral of monumental arches, a glowing shell, drifting silt in cold shafts of light. We loved it. It didn’t survive the edit; four worlds was the right rhythm for a spot this length, five was one too many. Sometimes the strongest cut is the one that hurts.

The end: the obsidian void

The closing beat of the film. A tiny figure on a floating island of cracked stone, colossal obsidian pillars rising into black, and a roar in the distance at exactly the 4-second mark. Yes, you can timecode a sound effect too.

Make your own AI commercial: the workflow

Six worlds, one character, one finished TV commercial, 24 hours. The structure that held it together:

1. Sound design in the first line of every prompt
2. Timecoded beats for every shot and every sound event
3. Character references (@img) plus wardrobe continuity written into the prompt
4. Explicit stability instructions for backgrounds
5. One light source, one grade, repeated across every scene
6. Voice and score generated in the same Space, timecoded like the shots

Watch the full build on YouTube, then open Spaces, copy any prompt, and make it yours.

Start making

FAQ

Can you really make a TV commercial with AI in 24 hours?

Yes. This spot went from concept to finished edit in one day: four cinematic scenes, a consistent character, an ElevenLabs voiceover and an original score, all generated in a single Magnific Space. The full process is in the video above, and every prompt is in this post.

How do you write a cinematic AI video prompt?

Lead with the sound design, break the shot into timecoded beats, name your camera behavior explicitly, and lock the look with one lighting source and one color grade. Every prompt in this post follows that structure. Copy one from the Space and start from there.

How do you keep a character consistent across AI video shots?

Attach the same reference images (@img1, @img2) in every prompt and write wardrobe and performance details into the text itself, in every beat where the character appears. Repetition inside the prompt is what holds continuity across cuts.

Can you timecode events inside an AI video prompt?

Yes. Beats like [00:04-00:08] control when actions happen, and you can pin single events to an exact second, from a bulb lighting up to a roar in the distance. The same applies to music: this film’s score places a piano note at second 10 and a crescendo at second 20.

What tools were used to make this AI commercial?

The video clips were generated in Magnific from a single Space, the voiceover with ElevenLabs v3 (Dorian Malice), and the score with ElevenLabs Music v2.