carat

Prompt Gallery

/

Prompts Shared by Carat

Login
Sign Up

Prompts Shared by Carat

Copy-and-run AI video prompts written and tested by the Carat team. See finished clips such as a rooftop stunt jump and a kitten flood on a train platform, each with its full prompt.

2026.09.07

Make “Prompts Shared by Carat” exactly the way you wantCreate your own “Prompts Shared by Carat”

Just ask in plain language — Carat generates optimized images and videos for you.

Start now with 100 free credits daily

·

Rooftop Jump, Rubber Duck Bounce

Seedance 2.5· Video
[LOOK] Shot on a phone by one of the friends on the roof, tall vertical composition, 24mm wide angle, one continuous unbroken take. Real extreme-sports phone footage at night, never a game engine and never CG. Handheld throughout with a 4-6 cm sway that turns into hard running jolts while the operator runs, and heavy motion blur whenever the subject moves fast. Night phone exposure with its real failures: the streetlights and lit signage below clip out to white cores with visible bloom, the shadow areas on the rooftop crush to noisy black, and the auto-exposure visibly rebalances a stop when the camera swings from the dark roof deck down into the lit street canyon and again when it lifts back to the sky. Focus hunts once as the subject leaves the roof and once as they come back at the lens. Slightly soft, mildly compressed image with visible luminance noise. The sense of real height is carried by accurate perspective and scale on the towers, not by lens distortion. Everything holds its shape across the whole take — no flicker, no morphing, no warping of the buildings around the moving subject. [EXCLUDE] No anime or cartoon rendering, no game-engine or CG look, no plastic skin, no cheap greenscreen compositing, no exaggerated light effects, no slow motion or speed ramp. No cuts, no cutaways, no camera teleporting, no third-person or aerial drone angle. No subtitles and no on-screen text of any kind. No dialogue and no voice-over — ambient and action sound only. No BGM and no musical score at any point. [REFERENCES] @Image1 — the protagonist. Use their face, hair and full styling including their wardrobe, carried through unchanged for the entire take. Do not use its background, its lighting or its framing. [PREMISE] A phone video shot by a friend on a Seoul high-rise rooftop at night: the protagonist sprints across the helipad, launches off the parapet into the street canyon between two office towers, falls the full height of the buildings, hits the back of a colossal inflatable yellow rubber duck parked in the road, is thrown straight back up off its stretched rubber skin, tucks and rotates through the air, and lands on their feet on the same rooftop they jumped from. [BEATS] 0-2s: The protagonist stands at the back of a dark green rooftop helipad with a painted yellow circle, facing a low concrete parapet at the far edge. Three friends are gathered at the left of the frame in puffer jackets and hoodies, phones already up. They go — a hard sprint straight across the circle, stride opening up, body tilting forward, arms swinging loose. The operator runs after them from directly behind, and the frame slams up and down on every footfall. Beyond the parapet, the tops of the office towers on the far side of the street stand at eye level, and far behind them on the skyline stands one enormous solitary supertall tower — Lotte World Tower, a smooth cone tapering over a hundred and twenty storeys with gently curved sides, its crown split open and glowing, lit at night by continuous vertical lines of light running its whole height. It is unmistakably taller than everything around it and it never moves from that spot on the skyline. <the flat wind-noise of an exposed rooftop, the low drone of HVAC condenser fans behind the camera, two sets of running shoes on painted concrete, traffic far below>. 2-3s: They plant their last step and drive off the low parapet, clearing it completely. As they leave the roof their arms open out to both sides and they tip forward on purpose, diving into the gap between the two towers. The operator sprints the last few metres to the parapet and pitches the phone hard downward, keeping them centred in frame. <the wind noise jumps as the phone clears the parapet edge, one shoe scuff on the lip of the wall>. 3-5s: The camera is now close to straight down. They drop fast along the glass of the two towers, their body rotating from a forward tilt to face down, arms held out to hold their line. Under them the street opens up: eight lanes with the blue bus-only lane stripe running down the middle, orange taxis stopped at the lights, a four-way crossing with its zebra bars, street lamps, and the stacked lit signage down the sides of the buildings. Filling the whole width of the road sits a colossal inflatable yellow rubber duck — the ordinary bath-toy shape blown up to the size of a building, taut glossy yellow skin, a flat orange bill, two round black eyes, sitting upright and squarely on the asphalt with barriers and work vehicles ringed around its base. Ground-mounted uplights throw hard yellow bounce off its belly onto the buildings on both sides of the street. Their fall line runs dead at the broad curved top of its back, between the shoulders. <wind roars and keeps rising, the traffic and a delivery motorbike passing come up to meet them, a siren somewhere far off>. 5-6s: They hit the duck’s back squarely between the shoulders. The stretched rubber takes them and drives down into a deep bowl under their body, radial creases running out across the skin from the point of contact, the whole duck squashing wider at the sides as it loads — and it does not burst, tear or lose its shape. They are compressed into the bottom of that bowl for a moment, and then the skin snaps back and fires them straight up. Contact, dip, compression and release all read as four separate stages in that order. <a huge low drum-skin boom on contact, air shoving inside the body, rubber creaking under load, then a deep elastic thump as it releases>. 6-8s: The operator is still at the parapet; for a moment their own two shoes appear at the bottom edge of the frame against the drop, which sells the height. The protagonist rises fast up the same vertical line they fell down, and the camera lifts continuously from pointing straight down to pointing up at the sky, holding them as they climb back through the gap between the towers — and as the frame comes up, Lotte World Tower rises into it on the skyline behind them. Far below, the duck goes on wobbling and rippling from the hit, its skin still travelling. On the way up they pull their knees to their chest and wrap their arms around their shins into a tight tuck and turns over continuously and fast, the rotation natural and unbroken rather than locked to a set number. As they near roof height they open out: legs release, head comes up looking for the roof, feet swing down under them, arms spread to both sides to steady the line. <wind stripping past the phone mic, the city noise dropping away below, rubber still slapping and settling far underneath, the friends starting to shout>. 8-9s: They come at the camera fast, cross back over the parapet and onto the helipad. Both feet hit the green surface together and their knees and hips fold deep into a low squat to take the load — soles landing, legs compressing, body sinking, one small bounce back up. The landing carries real weight; they do not float down or land standing straight. The operator takes a fast half step backward to avoid being run into and keeps them framed from the front. 9-10s: They come up out of the squat and turn to the lens with their mouth open, roaring. The friends pile into frame from the left with their phones still up, grabbing their shoulders, throwing punches in the air, shouting over each other. The frame shakes as they crowd the camera, and the recording ends in that state. <wordless overlapping shouting and whooping, hands slapping shoulders, wind, the city underneath>. [CONSISTENCY] One unbroken take from 0s to 10s: no cuts, no transitions, no cutaways, no change of viewpoint, and the camera stays in the hands of the friend on the roof for the entire take — the operator is never on the ground and never becomes an aerial or third-person view. The stunt runs as one fixed chain and no link is skipped or reordered: rooftop sprint → clears the parapet → falls → hits the duck’s back → the rubber dips deep → it rebounds hard → they rise fast → tucks and rotates → opens out → both feet return to the same rooftop → deep-knee landing → celebration. The rebound comes only from the physical stretch and recoil of the duck’s inflated rubber skin. They do not fly, hover, get pulled by any cable, or move under their own power in the air, and they never start moving upward before they touch the duck. The skin visibly dips under them on contact. The duck never bursts, never deflates, never tears, never changes shape permanently and never disappears, and it stays in exactly the same spot in the road for the whole take. They fall and rise along the same single vertical line, and the rooftop they land on is the same rooftop they jumped from. They land on the broad curved back of the duck, never on its head, bill, tail or the barriers around it, never against a building, never against the parapet, and they never pass through the rubber skin. Lotte World Tower stays in one fixed position on the skyline in every beat where the sky is visible, always far behind the two office towers, always the tallest thing in frame, and it is never approached, never entered and never jumped from. Their body stays anatomically correct through the rotation — no twisted limbs, no extra arms or legs, no reversed joints, no melting, no duplicate of them anywhere in frame. Nobody is hurt, there is no fall, no injury and no blood. The buildings stay straight, the vehicles stay on the road, the phones in the friends’ hands keep their shape, and the friends stay out of the lens until 9s. The protagonist is exactly as in @Image1 for the whole take and their styling never drifts or changes. Signage on the buildings and along the street stays out of focus and unreadable at all times.

Kitten Flood on a Train Platform

Seedance 2.5· Video
[LOOK] Shot on a handheld smartphone in a tall vertical composition, held at chest height, one continuous unbroken take — found footage, not cinema. No stabilization at any point: the frame breathes and drifts 3-5 cm constantly and jerks hard whenever the operator reacts. Fast pans smear and lean with rolling-shutter skew. Autofocus hunts once when the far distance changes and again when something crosses close to the lens. Auto-exposure visibly chases the light, dimming and re-brightening about a stop as the camera tilts between the bright glazed canopy overhead and the shaded platform floor. Bright midday daylight pours through the canopy glazing in hard slanted patches, with deep shade under the steel ribs; that daylight is the only light source. Slightly soft, mildly compressed image, visible noise in the shadows, flat and ungraded — exactly the quality of a phone recording somebody posted without touching it. Everything holds its shape across the whole take: no flicker, no morphing, no warping in the background around the moving mass. [EXCLUDE] No cinematic grading or colour styling. No slow motion, no speed ramp. No stabilised or gimbal-smooth movement. No drone, crane or second camera. No cuts, no cutaways, no transitions. No subtitles and no on-screen text of any kind. No BGM and no musical score at any point — ambient and action sound only. [REFERENCES] @Image1 — the location. Use its architecture, platform layout, canopy structure, track bed, catenary masts, signage placement, seating and lighting fixtures exactly. Do not use its people, its wardrobe, its framing or its time of day. [PREMISE] A first-person phone video from a mainline railway platform in Seoul at midday: a deep unexplained boom rolls up the line and startles the whole platform, and only afterwards does the thing that caused it arrive — a flood of tens of thousands of kittens pouring out of the far distance, down the track bed and over the platform, until it swallows the camera. A real-looking, unedited, single-take clip. [BEATS] 0-1s: The camera looks straight down the platform along the yellow safety line, which runs from the lower left of the frame to a vanishing point far away where the rails and the catenary masts converge. Passengers stand in a still, patient line near the edge — a man in a dark suit with a backpack, an older woman with one hand resting on the handle of a rolling suitcase, a couple in outdoor hiking jackets holding paper coffee cups. Nobody is looking at anything. <hollow room tone under a covered platform with a long reverb tail; one suitcase wheel rattling past close to the lens; a two-note announcement chime>. A station announcement in standard Korean, a calm recorded female voice over the platform PA with heavy reverb, is already halfway through a sentence: {잠시 후 부산행 열차가—} Half a second in, a single deep boom arrives from far up the line — almost no high frequency in it at all, more felt through the floor than heard, the sound of something very large landing a long way away. The camera sinks about fifteen centimetres and tips down four or five degrees as the operator’s hands drop, so the whole frame slides upward. Everyone on the platform flinches at once: heads snap toward the far end of the track, the man in the suit half-crouches, the older woman closes her fist on the suitcase handle, one of the hiking couple’s coffee slops over the rim. 1-2s: The boom’s tail rings out through the concrete for a long moment and dies. Nothing has happened. The recorded PA voice, completely untouched by it, finishes its sentence in the same calm tone in standard Korean: {—들어오겠습니다. 안전선 밖으로 한 걸음 물러나 주시기 바랍니다.} The line of passengers has broken up: people are turned toward the far end now, several have taken a half step back from the edge, and a low murmur runs along the platform as they ask each other what that was. <the room tone comes back thinner than before; scattered footsteps; overlapping low Korean voices, none of them clear enough to make out>. The camera comes back up only partway and never returns to where it started; it drifts along the track toward the vanishing point, hunting for whatever made the noise. 2-3s: At the vanishing point, exactly where a train would appear, a small brown smudge forms on the track bed. It is too far and too small to read as anything. <a low rumble begins far down the corridor and rises — at first indistinguishable from an approaching train, but with no rail squeal and no wind push in front of it; underneath it the platform concrete starts to buzz>. A man’s voice behind the camera, in standard Korean, uncertain: {저거 뭐야?} At 3s the operator reacts: a fast, ugly whip of the camera to the left and back, one beat of pure smear where nothing in the frame is legible, then the frame snaps back onto the track — over-rotated, correcting. The person holding the phone, quietly, in standard Korean, under their breath: {어?} 3-4s: The smudge now fills the far end of the corridor from the track bed up to the catenary wire, spilling sideways onto both platforms, and it is coming. <the rumble breaks apart into texture — tens of thousands of tiny paws hitting concrete and gravel, dense and fast like heavy rain drumming on a tin roof, with a dry scratching layer of claws over it; threaded through that, thousands of thin high meows, so many and so overlapped that they fuse into one continuous ringing tone; the whole thing doubled and smeared by the canopy reverb>. The murmur on the platform stops dead. 4-5s: The mass resolves into what it is — kittens. Tens of thousands of them, ginger, calico, tabby, grey, white-socked, pouring down the track bed and surging up over the platform edge in one unbroken running carpet. The platform breaks. The man in the suit takes three fast steps backward and puts his back against a pillar. One of the hiking couple drops her coffee cup and it bursts on the concrete. A woman further down screams. <the paw-rain doubles in volume, a coffee cup bursting, running footsteps on concrete>. A woman’s voice in standard Korean, high and disbelieving: {어머어머어머!} A man’s voice in standard Korean, shouting: {뭐야 저거!} 5-6s: The carpet fills the entire width of the frame and closes on the camera fast. The older woman yanks her suitcase in against her legs and it goes over on its side. The camera lurches as the operator backs up a step and the horizon tips. <the sound thickens into a wall — the paw-rain, the claw-scratch and the fused meowing now equally loud, the suitcase hitting the ground, a bag strap slapping>. The person holding the phone, in standard Korean, fast and half-laughing: {뭐야, 뭐야 뭐야—} 6-8s: The leading edge reaches the camera. The frame tips down and blue jeans and worn white sneakers enter the bottom of the frame; kittens break around the shins and stream past on both sides, a few brushing right across the lens and blurring it, one hooking a claw briefly on the denim. The camera is knocked sideways and does not fully recover. <sound peaks here — a dense roar of meowing and paws that briefly overloads the phone microphone and distorts into clipping, with soft fur-on-fabric friction close to the mic and small bodies thudding against the shins>. The person holding the phone lets out a sharp involuntary noise and then starts laughing, breathless, unable to stop, in standard Korean: {아 씨— 야, 야 이거 뭐야, 아하하하—} 8-10s: The camera stays pointed down. The platform floor has disappeared entirely under a moving carpet of kittens flowing past the feet like water. Higher in the frame, two people do nothing at all. The man in the suit has already gone back to looking at his phone and does not look up once. The older woman glances down, lifts her fallen suitcase upright and then five centimetres clear of the ground so the kittens can pass underneath it, and holds it there. Both of them keep their mouths closed and say nothing. Only the operator is still audible, laughing more quietly now, running out of breath. <the wall of meowing holds steady and unchanging; under it, the canopy reverb makes it sound like the noise is coming from every direction>. The same calm recorded female voice comes over the PA in standard Korean, unchanged in tone: {열차가 도착하고 있습니다. 승객 여러분께서는 안전선 밖으로 물러나 주시기 바랍니다.} 10-12s: The camera loses it. Kittens climb the operator; the horizon swings and tilts and never comes back to level. Focus fails and hunts and does not find anything. Fur fills the frame. <as the fur covers the microphone the entire soundscape goes muffled and dull in an instant — the roar drops to a thick blanketed rumble with all the high frequencies gone, and against that muffling, two or three individual meows land very close and very clear, right on the microphone>. The person holding the phone, muffled through fur, in standard Korean: {야 잠깐, 잠깐—} The last second is out-of-focus ginger fur pressed flat against the lens with one dark ear crossing the top of the frame. The recording does not resolve — it simply stops. [CONSISTENCY] One unbroken take from 0s to 12s: no cuts, no transitions, no cutaways, and the camera never leaves the operator’s hands. The same platform throughout, the same time of day, the same daylight direction. The boom in 0-1s happens once and never repeats, and nothing is ever shown or explained to be its cause. After the camera sinks in 0-1s it never returns to its opening height for the rest of the take. Kittens only — no adult cats — all roughly the same young size, in ordinary mixed domestic coat patterns. The flow never reverses direction: it always runs from the far vanishing point toward the camera. The man in the suit, the older woman with the suitcase and the couple in hiking jackets are the same four people throughout, in the same clothes. From 8s onward the man in the suit and the older woman are the only bystanders in frame and neither of them speaks, screams or runs — their mouths stay closed. Every spoken line in this video is in standard Korean with no regional accent. The operator is heard but never seen above the waist — no face, no arms, no hands and no phone appear in frame at any point. The sound never cuts to silence and never crossfades; it only changes because of what is physically happening to the microphone. Any Korean signage in the background stays completely out of focus and unreadable.

Kitten Flood on a Train Platform

Seedance 2.0· Video
[LOOK] Shot on a handheld smartphone in a tall vertical composition, held at chest height, one continuous unbroken take — found footage, not cinema. No stabilization at any point: the frame breathes and drifts 3-5 cm constantly and jerks hard whenever the operator reacts. Fast pans smear and lean with rolling-shutter skew. Autofocus hunts once when the far distance changes and again when something crosses close to the lens. Auto-exposure visibly chases the light, dimming and re-brightening about a stop as the camera tilts between the bright glazed canopy overhead and the shaded platform floor. Bright midday daylight pours through the canopy glazing in hard slanted patches, with deep shade under the steel ribs; that daylight is the only light source. Slightly soft, mildly compressed image, visible noise in the shadows, flat and ungraded — exactly the quality of a phone recording somebody posted without touching it. Everything holds its shape across the whole take: no flicker, no morphing, no warping in the background around the moving mass. [EXCLUDE] No cinematic grading or colour styling. No slow motion, no speed ramp. No stabilised or gimbal-smooth movement. No drone, crane or second camera. No cuts, no cutaways, no transitions. No subtitles and no on-screen text of any kind. No BGM and no musical score at any point — ambient and action sound only. [REFERENCES] @Image1 — the location. Use its architecture, platform layout, canopy structure, track bed, catenary masts, signage placement, seating and lighting fixtures exactly. Do not use its people, its wardrobe, its framing or its time of day. [PREMISE] A first-person phone video from a mainline railway platform in Seoul at midday: a deep unexplained boom rolls up the line and startles the whole platform, and only afterwards does the thing that caused it arrive — a flood of tens of thousands of kittens pouring out of the far distance, down the track bed and over the platform, until it swallows the camera. A real-looking, unedited, single-take clip. [BEATS] 0-1s: The camera looks straight down the platform along the yellow safety line, which runs from the lower left of the frame to a vanishing point far away where the rails and the catenary masts converge. Passengers stand in a still, patient line near the edge — a man in a dark suit with a backpack, an older woman with one hand resting on the handle of a rolling suitcase, a couple in outdoor hiking jackets holding paper coffee cups. Nobody is looking at anything. <hollow room tone under a covered platform with a long reverb tail; one suitcase wheel rattling past close to the lens; a two-note announcement chime>. A station announcement in standard Korean, a calm recorded female voice over the platform PA with heavy reverb, is already halfway through a sentence: {잠시 후 부산행 열차가—} Half a second in, a single deep boom arrives from far up the line — almost no high frequency in it at all, more felt through the floor than heard, the sound of something very large landing a long way away. The camera sinks about fifteen centimetres and tips down four or five degrees as the operator’s hands drop, so the whole frame slides upward. Everyone on the platform flinches at once: heads snap toward the far end of the track, the man in the suit half-crouches, the older woman closes her fist on the suitcase handle, one of the hiking couple’s coffee slops over the rim. 1-2s: The boom’s tail rings out through the concrete for a long moment and dies. Nothing has happened. The recorded PA voice, completely untouched by it, finishes its sentence in the same calm tone in standard Korean: {—들어오겠습니다. 안전선 밖으로 한 걸음 물러나 주시기 바랍니다.} The line of passengers has broken up: people are turned toward the far end now, several have taken a half step back from the edge, and a low murmur runs along the platform as they ask each other what that was. <the room tone comes back thinner than before; scattered footsteps; overlapping low Korean voices, none of them clear enough to make out>. The camera comes back up only partway and never returns to where it started; it drifts along the track toward the vanishing point, hunting for whatever made the noise. 2-3s: At the vanishing point, exactly where a train would appear, a small brown smudge forms on the track bed. It is too far and too small to read as anything. <a low rumble begins far down the corridor and rises — at first indistinguishable from an approaching train, but with no rail squeal and no wind push in front of it; underneath it the platform concrete starts to buzz>. A man’s voice behind the camera, in standard Korean, uncertain: {저거 뭐야?} At 3s the operator reacts: a fast, ugly whip of the camera to the left and back, one beat of pure smear where nothing in the frame is legible, then the frame snaps back onto the track — over-rotated, correcting. The person holding the phone, quietly, in standard Korean, under their breath: {어?} 3-4s: The smudge now fills the far end of the corridor from the track bed up to the catenary wire, spilling sideways onto both platforms, and it is coming. <the rumble breaks apart into texture — tens of thousands of tiny paws hitting concrete and gravel, dense and fast like heavy rain drumming on a tin roof, with a dry scratching layer of claws over it; threaded through that, thousands of thin high meows, so many and so overlapped that they fuse into one continuous ringing tone; the whole thing doubled and smeared by the canopy reverb>. The murmur on the platform stops dead. 4-5s: The mass resolves into what it is — kittens. Tens of thousands of them, ginger, calico, tabby, grey, white-socked, pouring down the track bed and surging up over the platform edge in one unbroken running carpet. The platform breaks. The man in the suit takes three fast steps backward and puts his back against a pillar. One of the hiking couple drops her coffee cup and it bursts on the concrete. A woman further down screams. <the paw-rain doubles in volume, a coffee cup bursting, running footsteps on concrete>. A woman’s voice in standard Korean, high and disbelieving: {어머어머어머!} A man’s voice in standard Korean, shouting: {뭐야 저거!} 5-6s: The carpet fills the entire width of the frame and closes on the camera fast. The older woman yanks her suitcase in against her legs and it goes over on its side. The camera lurches as the operator backs up a step and the horizon tips. <the sound thickens into a wall — the paw-rain, the claw-scratch and the fused meowing now equally loud, the suitcase hitting the ground, a bag strap slapping>. The person holding the phone, in standard Korean, fast and half-laughing: {뭐야, 뭐야 뭐야—} 6-8s: The leading edge reaches the camera. The frame tips down and blue jeans and worn white sneakers enter the bottom of the frame; kittens break around the shins and stream past on both sides, a few brushing right across the lens and blurring it, one hooking a claw briefly on the denim. The camera is knocked sideways and does not fully recover. <sound peaks here — a dense roar of meowing and paws that briefly overloads the phone microphone and distorts into clipping, with soft fur-on-fabric friction close to the mic and small bodies thudding against the shins>. The person holding the phone lets out a sharp involuntary noise and then starts laughing, breathless, unable to stop, in standard Korean: {아 씨— 야, 야 이거 뭐야, 아하하하—} 8-10s: The camera stays pointed down. The platform floor has disappeared entirely under a moving carpet of kittens flowing past the feet like water. Higher in the frame, two people do nothing at all. The man in the suit has already gone back to looking at his phone and does not look up once. The older woman glances down, lifts her fallen suitcase upright and then five centimetres clear of the ground so the kittens can pass underneath it, and holds it there. Both of them keep their mouths closed and say nothing. Only the operator is still audible, laughing more quietly now, running out of breath. <the wall of meowing holds steady and unchanging; under it, the canopy reverb makes it sound like the noise is coming from every direction>. The same calm recorded female voice comes over the PA in standard Korean, unchanged in tone: {열차가 도착하고 있습니다. 승객 여러분께서는 안전선 밖으로 물러나 주시기 바랍니다.} 10-12s: The camera loses it. Kittens climb the operator; the horizon swings and tilts and never comes back to level. Focus fails and hunts and does not find anything. Fur fills the frame. <as the fur covers the microphone the entire soundscape goes muffled and dull in an instant — the roar drops to a thick blanketed rumble with all the high frequencies gone, and against that muffling, two or three individual meows land very close and very clear, right on the microphone>. The person holding the phone, muffled through fur, in standard Korean: {야 잠깐, 잠깐—} The last second is out-of-focus ginger fur pressed flat against the lens with one dark ear crossing the top of the frame. The recording does not resolve — it simply stops. [CONSISTENCY] One unbroken take from 0s to 12s: no cuts, no transitions, no cutaways, and the camera never leaves the operator’s hands. The same platform throughout, the same time of day, the same daylight direction. The boom in 0-1s happens once and never repeats, and nothing is ever shown or explained to be its cause. After the camera sinks in 0-1s it never returns to its opening height for the rest of the take. Kittens only — no adult cats — all roughly the same young size, in ordinary mixed domestic coat patterns. The flow never reverses direction: it always runs from the far vanishing point toward the camera. The man in the suit, the older woman with the suitcase and the couple in hiking jackets are the same four people throughout, in the same clothes. From 8s onward the man in the suit and the older woman are the only bystanders in frame and neither of them speaks, screams or runs — their mouths stay closed. Every spoken line in this video is in standard Korean with no regional accent. The operator is heard but never seen above the waist — no face, no arms, no hands and no phone appear in frame at any point. The sound never cuts to silence and never crossfades; it only changes because of what is physically happening to the microphone. Any Korean signage in the background stays completely out of focus and unreadable.

Rooftop Jump, Rubber Duck Bounce

Seedance 2.0· Video
[LOOK] Shot on a phone by one of the friends on the roof, tall vertical composition, 24mm wide angle, one continuous unbroken take. Real extreme-sports phone footage at night, never a game engine and never CG. Handheld throughout with a 4-6 cm sway that turns into hard running jolts while the operator runs, and heavy motion blur whenever the subject moves fast. Night phone exposure with its real failures: the streetlights and lit signage below clip out to white cores with visible bloom, the shadow areas on the rooftop crush to noisy black, and the auto-exposure visibly rebalances a stop when the camera swings from the dark roof deck down into the lit street canyon and again when it lifts back to the sky. Focus hunts once as the subject leaves the roof and once as they come back at the lens. Slightly soft, mildly compressed image with visible luminance noise. The sense of real height is carried by accurate perspective and scale on the towers, not by lens distortion. Everything holds its shape across the whole take — no flicker, no morphing, no warping of the buildings around the moving subject. [EXCLUDE] No anime or cartoon rendering, no game-engine or CG look, no plastic skin, no cheap greenscreen compositing, no exaggerated light effects, no slow motion or speed ramp. No cuts, no cutaways, no camera teleporting, no third-person or aerial drone angle. No subtitles and no on-screen text of any kind. No dialogue and no voice-over — ambient and action sound only. No BGM and no musical score at any point. [REFERENCES] @Image1 — the protagonist. Use their face, hair and full styling including their wardrobe, carried through unchanged for the entire take. Do not use its background, its lighting or its framing. [PREMISE] A phone video shot by a friend on a Seoul high-rise rooftop at night: the protagonist sprints across the helipad, launches off the parapet into the street canyon between two office towers, falls the full height of the buildings, hits the back of a colossal inflatable yellow rubber duck parked in the road, is thrown straight back up off its stretched rubber skin, tucks and rotates through the air, and lands on their feet on the same rooftop they jumped from. [BEATS] 0-2s: The protagonist stands at the back of a dark green rooftop helipad with a painted yellow circle, facing a low concrete parapet at the far edge. Three friends are gathered at the left of the frame in puffer jackets and hoodies, phones already up. They go — a hard sprint straight across the circle, stride opening up, body tilting forward, arms swinging loose. The operator runs after them from directly behind, and the frame slams up and down on every footfall. Beyond the parapet, the tops of the office towers on the far side of the street stand at eye level, and far behind them on the skyline stands one enormous solitary supertall tower — Lotte World Tower, a smooth cone tapering over a hundred and twenty storeys with gently curved sides, its crown split open and glowing, lit at night by continuous vertical lines of light running its whole height. It is unmistakably taller than everything around it and it never moves from that spot on the skyline. <the flat wind-noise of an exposed rooftop, the low drone of HVAC condenser fans behind the camera, two sets of running shoes on painted concrete, traffic far below>. 2-3s: They plant their last step and drive off the low parapet, clearing it completely. As they leave the roof their arms open out to both sides and they tip forward on purpose, diving into the gap between the two towers. The operator sprints the last few metres to the parapet and pitches the phone hard downward, keeping them centred in frame. <the wind noise jumps as the phone clears the parapet edge, one shoe scuff on the lip of the wall>. 3-5s: The camera is now close to straight down. They drop fast along the glass of the two towers, their body rotating from a forward tilt to face down, arms held out to hold their line. Under them the street opens up: eight lanes with the blue bus-only lane stripe running down the middle, orange taxis stopped at the lights, a four-way crossing with its zebra bars, street lamps, and the stacked lit signage down the sides of the buildings. Filling the whole width of the road sits a colossal inflatable yellow rubber duck — the ordinary bath-toy shape blown up to the size of a building, taut glossy yellow skin, a flat orange bill, two round black eyes, sitting upright and squarely on the asphalt with barriers and work vehicles ringed around its base. Ground-mounted uplights throw hard yellow bounce off its belly onto the buildings on both sides of the street. Their fall line runs dead at the broad curved top of its back, between the shoulders. <wind roars and keeps rising, the traffic and a delivery motorbike passing come up to meet them, a siren somewhere far off>. 5-6s: They hit the duck’s back squarely between the shoulders. The stretched rubber takes them and drives down into a deep bowl under their body, radial creases running out across the skin from the point of contact, the whole duck squashing wider at the sides as it loads — and it does not burst, tear or lose its shape. They are compressed into the bottom of that bowl for a moment, and then the skin snaps back and fires them straight up. Contact, dip, compression and release all read as four separate stages in that order. <a huge low drum-skin boom on contact, air shoving inside the body, rubber creaking under load, then a deep elastic thump as it releases>. 6-8s: The operator is still at the parapet; for a moment their own two shoes appear at the bottom edge of the frame against the drop, which sells the height. The protagonist rises fast up the same vertical line they fell down, and the camera lifts continuously from pointing straight down to pointing up at the sky, holding them as they climb back through the gap between the towers — and as the frame comes up, Lotte World Tower rises into it on the skyline behind them. Far below, the duck goes on wobbling and rippling from the hit, its skin still travelling. On the way up they pull their knees to their chest and wrap their arms around their shins into a tight tuck and turns over continuously and fast, the rotation natural and unbroken rather than locked to a set number. As they near roof height they open out: legs release, head comes up looking for the roof, feet swing down under them, arms spread to both sides to steady the line. <wind stripping past the phone mic, the city noise dropping away below, rubber still slapping and settling far underneath, the friends starting to shout>. 8-9s: They come at the camera fast, cross back over the parapet and onto the helipad. Both feet hit the green surface together and their knees and hips fold deep into a low squat to take the load — soles landing, legs compressing, body sinking, one small bounce back up. The landing carries real weight; they do not float down or land standing straight. The operator takes a fast half step backward to avoid being run into and keeps them framed from the front. 9-10s: They come up out of the squat and turn to the lens with their mouth open, roaring. The friends pile into frame from the left with their phones still up, grabbing their shoulders, throwing punches in the air, shouting over each other. The frame shakes as they crowd the camera, and the recording ends in that state. <wordless overlapping shouting and whooping, hands slapping shoulders, wind, the city underneath>. [CONSISTENCY] One unbroken take from 0s to 10s: no cuts, no transitions, no cutaways, no change of viewpoint, and the camera stays in the hands of the friend on the roof for the entire take — the operator is never on the ground and never becomes an aerial or third-person view. The stunt runs as one fixed chain and no link is skipped or reordered: rooftop sprint → clears the parapet → falls → hits the duck’s back → the rubber dips deep → it rebounds hard → they rise fast → tucks and rotates → opens out → both feet return to the same rooftop → deep-knee landing → celebration. The rebound comes only from the physical stretch and recoil of the duck’s inflated rubber skin. They do not fly, hover, get pulled by any cable, or move under their own power in the air, and they never start moving upward before they touch the duck. The skin visibly dips under them on contact. The duck never bursts, never deflates, never tears, never changes shape permanently and never disappears, and it stays in exactly the same spot in the road for the whole take. They fall and rise along the same single vertical line, and the rooftop they land on is the same rooftop they jumped from. They land on the broad curved back of the duck, never on its head, bill, tail or the barriers around it, never against a building, never against the parapet, and they never pass through the rubber skin. Lotte World Tower stays in one fixed position on the skyline in every beat where the sky is visible, always far behind the two office towers, always the tallest thing in frame, and it is never approached, never entered and never jumped from. Their body stays anatomically correct through the rotation — no twisted limbs, no extra arms or legs, no reversed joints, no melting, no duplicate of them anywhere in frame. Nobody is hurt, there is no fall, no injury and no blood. The buildings stay straight, the vehicles stay on the road, the phones in the friends’ hands keep their shape, and the friends stay out of the lens until 9s. The protagonist is exactly as in @Image1 for the whole take and their styling never drifts or changes. Signage on the buildings and along the street stays out of focus and unreadable at all times.

What Are Prompts Shared by Carat?

This tag collects AI video prompts the Carat team wrote and actually ran, published exactly as they were used. Each card pairs the finished clip with the full prompt text, so you can read a line and see the shot it produced. They are meant to be copied as-is, or edited block by block.

Where You'll Use It

  • Learning structured long-form prompts (LOOK, EXCLUDE, BEATS) and second-by-second staging
  • Locking a character across a whole take from a single reference image
  • Specifying "shot on a real phone" texture: handheld sway, autofocus hunting, exposure shifts
  • Comparing the same prompt on Seedance 2.0 and Seedance 2.5
  • Writing ambient and impact sound into the prompt instead of accepting a silent clip

Why This Tag Is Useful

Only prompts with a verified result get posted here, so you are not re-running experiments to find out which phrasing produces the shot you want. The prompts are written in separate blocks — look, exclusions, timed beats, and consistency rules — which makes it easy to swap one block and keep the rest.

Tips for Crafting Prompts

  1. Keep the look and the exclusions in separate blocks. Mixed together, the model reads your bans as directions.
  2. Pin the action to timed beats so steps never reorder or drop out.
  3. To lock a character, name a reference image and make sure the pronouns and age in the text do not contradict it.
  4. Restate any physical chain that must happen in order (contact, compression, rebound) in a final consistency block.

Open a card to copy the full prompt. Cards that need a reference image are marked, so have a photo of the person or location ready before you run them.

Frequently Asked Questions

Related Prompts

Viral Video

19 prompts

Handheld Video

42 prompts

Surreal Video

59 prompts

Train Station

Train Station

5 prompts

Carat Prompt Gallery

Create trending AI prompts instantly

Prompts

  • Logo
  • PPT Design
  • Thumbnail
  • Poster
  • Web Novel Cover
  • ID Photo
  • Caricature
  • Product Detail Page
  • Banner

Models

  • GPT Image 2
  • Seedance 2.0
  • NanoBanana Pro
  • NanoBanana 2

Carat

  • Get started
  • About
  • Blog
  • Careers
  • Contact
  • Partnership
  • Terms
  • Privacy

© 2026 Paradot.Inc All rights reserved.