Video Prompts for Video Perfection

An ice moon at night: a space elevator tether rising from a domed city, a gas giant filling the sky behind it

Most AI video prompts fail for the same few reasons, and none of them are the adjectives. This page is the whole working library: the prompts behind every post on our account, the eleven-block structure the long ones share, the image-first method the short ones use, the five rules that decide whether generated footage reads as cinematic or as obviously AI, and the exact wording that rescued a shot that came back broken. Copy any of it.

The model and settings behind these video prompts

Seedance 2.0 — mode std, 1080p, 9:16, genre epic, high bitrate, native audio off, five seconds per shot. Native output is 1080×1920 at 24fps. Four shots, five seconds each, trimmed down to 3.7s apiece, for a 14.8-second finished cut.

The AI video prompt structure: eleven blocks, in this order

Every prompt below is the same eleven blocks in the same order. The order is doing real work: spatial rules come before camera style, optics come before aesthetic language, and lighting is written as a priority constraint rather than as decoration. Bury a placement rule inside style prose and the model will quietly ignore it.

SCENE CONTEXT one or two sentences, this shot only FORMAT MODE single take / real-time / no subtitles / no music LOCATION MAP camera position + facing, then fg / mg / bg FIRST FRAME AND everything required is already in frame one, SPATIAL BLOCKING nothing revealed late OPTICS diagonal field of view in degrees + the visible outcome CAMERA ONE move only, as physical operator behaviour ACTION TIMING time blocks: 0:00-0:02, 0:02-0:04, 0:04-0:05 PHYSICS mass, inertia, contact, delay, what the medium does LIGHTING sources, direction, what is exposed for, what crushes AUDIO silent POSITIVE CONSTRAINTS materials, palette, and the failure named positively

Five rules for writing AI video prompts

  1. 1Degrees, not millimetres. The model responds to observable lens results, not camera metadata. “84 degree diagonal field of view, camera 3 meters from the foreground truss” controls the shot. “35mm f/1.4” does not.
  2. 2One camera move per shot. An orbit into a dolly into a pan disorients. Drift, or tilt, or pace — pick one and hold it for the full five seconds. Whip-fast moves read as a transition preset, not a camera.
  3. 3States, not transitions. “two welder drones strike arcs at the rib joints” lands. “a drone approaches and begins to weld” collapses.
  4. 4Kill weak spatial words. Never near, around, beside, nearby. Always “40 meters off the port flank”, “3 meters from the lens”, “directly in the centre of frame”.
  5. 5Name the failure in positive form. “The frame stays clean of lettering, signage, watermarks” works. “Don’t add text” gets ignored, or inverted.

Shot 1 — Orbital shipyard

84° wide, slow lateral drift. White sun, cold blue planet-bounce, orange work lamps.

SCENE CONTEXT
A half-finished interstellar ship hangs in an orbital construction cradle above a ringed ice planet while small autonomous welder drones work along its exposed spine.

FORMAT MODE
Single continuous take. Real-time motion. No internal cuts. No subtitles. No music.

LOCATION MAP
Camera floats in open vacuum 40 meters off the ship's port flank, facing the hull broadside. Foreground: one lattice truss beam of the cradle crossing the lower-left of frame, 3 meters from the lens. Midground: the ship's ribbed spine and unplated hull sections, 40 meters away, running from the bottom of frame to the top. Background: the ringed ice planet fills the upper right, its ring plane cutting a hard diagonal across the sky. The sun is a small hard white point far off screen-left.

FIRST FRAME AND SPATIAL BLOCKING
The first visible frame already contains the truss beam in the lower-left foreground, the ship's spine running vertically through the centre third of frame, and the planet limb in the upper right. Four welder drones are already clinging to the spine at different heights and two more drift between the ribs. Nothing is revealed late. No empty establishing frame.

OPTICS
84 degree diagonal field of view, classic wide-angle lens character, camera 3 meters from the foreground truss. The foreground truss looms large and close, the ship and the planet stay visible to the frame edges, deep edge-to-edge focus, straight structural lines stay rectilinear, no fisheye curve. The camera remains physically close with wide spatial expansion for the whole shot.

CAMERA
One move only: the camera drifts laterally toward screen-right at a slow constant rate for the full five seconds. The foreground truss slides across frame faster than the ship behind it, the planet barely moves. Subtle organic settle as if mounted on a slow rig. Fixed focal length, no zoom, no roll.

ACTION TIMING
0:00 to 0:02 - two welder drones strike arcs at the rib joints, hard blue-white arc light flares and dies, throwing moving shadows across the ribs.
0:02 to 0:04 - a third drone releases a finished rib section and thrusts clear on a short puff of cold gas, the section drifting away with a slow tumble.
0:04 to 0:05 - the lateral drift carries the foreground truss out of frame left and opens the full length of the ship.

PHYSICS
Everything behaves in vacuum. Drones translate in straight lines and correct their attitude with visible thruster puffs. Released hardware holds constant velocity and slow rotation. Tether cables lag behind their anchor points and settle slowly. Weld sparks travel in straight lines and fade rather than falling.

LIGHTING
One hard distant white sun from screen-left. Surfaces facing left carry hot specular highlights, everything facing right falls into deep near-black shadow filled only by cold blue planet-bounce. Exposure is set for the sunlit metal so the shadow side crushes. Orange sodium work lamps mark the drone worksites. The arc welds are the only other light and they are brief. No frontal key, no beauty fill, no ambient glow.

AUDIO
Silent.

POSITIVE CONSTRAINTS
Photoreal hard-science-fiction engineering: load-bearing lattice, panel seams, radiator fins, scorch marks, machined tolerances, hull plating in progress. Palette limited to white sunlight, cold blue planet-bounce and orange work lamps. Repeated structural members hold even spacing through perspective. The frame stays clean of lettering, signage, subtitles, watermarks and interface graphics.

Shot 2 — Walking foundry

107° wide rectilinear, locked off at ground level. Rust red, dust ochre, one molten orange line.

SCENE CONTEXT
A colossal six-legged construction machine walks slowly across a rust-red plain on an alien world, extruding a ribbed habitat arch behind it as it goes.

FORMAT MODE
Single continuous take. Real-time motion. No internal cuts. No subtitles. No music.

LOCATION MAP
Camera sits on the ground, lens 30 centimeters above the red dust, facing across the plain. Foreground: coarse gravel and drift ridges of red dust running from the bottom edge of frame to 2 meters out. Midground: the walking machine, 60 meters tall, crossing from screen-left to screen-right at a distance of 120 meters, its six articulated legs planting one at a time. Background: the finished section of habitat arch it has already printed, receding toward the horizon on the left, and a pale dust-hazed sky with a small bright sun high on the right.

FIRST FRAME AND SPATIAL BLOCKING
The first visible frame already contains the foreground dust ridges along the bottom, the walking machine standing in the centre third with two legs mid-stride, and the completed arch stretching away behind it to screen-left. The extrusion head is already laying material. Nothing is revealed late.

OPTICS
107 degree diagonal field of view, wide rectilinear lens character, camera 0.5 meters from the nearest foreground gravel. Immediate foreground looms large, the plain and the machine spread wide to all frame edges, deep edge-to-edge focus, straight lines remain straight, no fisheye bubble, no circular vignette.

CAMERA
One move only: the camera holds locked off on the ground for the full five seconds with a faint settle as ground vibration reaches it. The machine moves through the frame, the camera does not chase it. No pan, no zoom, no crane.

ACTION TIMING
0:00 to 0:02 - the machine's forward right leg swings through and plants, driving a ring of red dust outward across the plain in a low expanding wall.
0:02 to 0:04 - the extrusion head at its rear lays a new rib of the arch, molten material glowing orange as it is placed and darkening to grey within a meter.
0:04 to 0:05 - the dust wall from the footfall reaches the foreground and drifts through the lower frame in thin sheets.

PHYSICS
The machine carries real mass: each leg accelerates, decelerates and lands with impact, the body settles on the loaded leg, hydraulic travel is visible. Dust thrown by a footfall rises, expands and falls slowly in low gravity, blowing with a consistent wind direction. The extruded material sags very slightly before it sets. Cables and hoses on the machine lag behind its motion and swing on their own delay.

LIGHTING
One small hard sun high on screen-right through thin dust haze. Long dark shadows stretch toward the camera on the left. Suspended dust scatters the light into visible shafts. The freshly extruded rib is the only warm light source at ground level. Exposure is set for the sunlit dust, shadow sides go deep and cold. No frontal fill, no studio softness.

AUDIO
Silent.

POSITIVE CONSTRAINTS
Photoreal hard-science-fiction heavy industry: worn paint, dust-caked joints, hydraulic rams, cooling fins, cable runs, ceramic printing head. Palette limited to rust red, dust ochre, machine grey and one orange molten line. The arch's repeated ribs hold even spacing as they recede toward the horizon. The frame stays clean of lettering, signage, subtitles, watermarks and interface graphics.

Shot 3 — Ice-moon space elevator

84° wide, very slow tilt up the tether. Blue ice, amber interiors, grey-gold cloud bands.

SCENE CONTEXT
A space elevator tether rises from a domed city on an ice moon, with a climber car ascending the tether while the gas giant the moon orbits hangs above the horizon.

FORMAT MODE
Single continuous take. Real-time motion. No internal cuts. No subtitles. No music.

LOCATION MAP
Camera stands on the ice plain 400 meters from the city's edge, lens 2 meters above the surface, facing the city. Foreground: ridged blue ice and wind-carved sastrugi running from the bottom edge of frame to 20 meters out. Midground: the domed city, a cluster of pressurised glass domes and lit habitat blocks, its base anchor tower directly in the centre of frame. Background: the tether, a single fine ribbon rising from the anchor tower straight up through the whole height of frame and out of the top; the banded gas giant occupying the upper right of the sky behind it; a small hard sun low on the left.

FIRST FRAME AND SPATIAL BLOCKING
The first visible frame already contains the foreground ice ridges along the bottom, the lit domes across the centre band of frame, the anchor tower in the exact centre, the tether ribbon running vertically up the centre of frame, and the gas giant in the upper right. A climber car is already visible on the tether at about one third of the frame height. Nothing is revealed late.

OPTICS
84 degree diagonal field of view, classic wide-angle lens character, camera 2 meters from the nearest foreground ice ridge. Foreground ice presence looms large and close, the city and the sky remain visible to the frame edges, deep edge-to-edge focus, the tether stays a straight rectilinear vertical line, no fisheye curve.

CAMERA
One move only: a very slow tilt upward for the full five seconds, following the tether from the city toward the climber car. Constant rate, no acceleration. Subtle organic settle. No zoom, no roll, no dolly.

ACTION TIMING
0:00 to 0:02 - lights inside the domes flicker and shift as the camera holds low on the city; thin ice crystals drift across the foreground on a steady wind.
0:02 to 0:04 - the tilt lifts past the anchor tower and follows the tether upward, the climber car sliding steadily up the ribbon.
0:04 to 0:05 - the tilt brings the gas giant's banded cloud tops fully into the upper frame with the tether crossing them.

PHYSICS
The tether is under tension and holds a dead-straight line with only a very slow lateral oscillation. The climber car moves at constant velocity with no jitter. Ice crystals travel with a single consistent wind direction and settle. Dome glass carries real reflections of the sun and the gas giant. Structures cast hard shadows onto the ice consistent with the low sun.

LIGHTING
One small hard low sun on screen-left throwing long blue shadows across the ice toward the camera. The gas giant adds a broad cold fill from the upper right. The city's own warm interior lighting is the only warm source in frame and it spills onto the ice around the domes. Exposure is set for the sunlit ice so the unlit sides of structures fall dark. No frontal fill, no ambient glow, no lens-flare starbursts.

AUDIO
Silent.

POSITIVE CONSTRAINTS
Photoreal hard-science-fiction colony engineering: geodesic dome framing with even panel spacing, insulated pipe runs, anchor tower guy cables, service lighting on the tower at regular intervals. Palette limited to blue ice, cold sunlight, warm amber city interiors and the grey-gold cloud bands of the gas giant. Repeated structural members hold even spacing through perspective. The frame stays clean of lettering, signage, subtitles, watermarks and interface graphics.

Shot 4 — Vessel under drive

47° normal, camera pacing the ship. This one is take 2 — the fixes are below.

SCENE CONTEXT
A finished interstellar vessel runs its main drive at full power and travels across the night side of a banded gas giant, seen from a camera pacing alongside it.

FORMAT MODE
Single continuous take. Real-time motion. No internal cuts. No subtitles. No music.

LOCATION MAP
Camera is in vacuum 120 meters off the vessel's starboard flank, level with its spine, pacing it at matched speed. Foreground: nothing between camera and vessel. Midground: the vessel, seen three-quarter from behind and to the side, its long hull running diagonally from lower-left of frame to upper-right, drive nozzles at the lower-left end. Background: the gas giant fills the entire sky behind, horizontal cloud bands clearly readable in cold grey-blue, the terminator crossing the upper third of frame.

FIRST FRAME AND SPATIAL BLOCKING
The first visible frame already contains the whole vessel occupying roughly half the frame diagonally, the drive plume already burning at the lower-left, and the banded cloud tops filling the background. The vessel keeps this size in frame for the entire shot because the camera paces it. Nothing is revealed late.

OPTICS
47 degree diagonal field of view, standard normal lens character, camera 120 meters from the vessel, natural perspective, no obvious distortion, comfortable depth, the cloud bands behind stay readable and are not exaggerated. This lens character holds for the whole shot with no drift.

CAMERA
One move only: the camera paces the vessel at matched velocity for the full five seconds, holding it at constant size in frame, with a very slow drift toward the vessel's tail. No zoom, no roll, no rack focus, no pull back, no orbit.

ACTION TIMING
0:00 to 0:02 - the drive plume burns steadily behind the nozzles, trailing straight back and away from the vessel toward the lower-left corner, well clear of the lens.
0:02 to 0:04 - the vessel crosses into the planet's shadow, the cold background light drops, and the hull is read by its own orange running lights and the blue rim thrown by the plume.
0:04 to 0:05 - radiator panels along the spine cool from faint orange to dark, the plume holds steady, the vessel still fills the same portion of frame.

PHYSICS
Motion is smooth and continuous with no banking and no jerk, because there is no atmosphere. The plume is a straight collimated column with visible shock diamonds, narrow and directional, and it does not spray sideways across the frame. Radiator panels glow at their roots and fade toward their tips. Light from the plume falls off with distance along the hull.

LIGHTING
Controlled exposure held for the hull. The gas giant's cloud bands sit mid-grey and stay visible as texture for the whole shot. The plume is a hard blue-white source that rims the vessel's tail only. Orange running lights mark the hull edges. Surfaces facing away from both sources go near-black. The image stays legible and dark. Highlights stay contained within the nozzle bells and the plume column.

AUDIO
Silent.

POSITIVE CONSTRAINTS
Photoreal hard-science-fiction spacecraft: radiator fins, thermal blanketing, docking hardpoints, machined nozzle bells, dark hull, evenly spaced panel lines running the full length. Palette limited to blue-white plume, cold grey-blue cloud bands and small orange running lights. The vessel remains the clear subject of the frame at all times. The frame stays clean of lens-flare starbursts, white bloom washing over the image, lettering, signage, subtitles, watermarks and interface graphics.

Every prompt behind every KitCrew post

The four shots above are the long-form Seedance prompts behind the first reel. Everything below is the working library: the actual prompt pairs behind every post on the account, in the order they went out. These are shorter by design. They run on Grok Imagine, image-first, and the job of the second prompt is to stop the model reinventing the subject halfway through the shot.

How the image-first method works. Generate four stills from one prompt, pick the frame that already looks like the shot, then animate that single still. The still carries the design, so the animate prompt only has to carry the motion, the sound and the promise that nothing changes. That last part is not optional: image-to-video only knows the one frame it was handed, and anything the camera reveals afterwards gets invented.

Camera moves. One slow continuous move per shot, and never an orbit. A 180-degree wrap kept its camera move perfectly and lost the subject on the far side of the arc - a smooth capsule arrived as a faceted wedge. If the design matters, keep the camera on one line and re-describe the hardware inside the animate prompt.

The Blue Gauntlet

A drive-through car wash, called as round seven of a world championship.

Grok Imagine, image-first per S279 doctrine. Stills first (Image mode, Quality, x4, 9:16), live seat picks one, then animate that single chosen still. Blue is the controlled variable this round.

Still prompt (Image mode, 9:16, Quality, x4)
Inside a modern automatic drive-through car wash tunnel, one ordinary sedan is midway
through the wash cycle. Thick blue-tinted foam coats the hood, roof and windshield.
Rows of spinning fabric wash mitts and rotating brushes flank the car on both sides.
Overhead, a continuous strip of cobalt-blue LED tunnel lighting runs the length of the
bay, throwing hard blue reflections across the wet, glossy floor and the car's
foam-covered paint. Fine mist and blue-lit water spray hang in the air. The wash
tunnel's steel rail and guide track run under the car's wheels. Shot low and wide,
straight down the tunnel, car centered, foam and blue light dominating the frame.
Photoreal, cinematic, high dynamic range, sharp focus on the car, shallow falloff into
the blue haze behind it. Vertical 9:16 frame. Blue is the dominant color of the image,
not an accent.
Animate prompt
Continue this exact car wash tunnel: the same sedan under the same cobalt-blue LED
strip lighting, the same rotating fabric wash mitts and brushes at either side, the
same blue-tinted foam sheeting off the hood and roof, the same wet reflective floor
throwing blue light back up into frame, the same steel guide rail under the wheels.
One slow continuous camera move only: a slow forward push down the tunnel alongside
the car, foam sliding off the glass as it passes under the next brush station. No
cuts, no scene changes, no orbit, no camera spin. Audio: driving garage rock, live
drums, distorted electric guitar, roughly 150 bpm, no vocals, running continuously for
the full length of the take. Identical materials, proportions and colours from the
first frame to the last.

Hour Twenty-Three

An ice resurfacer, called as the twenty-third hour of a twenty-four-hour endurance race.

Chris's seat is Grok Imagine (SuperGrok, free). Stills first, pick one, then animate that still. 720p is the plan ceiling; the finished take gets lanczos-upscaled 1.5x into the 1080x1920 canvas at edit time - that is expected, not a defect.

Still prompt - Image mode, Quality, x4, 9:16
A large ice resurfacer machine gliding slowly across a vast, completely empty ice hockey
rink at night, seen from low ice level. The machine is recognisable instantly: white and
blue body, wide front blade, rotating brush housing, a fresh wet resurfaced track gleaming
behind it on the ice. The arena is lit entirely in cold blue light - blue LED floodlights
overhead, a blue cast on the ice surface, faint blue rink boundary lines glowing up through
the fresh water sheen, blue haze hanging in the empty stands behind the boards. No players,
no crowd, no other vehicles. Wide cinematic angle, shallow depth of field, photorealistic,
broadcast sports-photography lighting, vertical 9:16 frame, blue as the dominant colour of
the entire image, not an accent.
Animate prompt - one continuous take, no cuts, no orbit
Animate this exact ice resurfacer on the ice rink: same white and blue machine, same wide
front blade, same rotating brush housing, same fresh wet track gleaming behind it, same
cold blue arena lighting washing the whole rink blue, same empty stands in blue haze. One
slow continuous camera move only - a slow forward push low along the ice, tracking the
machine as it works. No cuts, no scene changes, and no orbit around the machine - the
camera stays on one line the whole take. Audio: driving rock music, live drums, distorted
electric guitar, tempo around 145 bpm, no vocals, energy held steady and consistent through
the entire take, no drops, no breakdowns. Identical materials, proportions and colours on
the machine and the ice from the first frame to the last - nothing about the resurfacer or
the rink changes shape, colour or scale as the camera moves.

The Final Ascent

A shopping-mall escalator, narrated as a high-altitude summit push.

Grok Imagine, Chris's seat, image-first. Stills first (Image mode, Quality, x4, 9:16), the live seat picks one, then animate that chosen still. See BUILD_KIT_SPEC.md Section 3 for the protocol this follows.

Still prompt (Image mode, Quality, x4, 9:16)
A long modern shopping mall escalator rising through a bright glass atrium, shot straight on
from a low angle so the escalator climbs from the bottom of the frame toward a distant upper
level. The handrail and the leading edge of every step glow with cool blue LED lighting, the
strongest color source in the frame. Overhead, tall glass panels let in a flat blue-white
daylight that washes the polished floor and chrome railings in the same cool blue tone. A
handful of ordinary shoppers ride the escalator and stand on the level above, carrying bags,
in plain everyday clothing, no special gear. The escalator itself is unmistakable and fills
the center of the frame within the first glance: metal steps, moving handrail, safety combs,
floor-level skirting panel. Wide-angle architectural photography, sharp focus, high dynamic
range, cinematic but naturalistic, no text, no logos, no people looking at camera. Blue is
the dominant color of the entire image, not an accent.
Animate prompt (image-to-video, one continuous take, <=15s)
Animate this exact still. The escalator, its metal steps, the glowing blue LED handrail and
step-edge lighting, the blue-white glass atrium light, and the ordinary shoppers riding it
must stay identical to the starting frame. One slow continuous camera move only: a slow
forward push toward the top of the escalator, tracking low and steady, as if the camera
itself is riding the ascent. No cuts, no orbit, no scene change, no new camera angle. The
shoppers keep riding at a natural, unhurried pace. The blue LED glow on the handrail and
steps stays constant and dominant throughout the shot. Audio: driving rock music, live drums,
distorted electric guitar, roughly 140 bpm, building in intensity across the full take, no
vocals, no lyrics. Identical materials, proportions and colours from the first frame to the
last.

Recovery At Depth

A robotic pool cleaner, narrated as a deep-ocean salvage dive.

Two prompts, ready to paste into Grok Imagine. Stills first (Image mode), pick one, then animate that single chosen still. Do not skip the still step.

Still prompt - Image mode, 9:16, Quality, x4
A small robotic pool cleaner resting on the tiled floor of a lit backyard swimming pool at
night, seen from a low diver's-eye angle close to the tile. Its float cable trails up out
of frame toward the surface. A pool ladder and the tile line are visible within the first
second so the setting reads instantly as a residential swimming pool. The frame is
dominated by chlorine blue water and blue underwater pool lights, with blue caustic light
rippling across the tile and across the machine's white-and-blue segmented shell. Water
clarity is high, a little particulate drifting slowly through the light beam. Sharp focus
on the machine, cinematic, photoreal. No text, no overlays, no logos. 9:16 vertical.
Animate prompt - one continuous take
Animate this still with one slow continuous camera move only: a slow forward drift toward
the robotic cleaner. No cuts, no orbit, no scene change. Re-describe the hardware as it
holds its position on the tiled pool floor: a compact robotic cleaner unit with a
white-and-blue segmented shell, twin drive tracks, a mesh intake at the front, and a float
cable trailing straight up out of frame toward the surface. The water stays chlorine blue
for the whole shot, lit by blue underwater pool lights, with blue caustic light rippling
across the tile and across the machine's shell - blue is the dominant color of the entire
frame, not an accent. Score the shot with driving rock music: live drums, distorted
electric guitar, no vocals, tempo holding steady around 135 bpm for the full length of the
take. Identical materials, proportions and colours from the first frame to the last.

All Stations Are Go

A 24-hour laundromat at 3am, narrated as a mission-control launch sequence.

Chris's seat is Grok Imagine (SuperGrok, free tier). Stills first, Image mode, Quality, x4, 9:16. Pick one still, then animate it in a single continuous take. Paste these prompts as-is.

Still prompt (Image mode, 9:16, Quality, x4)
A photorealistic empty 24-hour laundromat at three in the morning. An unbroken wall of a dozen
chrome-front commercial front-loading washing machines runs the length of the room, every drum
mid-cycle, spinning fast and blurred behind round glass doors, water visibly sloshing with a
blue tint inside each one. Overhead blue-white fluorescent tube lighting is the dominant light
source in the frame, casting a cold blue-white wash across the walls, the machines and the
floor. Each washer's control panel glows with a small blue LED display. In the foreground, a
stainless steel folding table and a row of empty orange plastic chairs sit under the lights,
and a bank of coin slots is visible on the machine fronts. Speckled linoleum floor reflecting
the blue light. Wide static angle, deep depth of field, no people, cinematic realism, slight
lens vignette. Vertical 9:16 frame.
Animate prompt (single continuous take, image-to-video, from the chosen still)
Animate this exact laundromat: the same wall of chrome-front commercial washing machines, the
same drums still spinning fast behind round glass doors with blue-tinted water sloshing inside,
the same cold blue-white fluorescent overhead lighting, the same glowing blue LED control panel
displays, the same stainless folding table and empty orange plastic chairs, the same coin slots
and speckled linoleum floor. One slow continuous camera move only: a slow forward push down the
row of machines. No orbit, no cuts, no scene changes. Add a driving rock soundtrack under the
shot: live drums, distorted electric guitar, roughly 150 beats per minute, steady through the
whole take, instrumental with no vocals. Identical materials, proportions and colours from the
first frame to the last frame.

The Northern Excavator

A bucket-wheel excavator at golden hour, narrated as a migrating animal.

Note: EXCAVATOR_01 ("The Northern Excavator," live at instagram.com/reel/DbtySv2pO5x) was built before this repo started saving Grok prompts alongside the finished kit. The prompts below were authored after the fact, from the finished cover frame and caption, to bring this post up to the current build-kit standard. They are NOT the exact prompts that generated the live reel, and they should be read as a faithful reconstruction, not a recovered original.

Still prompt (Image mode, 9:16, Quality, x4)
A colossal bucket-wheel excavator crossing open steppe grassland at golden hour, shot like a
nature documentary subject. Weathered rust-orange and dark steel construction, a huge toothed
digging wheel mounted at the end of a long lattice-truss boom, operator cabins and catwalks
stacked beneath the boom's crossbraces, the whole machine riding on two wide crawler tracks
that carve fresh ruts into dry tan earth. The low sun sits near the horizon on the left,
flooding the scene in warm ochre and amber light. A plume of dust and haze trails up and back
from the tracks, lit gold from behind. Foreground is dry steppe grass, bent and backlit. Shot
low and wide, machine centered and filling the frame, sky graduating from warm amber near the
horizon to dusty rose-grey above. Photoreal, cinematic, high dynamic range, sharp focus on the
machine, shallow falloff into the haze behind it. Vertical 9:16 frame. Warm ochre and amber
light dominate the image; no blue anywhere in the frame.
Animate prompt
Continue this exact scene: the same colossal bucket-wheel excavator, the same rust-orange and
dark steel hull, the same toothed digging wheel and lattice-truss boom, the same crawler
tracks carving ruts in the dry tan earth, the same low amber sun on the horizon throwing warm
light across the machine and the dust plume trailing behind it, the same dry steppe grass in
the foreground. One slow continuous camera move only: a slow lateral tracking drift alongside
the machine as it advances across the plain, matching its pace. No cuts, no scene changes, no
orbit, no camera spin. Audio: sweeping orchestral nature-documentary score, strings and low
brass, slow unhurried build, no vocals, running continuously for the full length of the take.
Identical materials, proportions and colours from the first frame to the last.

The Engineering Exists

A magnetic launch rail climbing a ridge. The continuity lesson lives here.

Grok Imagine, image-first per S279 doctrine. Stills first (Image mode, Quality, x4, 9:16), live seat picks one, then animate that single chosen still.

Still prompt (Image mode, 9:16, Quality, x4)
Elevated aerial view looking straight down the spine of a bare rock mountain ridge at
sunrise, the sun low on the horizon directly ahead, light rays fanning out through thin
high cloud. A raised mass-driver rail runs the length of the ridge on a chain of black
steel gantry towers, each one guyed to the bare rock with steel support cables. Built into
the rail bed, a continuous run of evenly spaced glowing pale blue coil rings recedes in
perspective toward the sun. Two-thirds of the way up the rail, sitting square in its
gantry cradle, is a white, hard-edged, faceted payload sled with a blunt nose and no
windows. The low sun throws long hard-edged shadows from the truss towers across the bare
grey rock. No vegetation, no buildings, no people. Distant hazy mountain ridgelines fade
out on the left horizon. Photoreal, cinematic, high dynamic range, sharp focus on the sled
and the near rail, soft atmospheric falloff into haze toward the sun. Vertical 9:16 frame.
Animate prompt
Continue this exact mountain ridge rail scene, same sunrise light, same low sun on the
horizon, same steel gantry towers and guy cables anchored to the bare rock.

--- ON-RECORD FRAGMENT (verbatim from README.md, gap marked) ---
the same white hard-edged faceted payload sled with a blunt nose and no windows, riding
the same heavy steel rail with evenly spaced glowing pale blue coil rings [[gap in the
record - bridging text authored below]] identical materials proportions and colours from
the first frame to the last.
--- END ON-RECORD FRAGMENT ---

The sled climbs slowly up the rail toward the sun, gaining speed, fine dust kicking up off
the rock in its wake as it passes under each gantry tower in turn. One slow continuous
camera move only: a slow push forward and upward along the rail, following the sled up the
ridge toward the light. No cuts, no scene changes, no orbit, no camera spin. Audio:
sparse cinematic synth score, slow building low drone with a rising pulse under it, no
vocals, running continuously for the full length of the take.

The soundtrack: a hard-rock guitar riff, generated in code

There is no music generator on the rail we use. So the riff under our megastructure reel is not licensed, not sampled and not prompted — it is synthesised from scratch by a Python script with no dependencies, and you can have it. Run it, change the notes, change the tempo, use the output.

It took three attempts, and the first two failures are the useful part

Version one was constant palm-muted eighth notes with drums. It was rhythmically correct and it sounded like a backing track, because a riff is mostly silence. Stabs that ring out into a hole, accents landing on the offbeat, a question bar answered by a second bar. Playing continuously is the single fastest way to sound like a loop instead of a hook.

Version two fixed the arrangement and still sounded synthetic, because it was built from sawtooth oscillators through a distortion curve. That is a synth patch, not a guitar, and no amount of EQ rescues it. Arrangement and timbre are separate problems and solving one does not touch the other.

The two things that make a sound read as electric guitar

  1. 1A plucked string, not an oscillator. Karplus–Strong: excite a delay line with a short noise burst, then feed it back through a damping filter. The delay length sets the pitch and the damping decides how fast the harmonics die. On a real string the high harmonics fade first and the fundamental rings on — a modelled string does that automatically, an oscillator never does, and that mismatch is what your ear hears as fake.
  2. 2A speaker cabinet. A guitar cab is not full range. It rolls off steeply above about 5 kHz, peaks around 2.5 kHz, and has nothing below roughly 90 Hz. Distortion generates harmonics all the way up the spectrum, and without a cabinet those stay in the signal and read as fizz. Cabinet simulation is most of what separates “amplifier” from “broken speaker”.

After that, the details: double-track the rhythm guitar with a slight detune and about 11 ms of offset for width, clip asymmetrically so you get even harmonics as well as odd, and mix the guitar as the lead voice with the rhythm section well behind it. Rock is iconic for the riff, not the kit.

Why AI video looks fake — and the three lines that fixed it

Take 1 of the final shot blew out into a white lens-flare wash by three seconds and shrank the ship to a dot — a dead payoff right where the reel loops. The wording changes that fixed it are the most useful thing on this page, because they are the general case: when a generation goes wrong, the fix is usually a physical instruction, not a prettier adjective.

What went wrongThe instruction that fixed it
Ship shrank to a dot“the camera paces the vessel at matched velocity, holding it at constant size in frame” — a pacing camera instead of a fixed one
Plume sprayed across the whole frame“trailing straight back and away from the vessel, well clear of the lens”, plus “a straight collimated column, narrow and directional”
White bloom ate the imageexposure held for the hull, cloud bands “sit mid-grey”, “highlights stay contained within the nozzle bells” — and the failure itself named in POSITIVE CONSTRAINTS

Editing AI video: the two rules that decide whether it looks amateur

Trim the edges. Every generated clip drifts at the head and the tail. These were cut 0.55s off the front and everything past 4.25s — 3.7 seconds kept out of five. Hard cuts only; dissolves make generated footage look softer than it is.

Match your timeline to the model, not the other way round. These render at 24fps natively, so the edit is 24fps. Conforming 24fps source to a 30fps timeline means an upscale plus frame duplication, which puts judder on slow pans before the platform’s own re-compression ever touches it.

AI video prompting: common questions

What makes an AI video prompt actually work?

Structure and physical specificity, not adjectives. Write the same blocks in the same order every time — scene, location map, first frame, optics, one camera move, timed action, physics, lighting, constraints — and replace every vague spatial word with a measurement. “40 meters off the port flank” controls the shot; “near the ship” does not.

Why does my AI video look fake?

Usually three causes: the exposure blows out into white bloom, the subject drifts out of frame because the camera is fixed instead of pacing it, or repeated man-made structures lose their rhythm — railings, seating, panel lines that melt or space unevenly. All three are fixed with physical instructions, not prettier adjectives. And check frames at the crop you will actually use, never from the wide.

Should I use camera terms like 35mm or f/1.4 in an AI video prompt?

No. Video models respond to observable lens results, not camera metadata. Use diagonal field of view in degrees plus the visible outcome — “84 degree diagonal field of view, camera 3 meters from the foreground, straight lines stay rectilinear, no fisheye curve”. Millimetres, f-stops, ISO and lens brand names are weak control by comparison.

How long should an AI video prompt be?

Long enough to specify every block once, which in practice runs about 3,000–3,500 characters for a five-second shot. Length is not the point, coverage is. A prompt that omits lighting and physics leaves both to the model, and the model will pick the flattest, safest version of each.

How many shots do you need for a 15-second video?

Four five-second generations, trimmed to about 3.7 seconds each, gives a 14.8-second cut. Generated clips drift at both the head and the tail, so losing roughly half a second off the front and everything past the four-second mark is normal, not waste.

Do you need a character sheet for a short AI video?

Only if a person or creature recurs across shots. Character sheets, identity models and spatial layout systems exist to hold one subject stable across hundreds of shots in a film. On a three or four shot clip with no returning face they cost time and money and buy nothing.

Why a marketing automation company is giving away video prompts

Because the prompt was never the hard part. Anyone can copy the four blocks above and get a good-looking shot tonight. Doing it every week — on schedule, on brand, checked before it goes out, with the failures caught before the audience sees them — is the part that actually costs people their evenings.

That is the job KitCrew does. If the reel is the thing you liked, take the prompts and go build something. If the every week, without you part is the thing you want, that’s what the crew is for.