Prompt Cheat Sheet
The words that actually change a generated image or clip — what each one does, when to reach for it, and builders that assemble them into something you can paste.
New to this? Start with the four rules below, build one prompt with the tool, and ignore the rest until you need something specific. The term list is for looking things up mid-project, not for reading straight through.
How an Image Prompt Is Built
An image model does not follow instructions so much as it responds to description. Everything below is vocabulary for describing more precisely. Four things matter more than any individual term:
Say what the picture is of, first and in plain words. “A weathered lighthouse on a rock shelf, waves breaking against the base.” No term on this page can supply a subject — they only shape one that is already there.
Add three to five terms, not twelve. A prompt carrying a dozen modifiers does not produce a twelve-modifier image. It produces an averaged one, because the terms contradict each other and the model splits the difference.
Take at most one from each group. One camera angle, one lighting, one atmosphere, one style. Two lighting terms fight. Two camera angles are incoherent.
When a result is close but wrong, change one term. Swap the lighting, keep everything else, regenerate. Rewriting the whole prompt throws away whatever was working — and something usually was.
Describe your subject, then hit Add on any term below. They assemble here in the right order.
Camera and Framing
Where the camera sits and how tightly it crops. This is the highest-leverage group on the page: the angle changes what an image means without changing anything in it. Pick exactly one.
Close-up
tight close-up, subject fills the frame
Crops in until texture and expression are the whole picture — pores, fabric weave, condensation on a glass.
Reach for it when detail or emotion is the point. Add “extreme” if you want to get closer than a person normally stands.
Wide shot
wide shot, wide-angle lens
Pulls back to include the surroundings, with slight edge stretching that makes a space feel larger than it is.
Reach for it when the viewer needs to know where they are. Landscapes, interiors, establishing shots.
Fisheye
fisheye lens, heavy barrel distortion
Extreme curvature — straight lines bow outward and the center bulges toward you.
Reach for it when you want skate-video energy or deliberate disorientation. It is never subtle.
Low angle
low camera angle looking up at the subject
Puts the viewer beneath the subject so it looms overhead. Reads instantly as power and scale.
Reach for it when something should feel monumental or threatening. Say “worm’s-eye view, camera on the ground” for the extreme version.
High angle
high camera angle looking down at the subject
Places the viewer above the subject, which shrinks it and makes it feel observed.
Reach for it when you want vulnerability, or need to show a scene’s layout without going fully overhead.
Overhead
bird’s-eye view, camera directly overhead
Straight down. Flattens the scene into a graphic arrangement of shapes rather than a space.
Reach for it when shooting flat-lays, table settings, maps of a room, or any pattern. Say “overhead flat lay” for product work.
Drone shot
aerial drone shot, high above the scene
High altitude at a slight tilt, so landscape reads as terrain instead of as a view.
Reach for it when the scale of a place is the subject: coastlines, city grids, crowds.
Side profile
side profile view
Camera perpendicular to the subject, so outline and silhouette dominate over depth.
Reach for it when the shape is what matters — profile portraits, or a product where thickness needs to read.
First person
first-person POV, as seen through the subject’s own eyes
Puts the viewer inside the scene, usually with hands or held objects at the frame edge.
Reach for it when you want the viewer to be the person rather than watch them.
Over the shoulder
over-the-shoulder shot
A shoulder in soft foreground frames whatever the person is looking at, creating depth in two planes.
Reach for it when showing someone looking at something — a screen, a view, another person.
Lens, Focus, and Motion
Optical and shutter behavior. These are the terms that make an image feel physically photographed rather than generated, and they are the most commonly skipped.
Shallow focus
shallow depth of field, creamy bokeh in the background
Subject razor sharp, background dissolved into soft blobs of light.
Reach for it when shooting almost any portrait or product. This is the single fastest way to make an image look professionally shot.
Anamorphic
anamorphic widescreen, horizontal blue lens flares
Cinema aspect ratio with oval bokeh and streaking flares across the frame.
Reach for it when the image should read as a film still rather than a photograph.
Film grain
visible film grain, analog texture
Adds fine noise, which undoes the plastic smoothness that gives generated images away.
Reach for it when anything looks too clean. Costs nothing and helps almost every photographic prompt.
Motion blur
motion blur, subject in motion
Streaks the moving parts while the rest stays put. Reads as speed.
Reach for it when movement matters more than clarity — traffic, dancers, a thrown ball.
Long exposure
long exposure, light trails
Moving lights smear into ribbons; water and clouds go silky and smooth.
Reach for it when shooting night cities, waterfalls, or star trails. Different from motion blur: the lights move, not the subject.
Frozen action
high shutter speed, frozen mid-action
Everything razor sharp at peak movement — splashes suspended, dust hanging in the air.
Reach for it when the moment matters more than the movement. The direct opposite of motion blur.
Light and Time of Day
The group worth learning first. Lighting decides whether an image reads as a snapshot or as a shot. Choose one time-of-day term and at most one lighting technique — stacking two light sources produces mush.
Golden hour
golden hour light, warm low sun, long shadows
Warm orange light raking across the scene from a low angle.
Reach for it when you do not yet know what light to ask for. It flatters nearly every subject and is very hard to make ugly.
Blue hour
blue hour, cool twilight just after sunset
Even blue ambient light, with artificial lights just beginning to register.
Reach for it when shooting cities and architecture. It is the narrow window where sky and windows balance each other.
Sunrise or sunset
sunset on the horizon
The sun itself in frame with color spread across the sky. Sunrise runs cooler and cleaner, sunset warmer and hazier.
Reach for it when you want the sun visible. If you only want the light, ask for golden hour instead — it is the better term.
Moonlight
moonlit, cool blue night light
Low, cold, directional light with deep shadows and little color.
Reach for it when shooting night exteriors that still need to be readable.
Soft light
soft diffused light, overcast
Shadows with gentle edges and nothing blown out.
Reach for it when shooting faces or food, or anything where a hard shadow would be unflattering.
Dramatic light
dramatic high-contrast lighting, deep shadows
One hard source with most of the frame falling away into black.
Reach for it when you want tension and weight. The opposite decision from soft light — you cannot have both.
Backlight
backlit, bright rim light outlining the subject
Light from behind draws a glowing edge that separates the subject from the background.
Reach for it when a subject needs to pop off a dark background. Works beautifully combined with fog or haze.
God rays
volumetric light, visible god rays through the air
Beams of light you can actually see, because there is dust or moisture for them to catch on.
Reach for it when shooting forests, cathedrals, warehouses. Needs something in the air, so pair it with haze.
Neon
neon lighting, saturated magenta and cyan
Colored light sources sitting inside the scene, reflecting off wet and glossy surfaces.
Reach for it when doing night city work. Name the two colors you want, or you will get all of them at once.
Studio
studio lighting, seamless backdrop
Fully controlled light against a plain background. No environment, all subject.
Reach for it when shooting products or clean portraits where the setting would only distract.
Cinematic
cinematic lighting and composition
A broad request: motivated light sources, wider framing, film-like color.
Reach for it when you are still exploring. Useful as a starting point, but replace it with a specific lighting term once you know what you actually want.
Atmosphere and Weather
What is in the air between the camera and the subject. These carry mood cheaply, and they are what make light visible — most images that get called cinematic simply have something in the air.
Fog
fog
Softens everything with distance, so depth reads as a stack of separated layers.
Reach for it when you want mood and clean separation between foreground and background. Say “light mist” if fog is too heavy.
Rain
rain, wet reflective streets
Falling water plus the wet surfaces it leaves behind — the reflections do most of the work.
Reach for it when you want mood and an excuse for reflected light. Add “at night” for the classic look.
Storm
stormy sky, heavy dark clouds
Dramatic cloud structure with flat, low-contrast light beneath it.
Reach for it when a landscape needs threat and scale.
Snow
falling snow, snow-covered ground
Simplifies a scene into white shapes and bounces light upward into the shadows.
Reach for it when you want quiet, cold, or a clear seasonal read.
Haze and smoke
atmospheric haze and drifting smoke
Gives the air physical substance, so light beams and depth become visible.
Reach for it when almost any lit scene. This is the quiet trick behind most images that look expensively shot.
Firelight
firelight, flames and glowing embers
A warm, flickering light source living inside the frame.
Reach for it when a night scene needs a warm anchor to balance all the cold light.
Settings and Environments
Where the scene happens. Combine one of these with a weather term and a time-of-day term and the location is fully specified.
Underwater
underwater, filtered blue light and floating particles
Blue-shifted light falling in shafts, with visible particulate suspended in the water.
Reach for it when shooting anything submerged. The particles are what sell it as water rather than blue air.
Space
outer space, stars and distant nebula
Black field with hard unfiltered light and no atmospheric falloff.
Reach for it when you want scale and isolation. Note that shadows in space are much harder than people expect.
Desert
desert, sand dunes and open sky
Vast, minimal, high contrast, with strong directional shadows across the sand.
Reach for it when you want emptiness, heat, or a clean horizon line.
Forest
dense forest, tall trees and undergrowth
Vertical structure with dappled light filtering down through a canopy.
Reach for it when you want enclosure and texture. Pairs unusually well with god rays and mist.
Mountains
mountain range, dramatic peaks
Layered ridgelines receding into atmospheric haze.
Reach for it when scale is the subject. The haze between layers is what creates the sense of distance.
Coast
coastline, sand and breaking waves
Open horizon, reflected sky, and motion at the waterline.
Reach for it when you want openness or a natural place to put motion in an otherwise still frame.
City
city skyline
Built environment at scale — geometry, glass, and repeated verticals.
Reach for it when you need an urban setting. Say “city street at night” instead to get down to eye level, where neon and reflections live.
Mood and Genre
Narrative worlds rather than visual treatments. Each one drags in its own props, wardrobe, and palette, which is why they collide badly. Use exactly one.
Dark and moody
dark and moody, muted desaturated palette
Pulls saturation and brightness down across the whole frame.
Reach for it when you want restraint and weight. Works as a modifier on top of most other genres.
Cyberpunk
cyberpunk, neon-lit rain-soaked future city
Dense neon, permanent rain, towering signage, deep shadow.
Reach for it when you want the specific near-future-dystopia look. It is a strong flavor and overrides most style terms.
Sci-fi
futuristic sci-fi, sleek advanced technology
Clean surfaces, restrained palette, engineered geometry.
Reach for it when you want the future without the grime. The opposite end from cyberpunk.
Post-apocalyptic
post-apocalyptic, overgrown ruins and decay
Structures being reclaimed by weather and plants; dust and rust throughout.
Reach for it when you want the aftermath rather than the event.
Fantasy
high fantasy, magical and otherworldly
Impossible architecture, luminous elements, heightened natural forms.
Reach for it when realism is not the goal and you want scale and wonder.
Medieval
medieval, stone and timber, period-accurate
Rough materials, firelight, hand-built construction.
Reach for it when you want a grounded historical setting. Add “period-accurate” or you will drift toward fantasy.
Feudal Japan
feudal Japan, samurai, cinematic
Specific architecture, wardrobe, and a restrained, deliberate compositional style.
Reach for it when you want that setting specifically. Unusually reliable for such a narrow term.
Film noir
film noir, hard shadows and venetian blind light
Black and white or near it, hard slashes of light, rain, cigarette haze.
Reach for it when you want mystery and moral weight. Say “detective” with it to bring in the wardrobe and props.
Superhero
epic superhero scene, heroic scale
Low angles, dynamic poses, dramatic backlighting, extreme scale.
Reach for it when you want comic-adjacent spectacle in a photographic frame.
Surreal
surreal and dreamlike, slightly uncanny
Recognizable elements arranged in ways that do not quite resolve.
Reach for it when you want unease rather than fantasy. The “slightly” matters — drop it and you get chaos.
Style and Aesthetic
Visual vocabularies borrowed from fashion, editorial, and design. Strong flavors that fight each other — one at a time.
Minimalist
minimalist, generous negative space, single focal point
Strips the frame to one subject and a great deal of empty room around it.
Reach for it when you need space for text, or the subject is strong enough to carry the frame alone.
Luxury
luxury, understated wealth, quality materials
Restrained palette, expensive surfaces, careful light. Quiet rather than flashy.
Reach for it when you want premium without gold and sparkle. Say “old money” to push it further toward heritage.
Vintage
vintage 1970s, faded warm color
Period color response with lifted blacks and a warm cast.
Reach for it when you want nostalgia. Always name the decade — “vintage” alone lands somewhere vague.
Editorial
editorial fashion photography, magazine quality
Styled, deliberate, and confident. Composed for a page rather than a moment.
Reach for it when you want polish and intent. Adds wardrobe and posing direction, not just a look.
Street style
street style, candid urban fashion
Looks caught rather than staged, shot at eye level in real surroundings.
Reach for it when you want authenticity. The direct opposite of the editorial treatment.
Photographic Formats
Specific capture media. Reach for these when you want the artifact of the format itself — the border, the grain, the flash falloff.
35mm film
shot on 35mm film, natural grain
Film color response and organic grain instead of digital cleanliness.
Reach for it when you want a photograph that feels made rather than rendered. The most broadly useful term in this group.
Polaroid
polaroid instant photo, white border, soft faded color
Square frame, thick white border, low contrast, slightly shifted color.
Reach for it when you want the object as much as the image — something that looks physically held.
Disposable camera
disposable camera, harsh direct flash
Blown-out foreground, black background, unflattering and immediate.
Reach for it when you want a party or documentary feel. The ugliness is the entire point.
Magazine cover
magazine cover layout, space reserved for a masthead
Subject positioned low and centered with deliberate empty room at the top.
Reach for it when you actually intend to put type on it. Asking for the layout gets you the space for it.
Illustration and Render Styles
Non-photographic treatments. These replace the photographic vocabulary rather than stacking with it — asking for a watercolor with shallow depth of field mostly wastes both terms.
Anime
anime style, cel shaded
Flat color areas with hard shadow boundaries and clean linework.
Reach for it when you want the Japanese animation look. “Cel shaded” is what keeps it from drifting into generic digital art.
Comic book
comic book illustration, bold ink lines and halftone dots
Heavy black outlines, limited palette, printed dot texture.
Reach for it when you want Western comics. The halftone is what makes it read as printed rather than drawn.
Oil painting
oil painting, visible brushwork and impasto
Thick, directional paint with texture standing off the surface.
Reach for it when you want gravity and craft. “Impasto” is what gets you actual paint thickness.
Watercolor
watercolor, soft bleeding edges and visible paper texture
Transparent washes bleeding into each other with the paper showing through.
Reach for it when you want lightness and suggestion. The paper texture is what stops it looking like a filter.
Pencil sketch
pencil sketch, visible construction lines
Graphite linework with the underdrawing left in.
Reach for it when you want something that reads as in-progress rather than finished.
Claymation
claymation, stop-motion clay figures with visible fingerprints
Handmade forms with tool marks and slightly imperfect surfaces.
Reach for it when you want tactile charm. The fingerprints are what separate it from a smooth 3D render.
3D render
3D render, soft global illumination
Clean CG surfaces with physically accurate bounced light.
Reach for it when you want product-visualization polish.
Isometric
isometric view, clean geometric perspective
No perspective convergence — parallel lines stay parallel at a fixed angle.
Reach for it when diagramming a space, or making game-style art where every part reads equally.
Tilt-shift miniature
tilt-shift miniature effect, tiny diorama
Extreme selective focus that tricks the eye into reading a real scene as a model.
Reach for it when you want a city or crowd to look like a toy. Works best shot from above.
Papercraft
layered paper cut-out, handmade craft
Stacked flat shapes with visible edges and soft shadows between layers.
Reach for it when you want depth built from flat elements — good for title cards and covers.
Materials and Effects
Surface treatments and post effects. Best used last, applied to a scene the earlier terms have already established. One is usually enough.
Glass
translucent glass, refraction and caustics
Light bending through the material and pooling into bright patterns behind it.
Reach for it when the material is the subject. “Caustics” is the term that produces the light pooling.
Chrome
polished chrome, mirror-like metal
Fully reflective surface that renders its entire surroundings.
Reach for it when you want a hard, expensive, slightly unreal object. It reflects the scene, so the scene has to be worth reflecting.
Holographic
holographic iridescent surface, shifting rainbow
Color that changes across the surface depending on viewing angle.
Reach for it when you want a synthetic, contemporary, slightly toylike finish.
Double exposure
double exposure, two images blended into one frame
Two subjects overlaid so each shows through the other.
Reach for it when you want to state a relationship between two things without a caption.
Silhouette
silhouette against a bright background
Subject reduced to a black shape with all detail gone.
Reach for it when the outline alone tells the story. Requires a genuinely recognizable shape.
Reflection
mirrored reflection, symmetrical composition
The scene doubled in water, glass, or a mirror.
Reach for it when you want symmetry or a visual rhyme. Still water is the easiest surface to ask for.
Glitch
digital glitch, RGB channel split and datamosh
Displaced color channels and torn scan lines.
Reach for it when you want digital decay. The channel split is what makes it read as glitch rather than damage.
Quality and Finish
Output modifiers. These raise the ceiling but supply no direction of their own — a prompt made only of these produces a very sharp picture of nothing in particular. Add them last, and never more than one line's worth.
Detail
highly detailed, sharp focus, high resolution
Pushes fine detail and edge definition throughout the frame.
Reach for it when the result comes back soft or vague. Covers what “4K” and “8K” were reaching for, more reliably.
Photorealistic
photorealistic
Pushes toward camera output rather than illustration.
Reach for it when the result drifts toward digital painting when you wanted a photograph. Say “hyperrealistic” to push past what a camera would actually capture.
The Same Subject, Four Ways
Every one of these starts from the identical subject line. Only the modifiers change, and each version is a different photograph.
SUBJECT (unchanged in all four)
A weathered lighthouse on a rock shelf, waves breaking against the base.
1 — MONUMENTAL
Low camera angle looking up at the subject. Dramatic high-contrast
lighting, deep shadows. Stormy sky, heavy dark clouds. Highly
detailed, sharp focus, high resolution.
2 — SERENE
Wide shot, wide-angle lens. Golden hour light, warm low sun, long
shadows. Light mist. Shot on 35mm film, natural grain.
3 — LONELY
Aerial drone shot, high above the scene. Blue hour, cool twilight
just after sunset. Dark and moody, muted desaturated palette.
4 — ILLUSTRATED
Watercolor, soft bleeding edges and visible paper texture. Fog.
Minimalist, generous negative space, single focal point.
The one thing worth internalizing: version 4 drops every photographic term. Depth of field, film grain, and shutter speed mean nothing to a watercolor, and asking for them anyway is how you get a muddy hybrid that reads as neither. Once you choose illustration, the photographic vocabulary goes in the bin.
When It Comes Back Wrong
Flat and lifeless
You have no lighting term. Add one — golden hour or dramatic high-contrast will fix most of these on the first try.
Obviously AI
Too clean. Add film grain, a shallow depth of field, and haze in the air. Physical imperfection is what sells it.
Generic and averaged
Too many terms pulling against each other. Cut back to three and rebuild one at a time.
Right idea, wrong feeling
Change only the lighting term and regenerate. Lighting carries most of the emotional weight in any image.
Subject keeps getting lost
Your modifiers outweigh your subject description. Make the plain-language part longer and more specific.
Cluttered background
Add shallow depth of field, or studio lighting with a seamless backdrop to remove the environment entirely.
Coming from the image side? Lighting, atmosphere, setting, and style vocabulary carries over unchanged — the Image tab is still the reference for those. What follows covers only what video adds: movement, time, and the syntax differences between models.
What Video Adds
A still prompt describes a state. A video prompt describes a state changing, and that difference creates two new slots that did not exist before: what the subject does over the clip, and what the camera does while it happens. Kling calls the first Subject Movement, Veo calls it Action, Sora called it Actions — every vendor has the slot, and leaving it empty is the most common way to get an unnervingly frozen or randomly flailing result.
One camera move, one subject action. Sora's guidance, and it holds everywhere. Two moves in a five-second clip is more than the model can resolve, and you get neither.
Describe the action concretely, with an endpoint. “Takes four steps to the window, pauses, then pulls the curtain in the final second” beats “walks around.” Runway, Sora, and Kling all say some version of this.
Say something in the world moves. Drifting dust, wind in fabric, steam. Without secondary motion a clip reads as a photograph that happens to last five seconds.
Never phrase anything as a negative. Runway's docs state outright that negative phrasing “may produce unpredictable or even opposite results.” Say locked camera, the camera remains still — never no camera movement. This is the sharpest difference from image prompting, where negatives are routine.
Pick a model first — it changes the output syntax, not the vocabulary. Then describe the subject and the action, and add terms below.
Camera Movement
The axis that does not exist in still images, and the one most people leave out. In a photograph the camera has a position; in a clip it has a path. Every vendor guide treats this as its own slot. Pick exactly one — Sora's guidance puts it plainly, and it holds everywhere: one clear camera move and one clear subject action per shot.
Push in
the camera slowly pushes in toward the subject
MiniMax Hailuo bracket command: [Push in]
Camera physically travels closer — also called a dolly in. Perspective shifts as it moves, so the background relationship changes.
Reach for it when you want growing intimacy or rising tension. The most reliable move in the list — nearly every model handles it well.
Pull out
the camera slowly pulls back away from the subject
MiniMax Hailuo bracket command: [Pull out]
Camera retreats, progressively revealing the surroundings. Also called a dolly out or pull back.
Reach for it when you want a reveal, or an ending beat that leaves the subject small in its context.
Pan
the camera pans slowly to the right
MiniMax Hailuo bracket command: [Pan right]
Camera rotates left or right on a fixed point. The body does not move — only the head turns.
Reach for it when surveying a space from one vantage point. Name the direction or you get an arbitrary one.
Tilt
the camera tilts up
MiniMax Hailuo bracket command: [Tilt up]
Camera rotates up or down on a fixed point. The vertical counterpart to a pan.
Reach for it when revealing height — up a building, down a figure. Pair with a low angle for scale.
Truck
the camera trucks left, moving parallel to the subject
MiniMax Hailuo bracket command: [Truck left]
Whole camera slides sideways through space. Not a pan — the position changes, not just the angle.
Reach for it when you want the parallax of moving past something. This is the move people ask for when they say “pan” and get the wrong result.
Pedestal
the camera rises straight up
MiniMax Hailuo bracket command: [Pedestal up]
Camera moves vertically without tilting. Framing stays level throughout.
Reach for it when you want to change height without changing what the lens points at.
Crane
a crane shot rising high above the scene
Sweeping vertical move on an arm, usually combined with a slight tilt. Grander than a pedestal.
Reach for it when opening or closing a sequence. Veo's own example prompt uses one to reveal a canyon.
Orbit
the camera orbits around the subject
Circles the subject completely, keeping it centered.
Reach for it when showing an object from all sides, or building energy around a static figure.
Arc
the camera arcs in a curved path around the subject
A partial curve rather than a full circle. Runway draws this distinction explicitly; most film crews use the two words interchangeably.
Reach for it when you want some of an orbit's motion without the full rotation.
Tracking
the camera tracks alongside the subject as it moves
MiniMax Hailuo bracket command: [Tracking shot]
Camera travels with a moving subject, holding it in frame. Also called a following or dolly shot.
Reach for it when the subject is going somewhere and you want to stay with them.
Handheld
handheld camera with natural shake
Organic instability — small corrections, slight drift, imperfect framing.
Reach for it when you want documentary immediacy. Also the cheapest way to break the eerie smoothness of generated motion.
Steadicam
smooth steadicam movement following the subject
Stabilized motion that glides while walking. Handheld's controlled cousin.
Reach for it when you want movement without the shake — following someone through a space.
Locked off
locked camera, the camera remains completely still
MiniMax Hailuo bracket command: [Static shot]
No camera motion at all. The subject moves; the frame does not.
Reach for it when the action alone should carry the shot. Phrase it positively like this — Runway's docs warn that “no camera movement” can produce the opposite of what you meant.
Zoom
the lens zooms in
MiniMax Hailuo bracket command: [Zoom in]
Focal length changes. The camera does not move, so perspective stays flat and compresses.
Reach for it when you specifically want the flattened, slightly artificial look. Otherwise use push in, which is almost always better.
Crash zoom
fast crash zoom in
Violently quick zoom. Comedic or shocking depending on context.
Reach for it when you want a punch — a reaction, a punchline, a jump scare.
Whip pan
whip pan to the right, heavy motion blur
Extremely fast pan that smears the frame into streaks.
Reach for it when you want an in-camera transition between two things, or frantic energy.
Dolly zoom
dolly zoom, the background warping while the subject stays the same size
Camera pushes in while the lens zooms out (or the reverse). The subject holds; the world distorts behind it.
Reach for it when you want dread or revelation. A famous effect, and unreliable — expect several attempts.
Camera shake
the camera shakes
MiniMax Hailuo bracket command: [Shake]
Sharp jolts rather than handheld drift.
Reach for it when something impacts — an explosion, a footfall, a collision.
Shot Size and Angle
Where the shot starts. This vocabulary carries over almost unchanged from still photography, and it is corroborated across Runway, Veo, Kling, and Sora's guides. The difference in video is that framing is now a starting condition that the camera move then changes.
Establishing shot
wide establishing shot
Opens on the whole location before going in.
Reach for it when the viewer needs to understand the space before you show them details.
Wide shot
wide shot
Full subject with substantial surroundings visible.
Reach for it when the relationship between subject and place matters.
Medium shot
medium shot
Roughly waist up. The workhorse framing for people doing things.
Reach for it when you want both gesture and expression readable.
Close-up
close-up
Face or object fills most of the frame.
Reach for it when the emotion or the detail is the content of the shot.
Extreme close-up
extreme close-up
A fragment — eyes, hands, a texture.
Reach for it when you want intensity or abstraction. Very short clips suit this well.
Two-shot
two-shot
Two subjects framed together in one shot.
Reach for it when the relationship between two people is the subject.
Over the shoulder
over-the-shoulder shot
Foreground shoulder framing what the person faces.
Reach for it when showing someone looking at something, with depth in two planes.
Low angle
low angle
Camera below the subject, looking up.
Reach for it when you want the subject to feel powerful or looming.
High angle
high angle
Camera above the subject, looking down.
Reach for it when you want the subject diminished or observed.
Eye level
eye level
Neutral human height. No editorial opinion.
Reach for it when you want the viewer to read the scene without being told how to feel about it.
Aerial
aerial view high above the scene
Far overhead, landscape reading as terrain.
Reach for it when the scale of a place is the point.
POV
first-person POV
The frame is the subject's own eyes.
Reach for it when you want immersion rather than observation.
Subject Motion and Beats
The genuinely new slot. An image prompt describes a state; a video prompt describes a change of state. Kling calls this Subject Movement, Veo calls it Action, Sora calls it Actions — every vendor has the slot, and omitting it is the most common way to get an unnervingly static or randomly-flailing clip.
One clear action
the subject slowly turns to face the camera
A single, physically simple thing happening over the clip.
Reach for it when always. Sora's rule — one camera move, one subject action per shot — is the single most useful constraint in video prompting.
Counted beats
the subject takes four steps to the window, pauses, then pulls the curtain in the final second
Specifies how many, in what order, and when the last beat lands.
Reach for it when you need the timing to be right. Sora's guide contrasts this directly against vague verbs like “walks,” and the specificity is what makes it land.
Environmental motion
dust drifts through the air and fabric stirs in the wind
Secondary movement in the world that is not the subject.
Reach for it when a clip feels dead. Something always moves in a real scene, and this is what separates a video from a photograph that happens to last five seconds.
Enter frame
the subject walks into frame from the left
The shot begins empty and the subject arrives.
Reach for it when you want an arrival beat, or a clean cut point at the head of the clip.
Exit frame
the subject walks out of frame to the right
The subject leaves before the clip ends.
Reach for it when you want a departure, or a clean handoff into the next shot.
Continuous shot
continuous, seamless shot with no cuts
Explicitly forbids the model from inventing an edit mid-clip.
Reach for it when you are getting unexpected cuts. Runway's docs name this as the fix, along with allowing more duration.
Reveal
the camera move reveals what was hidden behind the foreground
Ties the camera path to an information change — something becomes visible that was not.
Reach for it when you want the movement to mean something rather than just look like movement.
Pacing and Timing
Clips are short — most models give you between four and fifteen seconds. Timing is therefore a budget, not a detail. Runway and Veo both document timestamp notation for placing beats inside a single generation, which is the closest thing to editorial control these tools offer.
Slow motion
in slow motion
Stretches the action across the clip.
Reach for it when the movement itself is the subject — impacts, splashes, fabric.
Natural speed
at natural speed, real time
Ordinary playback rate.
Reach for it when you are getting unwanted slow motion, which some models drift toward by default.
Time lapse
time lapse, clouds racing overhead
Compresses a long span into the clip length.
Reach for it when you want to show change over time. Standard film language rather than documented vendor vocabulary — worth a test generation before you rely on it.
Timestamped beats
[00:00-00:03] the subject stands still. [00:03-00:06] the subject turns and walks away
Assigns specific actions to specific time ranges inside one clip.
Reach for it when you need more than one beat in a single generation. Documented by both Runway and Veo; treat it as unsupported elsewhere until you have tested it.
Lens and Focus
Depth-of-field vocabulary is the most broadly corroborated technical language across vendor guides — it appears in Runway, Veo, Kling, and Sora. One term here is video-only: focus that changes during the shot.
Shallow focus
shallow depth of field, background falling out of focus
Subject sharp, background soft.
Reach for it when you want the subject isolated. Works the same way it does in stills.
Deep focus
deep focus, foreground and background both sharp
Everything in the frame holds detail.
Reach for it when the environment is as important as the subject.
Rack focus
rack focus from the foreground to the subject behind
Focus shifts between two planes during the shot. Has no equivalent in a still image.
Reach for it when you want to redirect attention without moving the camera. One of the few genuinely video-native techniques here.
Soft focus
soft focus, gentle diffusion
Overall softening across the whole frame.
Reach for it when you want a dreamlike or period feel.
Anamorphic
anamorphic lens, horizontal flares and oval bokeh
Widescreen character with streaking flares.
Reach for it when the clip should read as cinema rather than video.
Named focal length
shot on a 35mm spherical prime
Names actual lens hardware rather than describing the effect.
Reach for it when you know exactly what optical character you want. Sora's guide recommends this; treat it as advanced and unverified on other models.
Motion blur
natural motion blur on the moving subject
Per-frame blur consistent with the movement speed.
Reach for it when movement looks unnaturally crisp and strobe-like.
Audio and Dialogue
Only some models generate sound, and the syntax for it is strictly vendor-specific — these are Veo conventions unless noted. On a model without native audio these lines are wasted tokens at best and confusing instructions at worst.
Spoken dialogue
the woman says, “We have to leave now.”
Quotation marks mark spoken lines for lip-synced delivery.
Reach for it when you need someone to speak. Veo and the former Sora both used quoted dialogue. Keep it short — long speeches break sync.
Sound effects
SFX: distant thunder, footsteps on gravel
The SFX: prefix marks discrete sound events.
Reach for it when specific sounds matter. Veo-specific syntax.
Ambient noise
Ambient noise: quiet rain against a window
The Ambient noise: prefix sets the background bed.
Reach for it when you want a sonic environment rather than individual effects. Veo-specific syntax.
Suppress audio
no subtitles, no BGM
Excludes captions and background music.
Reach for it when a model is adding unwanted text or music. Documented for Seedance, which supports negative phrasing only for these two things.
One Shot, Three Models
The same intent, written for three models. The vocabulary is identical. Everything else is not.
INTENT
A lighthouse keeper turns from the railing and walks to the door.
Camera pushes in. Handheld. Storm outside.
RUNWAY (prose, iterative, no negatives ever)
A lighthouse keeper in a heavy coat stands at a rain-slick railing.
He turns and walks toward the door, pausing in the final second.
The camera slowly pushes in toward the subject. Handheld camera with
natural shake. Rain lashes the gantry. Continuous, seamless shot with
no cuts.
VEO (cinematography first, native audio)
Medium shot, handheld with natural shake, camera slowly pushing in.
A lighthouse keeper in a heavy coat at a rain-slick railing. He turns
and walks toward the door, pausing in the final second. Storm at
night, practical lamp overhead. Ambient noise: heavy rain on metal,
wind. SFX: a door latch clicking.
MINIMAX HAILUO (bracket commands, no negative field)
A lighthouse keeper in a heavy coat at a rain-slick railing turns and
walks toward the door. Handheld camera with natural shake. Rain
lashes the gantry.
[Push in]
What that comparison is really showing: the words “push in,” “handheld,” and the action description are the same in all three, because film vocabulary transfers. What does not transfer is everything structural — where cinematography goes in the sentence, whether audio has its own syntax, and whether the camera move is prose or a bracket token. Learn the vocabulary once; look up the syntax every time you switch models.
The Seven Models
Every claim below traces to the vendor's own documentation, linked on each card. Verified 8 September 2026. Treat version numbers, durations, and resolutions as the fastest-rotting information here — the prompt structures and syntax quirks last considerably longer.
Kling
- Version
- Kling 3.0 (launched Feb 2026); the 2.x line remains available via API
- Clip length
- Up to 15s on 3.0; 5s or 10s on 2.x
- Resolution
- 1080p video
- Prompt structure
- Subject + Subject Movement + Scene + (Camera Language + Lighting + Atmosphere)
- Negative prompt
- Kling exposes a real
negative_promptAPI field, defaulting toblur, distort, and low quality. Note that Kling's own consumer-facing prompt guide never mentions it. - Model-specific syntax
- Standard and Professional modes;
cfg_scale; first- and last-frame image inputs. The API also accepts structured camera parameters, while the consumer UI teaches camera moves as prose. - Worth knowing
- Kling's own docs warn that it is weak on counts (“10 puppies” drifts) and on complex physics such as a bouncing ball or a thrown object's arc. It also notes that words like “Oriental” or “Asia” bias the aesthetic.
Runway
- Version
- Gen-4.5 is the current flagship; Gen-4 and Gen-4 Turbo remain available
- Clip length
- 2–10s
- Resolution
- 720p at 24 or 25fps
- Prompt structure
- No fixed template. Runway recommends building iteratively: foundational motion, then subject motion, then camera motion, then scene motion, then style.
- Negative prompt
- Runway's documentation is unusually direct: negative phrasing “is not supported and may produce unpredictable or even opposite results.” Say
locked camera, the camera remains still, neverno camera movement. - Model-specific syntax
- None — plain descriptive language throughout. Style is set with ordinary words like
cinematic live-actionorstop motion. - Worth knowing
- Runway lists its own common mistakes: overly conceptual language, restating static image details instead of describing motion, conversational phrasing, and multi-scene prompts. Unexpected cuts usually mean you need more duration or an explicit continuity instruction.
Google Veo
- Version
- Veo 3.1, in Standard, Fast, and Lite variants
- Clip length
- 4, 6, or 8s
- Resolution
- 720p or 1080p
- Prompt structure
- Cinematography + Subject + Action + Context + Style & Ambiance
- Negative prompt
- Google recommends rephrasing exclusions as positive description —
a desolate landscape with no buildingsrather thanno man-made structures. Whether an API-level negative field exists is not stated either way in the guide. - Model-specific syntax
- Native synchronized audio. Dialogue goes in quotation marks;
SFX:andAmbient noise:prefix sound. Supports[MM:SS-MM:SS]timestamp ranges for multi-beat shots, and start-image plus end-image transitions. - Worth knowing
- All output carries SynthID watermarking. The Add/Remove Object feature still runs on the older Veo 2 and produces no audio.
MiniMax Hailuo
- Version
- MiniMax-Hailuo-2.3, 2.3-Fast, and Hailuo-02
- Clip length
- 6s or 10s (1080p caps at 6s)
- Resolution
- 768p default, 1080p available
- Prompt structure
- No formal template. Guidance is camera-command-centric.
- Negative prompt
- No negative-prompt field exists in the API schema. Describe what you want instead.
- Model-specific syntax
- Fifteen bracket commands inserted directly into the prompt string:
[Truck left/right],[Pan left/right],[Push in],[Pull out],[Pedestal up/down],[Tilt up/down],[Zoom in/out],[Shake],[Tracking shot],[Static shot]. Combine up to three inside one bracket; sequence them with “then.” Aprompt_optimizerflag rewrites your prompt automatically and is on by default. - Worth knowing
- Prompts cap at 2,000 characters. Free-form descriptions work, but the docs state the explicit commands are more accurate — so on this model, use them.
Luma
- Version
- Ray3.2 on the Agents/Platform API. The older Dream Machine API still exposes only Ray2 and Ray2 Flash — they are different endpoints with different schemas.
- Clip length
- 5s or 10s per generation; Modify Video V2 handles up to 20s at 1080p
- Resolution
- 360p to 1080p on Ray3.2; the legacy Ray2 endpoint also lists 4K
- Prompt structure
- No formal template published.
- Negative prompt
- No negative-prompt field exists on either the legacy or the current API.
- Model-specific syntax
- Keyframe control differs by API: the legacy endpoint uses
frame0/frame1, while Ray3.2 usesstart_frame/end_frameplus a keyframe array supporting up to 64 anchors. A/concepts/listendpoint returns 33 supported camera-motion strings (push_in,orbit_left,dolly_zoom,tiny_planetand so on) that you drop into the prose prompt. Video extension, clip interpolation, and a seamlessloopflag are all supported. - Worth knowing
- Check which API you are actually calling. Docs for the two are separate and the keyframe schemas are not interchangeable.
Pika
- Version
- Pika 2.5. There is no official 3.0 — pages claiming one are third-party.
- Clip length
- 5s or 10s for text- and image-to-video; Pikaframes reaches 20–25s
- Resolution
- 480p on free tiers, up to 1080p on paid
- Prompt structure
- Pika publishes no official prompt formula. The structure recommended on this page is a reasonable default, not a vendor position.
- Negative prompt
- A
negative_promptfield exists for text-to-video, image-to-video, Pikaframes, Pikadditions, and Pikaswaps — but not for Pikaffects, which take only an effect name, source media, and seed. - Model-specific syntax
- Named features rather than prompt syntax. Pikaffects apply one of a fixed list of physics effects (Melt, Explode, Inflate, Deflate, Squish, Crush, Levitate, Peel, Crumble and many more) — a closed keyword vocabulary, not description. Pikaframes animates between two to five ordered keyframes. Pikadditions inserts a subject into existing video; Pikaswaps replaces one.
- Worth knowing
- Because Pikaffects is a keyword picker, most of the vocabulary on this page does not apply to it. Use the term list for Pika's ordinary text- and image-to-video modes.
Seedance
- Version
- Dreamina Seedance 2.5
- Clip length
- 4–30s — by a wide margin the longest single generation here
- Resolution
- 480p, 720p, or 1080p. Not 4K — that belongs to the earlier Seedance 2.0, and the 4K figure quoted around the web for 2.5 refers to input reference images.
- Prompt structure
- Subject + Location + Event + Genre/Style + Camera movement
- Negative prompt
- No general negative-prompt parameter. Exclusion phrasing is documented as working for exactly two things:
no subtitlesandno BGM. - Model-specific syntax
- Assets are referenced inline as
@Image1or@video1, with roles set to first frame, last frame, reference image, reference video, or reference audio. Parameters includegenerate_audio,seed,watermark, and aspect ratios through 21:9. Supports up to 30 reference images. - Worth knowing
- The widely-circulated “six-part formula” is a third-party summary, not ByteDance's own wording. The official guide gives the five-part structure above.
On Sora: it is not on this list because it no longer exists. OpenAI discontinued the Sora web and app experiences on 26 April 2026, and the Sora API shuts down on 24 September 2026. Guides published before mid-2026 still lead with it — a good reason to check dates on anything you read about this field.
When the Clip Comes Back Wrong
Nothing really moves
No action term. Add a concrete subject action and some environmental motion — drifting dust, wind, steam.
It cuts unexpectedly
Add “continuous, seamless shot with no cuts,” and give it more duration. Runway's docs name both as the fix.
Camera does the wrong move
You probably asked for a pan and wanted a truck, or a zoom and wanted a push in. Check the camera movement cards — those two pairs are the most confused in the list.
Asked for stillness, got motion
You phrased it negatively. Say “locked camera, the camera remains still” instead of “no camera movement.”
Motion looks uncanny
Add handheld shake and natural motion blur. Generated motion is unnaturally smooth, and imperfection is what fixes it.
Too much crammed in
One camera move, one subject action. If you need two beats, use timestamps — and only on Runway or Veo, which document them.
Original Source
This page was previously built in Claude as an original local reference for the AI Helper Library. The image vocabulary began as a public social carousel that circulated these as slash commands (/goldenhour, /lowangle). No model parses those — the slash is cosmetic and the words do all the work — so this version uses the phrasing you would actually type, merges entries that were duplicates under different names, and adds what each term is for.
The video vocabulary is drawn from vendor prompting documentation: Runway, Google Veo, Kling, MiniMax, Luma, Pika, and ByteDance Seedance. Terms marked as coming from a single vendor may not behave the same elsewhere. Where a term is ordinary cinematography language rather than something a vendor coined, it is described as such.
Effects described here are general behavior; results vary by model, by version, and by how much of your prompt is competing for attention. Nothing on this page substitutes for a test generation.