Prompt library
Prompt to video AI: what to actually type
Prompt to video AI comes down to naming five things in one or two sentences: the subject, one action, the camera move, the lens and light, and the duration. Every prompt in the library below follows that shape, and each one sits beside the slot its clip fills once that run is published.
Most pages about prompting for video are lists of adjectives. Cinematic, hyper realistic, 8K, award winning. Those words do very little. What changes the output is structure, because these models are answering a description of a shot and a description of a shot has parts. Learn the parts and you stop guessing.
The five slots
| Slot | What goes in it | What happens if you leave it out |
|---|---|---|
| Subject | The thing the shot is about, described concretely | The model picks something generic and usually puts a person in it |
| Action | One thing that happens, in physical terms | You get a near still image with a drifting camera |
| Camera | A named move: pan, orbit, dolly, tilt, zoom, static | A slow push in, every single time |
| Lens and light | Focal length, time of day, direction of light | Flat, evenly lit output that looks like everyone else's |
| Duration | How many seconds, stated | The tool default, and pacing that does not match your edit |
The template
[subject] [one action], camera [named move] [direction and amount], [focal length], [light and time of day], [n] seconds.
That reads as an ordinary sentence when you fill it in, which is the point. Write it the way you would say it, not as a list of tags.
The library
Six prompts, each covering a different shape of shot. Copy the text, swap the subject, keep the structure.
Camera led: a slow dolly in
A vintage motorcycle parked in an alley, camera dollies in slowly from three metres to one metre, 50mm, late afternoon light, four seconds.
The move is named and the distance is given. Both matter. Without the distance the model picks its own amount and usually overshoots.
Veo 3 · 16:9 · clip not uploaded yet
Subject led: one clear action
A border collie shakes water off its coat in slow motion, side on, backlit, water droplets catch the sun, static camera, three seconds.
One action, described physically. Slow motion, backlight and a static camera do the rest. Adding a second action here would halve the hit rate.
Kling 2.5 · 16:9 · clip not uploaded yet
Product: orbit on a plain background
A glass perfume bottle on black stone, camera orbits ninety degrees left, single hard key light with a soft fill, no text or logos, five seconds.
The most reliable product shot there is. Note the explicit no text or logos, which stops the model inventing branding on the object.
Kling 2.5 · 16:9 · clip not uploaded yet
Abstract: texture with no structure
Ink diffusing into water, macro, high shutter speed, black background, no camera movement, four seconds.
Nothing here has an identity the model must preserve, which is why abstract prompts almost never fail. Useful as filler and as B-roll.
Wan 2.5 · 16:9 · clip not uploaded yet
Vertical: framing stated in the prompt
Vertical 9:16. A skateboarder rolls into frame from the left and stops dead centre, camera at ground level, harsh midday sun, three seconds.
Setting the output ratio is not enough. Saying vertical 9:16 in the prompt changes how the model composes the shot.
Seedance 2.0 · 9:16 · clip not uploaded yet
Negative: saying what must not happen
A wheat field in wind, wide shot, static camera, overcast light. No people, no text, no lens flare, no camera movement.
The last sentence is doing the work. Models over animate and add people by default, so exclusions are as important as descriptions.
Veo 3 · 16:9 · clip not uploaded yet
Negative prompting
Video models add things. People appear in empty streets, lens flares arrive uninvited, and a static shot slowly drifts. Exclusions fix all three and cost you nothing.
- No camera movement. The single most useful exclusion. A truly static shot has to be asked for.
- No text. Stops invented signage and packaging, which never survives a clip anyway.
- No people. Landscapes and interiors fill with figures otherwise.
- No lens flare. Models associate flare with cinematic and add it whether or not it fits.
- No slow motion. Worth stating when you want real time, because dramatic prompts tend to come back slowed down.
Mistakes that cost you takes
- Two actions in one prompt. She picks up the cup and walks to the window is two shots. Ask for one and the hit rate roughly doubles.
- Quality words instead of specifics. Masterpiece and 8K do nothing. Golden hour, 50mm, backlit all do something.
- Naming a style you cannot check. In the style of a named director produces an average of everything the model associates with that name.
- Asking for text in the frame. It will not be readable. Add it in post.
- Front loading adjectives. Put the subject first. The start of a prompt carries the most weight and adjectives are not the subject.
The modifier reference
These are the words that reliably change output, grouped by what they control. Pick one or two from each group rather than stacking five, because a prompt with fifteen modifiers loses the ones at the end.
| Group | Words | Effect |
|---|---|---|
| Camera move | pan, tilt, orbit, dolly in, dolly out, tracking, static, handheld, crane | The single highest impact group. Always pick one. The camera movement prompts page has the exact wording for each move. |
| Framing | wide, medium, close up, macro, over the shoulder, top down, low angle | Decides how much of the subject is in frame and how much can go wrong. |
| Lens | 24mm, 35mm, 50mm, 85mm, shallow depth of field, deep focus | Changes compression and background separation. 50mm is a safe default. |
| Light | golden hour, overcast, backlit, hard key light, soft window light, night, neon | The biggest driver of whether output looks expensive or generic. |
| Motion quality | slow, fast, sudden, drifting, real time, slow motion | Controls pace. Say real time if you do not want slow motion, because you often get it. |
| Atmosphere | haze, fog, dust in the air, rain, steam, smoke | Cheap depth. Adds a lot for one word and hides small artefacts. |
| Exclusions | no text, no people, no camera movement, no lens flare, no slow motion | Removes the defaults. As important as anything you ask for. |
Six more prompts, no clips attached
These follow the same five slot structure and cover situations the library above does not. They have no clip beside them because they have not been run for this page, which is why they are in a different section rather than mixed in with the ones that have.
- An empty theatre seen from the stage, camera tilts slowly upward to the balcony, 24mm, single work light, dust in the air, five seconds. No people.
- A hand places a ceramic bowl onto a wooden table, top down, static camera, 50mm, soft window light from the left, three seconds.
- Fog moving through pine trees on a hillside, wide, camera tracks slowly right, overcast light, real time, six seconds. No people, no animals.
- A subway train arrives at a platform, camera static at platform level, 35mm, fluorescent light, motion blur on the carriages, four seconds. No readable signage.
- Close up of a mechanical watch face, camera orbits slightly right, macro, hard key light with a soft fill, four seconds. No text on the dial.
- A curtain moving in an open window, medium shot, static camera, 85mm, late afternoon backlight, slow, five seconds. Nothing else in the room moves.
Iterating on a prompt that nearly works
Change one thing
Change the camera move or the light, not both. With randomness in every run you cannot attribute a change to two edits at once.
Run the same prompt twice before editing it
Half the time the prompt was fine and the run was unlucky. Two takes tell you whether you have a prompt problem or a seed problem.
Move the important words earlier
The start of a prompt carries the most weight. If the model is ignoring something, it is usually at the end of the sentence.
Cut before you add
When a prompt stops working after several edits, it is usually too long. Delete half of it and the result often improves.
Keep the version that worked
Save prompts in a file with the model name next to each one. Prompts are the asset you build in this category, and they outlive the clips.
Prompting for each job
The grammar is the same everywhere, but each job puts weight on a different slot. These pages go into the specifics.
- Image to video puts the weight on the camera slot, because the photo already answers subject and light.
- Vertical output puts the weight on framing, which has to be stated in the prompt as well as in the settings.
- B-roll wants prompts that are deliberately boring, so the insert does not pull attention off the voice.
- Script to video repeats the lens and light slots across every scene, which is what makes separate generations look like one shoot.
- Native audio adds a sixth consideration, because what you describe hearing affects what gets generated.
- The full build shows these prompts inside a finished video rather than on their own.
For the general capabilities and limits behind all of it, the overview is the video AI generator page.
Prompting questions
How long should a video prompt be?
One to three sentences, around twenty five to fifty words. Shorter and the model fills the gaps itself. Longer and it starts dropping details, usually the ones at the end, so put what matters most first.
What should a prompt to video AI prompt include?
Five things: the subject, one action, the camera move, the lens and lighting, and the duration. Prompts that miss the camera move are the most common reason output looks generic, because the model defaults to a slow push in.
Do negative prompts work in video generation?
Yes, and they matter more than in image generation because these models over animate. Writing no camera movement, no text, no people at the end of a prompt reliably removes those things.
Why does the same prompt give different results each time?
Generation is random by design and every run starts from different noise. Some tools expose a seed you can fix, which reproduces a result exactly. Without a seed, plan on running the same prompt several times and picking.
Should I write prompts in one sentence or in sections?
One flowing sentence with commas works better than a list of tags. These models were trained on captions and descriptions, not on keyword strings, so writing the way a person would describe the shot gets closer to what you meant.
Do prompts transfer between models?
The structure does, the details do not. A prompt tuned on one model usually gives a reasonable result on another, then needs its camera and lighting terms adjusted. Keep your prompts and re-tune them rather than starting over.