AI Video Prompts

What a video prompt has to settle

A video model fills every gap you leave, and a few seconds of footage has a lot of them. If you don't decide what moves, where the camera goes and what happens first, the model decides at random, and differently each run. Six things close most of the gaps:

1. The opening picture. Subject, place and time of day, written like a still photo. Everything later grows out of that first frame.
2. One action and its pace. Steam drifts, a cat crosses the street, dough folds. Say how fast. «Slowly» is an instruction; «dynamic» isn't.
3. One camera move. Locked off, a slow push-in, a pan to the right, a crane rising. Two moves in one short clip usually fight each other.
4. Timing. The length in seconds, and the order of events if there's more than one: «for the first three seconds…, then…». Runway's own prompting guide also documents rough timestamps such as [00:01] and [00:03].
5. Light and look. Where the light comes from, the palette, and a genre the model knows: a nature documentary, a product ad, a hand-drawn anime.
6. Sound, if the tool makes it. One line that names what you hear and what you don't: «Sound: rain on glass, distant traffic, no music».

Put together, a pattern you can type from memory:

[Shot type] of [subject] in [place], [time of day].
[What moves], [how fast]. Then [the one change, if any].
Camera: [one move], smooth.
Light: [direction]. Look: [palette or genre].
Sound: [what you hear], no music.
[Seconds] seconds, [16:9 or 9:16]. No text, no logos, no watermark.

Every block on this page follows that order and closes with a «No…» line against captions, logos and stray signs. The blocks are written silent, so add the sound line yourself when your tool generates audio.

Text-to-video or image-to-video

The two modes need different prompts, and mixing them up is the quickest way to waste a render.

In text-to-video, the words are all the model has. You describe the picture and the motion: subject, setting, light, style, then what moves and what the camera does.

In image-to-video, your picture is the first frame. Runway's image-to-video guide puts it plainly: the image already sets the composition, subject, lighting and style, so the prompt should describe motion. Refer to things in general terms («the woman», «the cup») and spend your words on what moves, the camera and the timing.

The camera [slowly pushes in]. [The subject] [one action].
[What else moves in the scene]. Everything else stays as it is.

Two things about the source image matter more than the prompt. Blurry hands, warped faces and other flaws tend to get worse once the frame moves, so fix them first. And an image can carry implied motion, such as motion blur, a mid-stride pose or flying dust, which pulls the clip in that direction whatever you write. The Animate A Still Photo block is written for this mode and allows only motion that already belongs in the picture.

Which prompt for which clip

Pick by what the clip is for, then fill the brackets.

A weak prompt, written as a shot

This is the kind of request that gets typed into a video tool first:

Cozy bakery at sunrise, cinematic, aesthetic, 4K, trending.

It names a place and a mood and nothing that happens. Here's the same idea as a shot. We wrote it for this page to show the shape; it isn't output from any model:

Static wide shot of a small corner bakery seen from across an empty
street at dawn, blue light, wet cobblestones. For the first three
seconds nothing moves except a pigeon on the curb. Then the lights
inside the bakery switch on, one row at a time, and warm light spills
onto the pavement. The camera begins a slow push-in toward the window
and stops at the glass. Warm amber inside, cool blue outside, slight
film grain. Sound: quiet street ambience, one bird, the soft click of
the lights, no music. 8 seconds, 16:9. No people, no readable signage,
no text, no watermark.

Three changes carry it. The clip now has an event, the lights coming on, so the eight seconds have a beginning and an end. The timing is spelled out, and the camera only starts moving after the event, so two things never compete for the same second. And the sound line says what to hear and rules out music. Drop it in a tool that makes no audio.

Lines that fix a clip

Run the same prompt two or three times before you change it, since results vary a lot between runs. If the same fault keeps coming back, add the line on the right.

Where AI video goes wrong

The weak spots are shared across tools, and writing around them saves renders.

• Motion you didn't ask for. Video models are built to move things. A prompt that only describes a scene often gets a drifting camera or a sudden zoom. Say what moves, and say what stays still.
• Too many beats for the length. Three events in five seconds come out rushed and half-finished. Give each event two or three seconds, or make two clips and join them.
• Contact. Pouring, gripping and anything where two objects touch can melt or pass through each other. That's why several blocks here say «no hands».
• Words on screen. Signs, labels and titles come out garbled. Add text in an editor, where you can spell it.
• Speech. Tools that make dialogue can still miss lip movement or mangle a line. If you don't need a voice, write «no one speaks».
• Real people. Our text-to-video blocks keep faces out of frame or turned away, and none names a real person. Don't use these tools to put words or actions on someone who exists; Google's Gemini help page, for one, bars using its video feature to deceive, harass or harm others.

What the current tools take, as of October 2026

From each vendor's own documentation. These change often, so check your tool's controls before writing a length into the prompt.

What that changes in the prompt: Google's Veo documentation asks for dialogue in quotation marks and for sound effects and ambience described outright, so write «Sound: …» as its own sentence. Gemini Omni Flash, which Google's help center says now makes video in the Gemini app, edits a finished clip from a short instruction. Keep those edits brief and add «keep everything else the same». Kling's 15 seconds fit a sequence of beats; the other tools here want one or two. Runway's Gen-4.5 needs the Standard plan or higher, and its duration and frame rate are set in the controls, so delete those words from the prompt there. Our reviews of Google Flow, Kling, Runway, Luma and Pika cover prices, and the AI video generators page compares them. If you're coming from Sora, which OpenAI has shut down, the Sora prompts page keeps more shot types in the same shape, ready to paste into any of these.