Write effective video prompts

Updated November 3, 2026. Based on the Video Overview and Video Generation pages from xAI and its image prompting guide: xAI asks you to describe the entire shot, including camera movement, light, and sound. The “two or three sentences max” rule from the previous version does not appear on any xAI page; in fact, the image guide recommends four to six sentences for a single image. This lesson also covers 1080p for 1.5 models, the nine aspect ratios, and the audio track generated by default. Sources: Video Overview · Video Generation · Image Prompting Guide.

The art of the video prompt

The quality of the generated video depends directly on the precision of your prompt, and a video prompt requires more than an image prompt. Where a static description suffices to produce an illustration, a video prompt must also carry the movement, temporality, and dynamics of the scene—and, since the video outputs with its soundtrack, what you hear in it. A prompt that perfectly describes a setting without ever saying what moves risks producing a nearly frozen shot, technically animated but visually lifeless. This lesson teaches you to structure your descriptions to get the best out of the Grok Imagine Video models.

The four components of a good description

An effective video prompt articulates four layers of information, and you work on each separately.

The first is the subject: the main element or elements of the scene, described with enough precision that the model does not have to invent. Writing “a bird” leaves the field wide open; writing “a hummingbird with iridescent blue and green feathers” locks in the appearance and makes your generations more consistent from one attempt to the next.

Next comes the action, the most decisive component in video. It is not enough to say that something moves; you must specify the movement, its direction, and its intensity. “A bird flying” is an intention, not a staging instruction. “A hummingbird feeding on a red hibiscus flower, wings beating rapidly” gives the model a gesture, a target, and a rhythm.

The third layer is the environment: the place, light, weather, and mood. A simple “in a garden” leaves the model to choose the time and season for you. “In a lush tropical garden, golden late afternoon light filtering through the foliage” sets the atmosphere and, incidentally, the direction of light on your subject.

Finally, the visual style, which determines the render: cinematic, documentary, animated. A phrase like “cinematic style, shallow depth of field, warm colors” establishes an advertising register, while “stylized 3D animation, pastel palette” takes you into a completely different universe with the same input scene.

Order matters. The xAI image prompting guide, written for still images but useful here, observes that a prompt opening with mood, film grain, or a list of prohibitions produces a polished scene where the subject is merely a background extra. So start with the subject and its action, and keep the style for the end.

Two layers specific to video: camera and sound

For text-to-video, xAI summarizes its instruction in one sentence: describe the shot, including camera movement, light, and sound. Two of these elements do not exist in still images.

The camera is controlled with filming vocabulary: static shot, slow zoom, tracking shot, drone view, timelapse. The examples in xAI documentation use them as is: a slow zoom on a runway podium, a camera gliding from a sunny door to a window, a drone shot between peaks, a timelapse of a flower blooming. If you say nothing, you leave this choice to the model.

Light benefits from being described as the xAI image guide does: its source, direction, and hardness, strong enough that the subject remains legible. “Hard afternoon sun from the left” tells the model what to illuminate; “nice light” tells it nothing.

Sound is part of the result: a generated video includes an audio track by default, with lip-synced speech on grok-imagine-video-1.5 and its -lite version. So describe the soundscape you want—rain on pavement, room murmur, soft music—or cut the track with "generate_audio": false if you will lay your own soundtrack in editing.

One last operational detail changes how you write: on 1.5 models, text-to-video first renders a still image from the prompt, then animates it. The starting scene must therefore read as a sharp image—subject, decor, light—and the movement as what happens next.

Three prompts, three styles

These layers combine differently depending on the clip’s destination. For advertising use, you care for material and light, and the action remains minimal but perfectly described:

A steaming cup of coffee on a rustic wooden table,
steam rising in gentle spirals. Natural side light from
the left, very slow zoom toward the cup, shallow depth of
field, advertising style. Muted coffee shop ambiance.

For data visualization intended for a presentation, movement becomes structural—it carries the information—and the style becomes deliberately sober. This video will play under live commentary: generate it mute, with "generate_audio": false.

A slowly rotating globe, bright lines
connecting major cities. Dark background, static camera, style
clean modern infographic.

In creative animation finally, the subject can be abstract, provided the transformation is described precisely:

Abstract geometric shapes transforming fluidly
into one another, vivid color palette on black background.
Camera slowly orbits the shapes, contemporary motion design
style, synth pads.

Take a moment to identify, in each of these three blocks, where the subject, action, environment, and style are, then the camera and sound. This reading exercise makes writing natural afterward.

Matching parameters to content

The prompt does not work alone: duration, ratio, and resolution must answer to it. Duration ranges from 1 to 15 seconds, 8 by default. A duration of 1 to 3 seconds suits a single action, a transition, or a GIF-style render; from 4 to 7 seconds, you get a short scene with a beginning and an end; between 8 and 15 seconds, you enter narrative sequence with evolution. The longer the video, the more generation time it takes, and price is counted per second: start short and increase progressively.

Choose the ratio based on the final destination. The API offers nine: 16:9 (default) for widescreen, web, and presentations, 9:16 for mobile, stories, and Reels, 1:1 for social posts and thumbnails, 4:3 and 3:4 for presentations and portraits, 3:2 and 2:3 for a photographic render, 21:9 for an ultra-wide cinema format, 5:2 for a banner or site header. In image-to-video, this choice disappears: the output follows the image ratio.

As for resolution, keep 480p for previews and drafts—this is the default value, and render is faster—then move up to 720p or 1080p when the clip goes into production. 1080p exists only on 1.5 models: native on grok-imagine-video-1.5, obtained by upsampling 720p on its -lite version, while grok-imagine-video stops at 720p. For simple videos, xAI recommends grok-imagine-video-1.5-lite, the cheapest per second, and reserves grok-imagine-video-1.5 for reference images, voices, imposed images, and native 1080p.

What makes a prompt fail

Five pitfalls recur often. The first is the prompt too vague: “a beautiful scene” contains no usable information. The second is the prompt that wants to tell everything—multiple subjects, multiple locations, multiple scenes in a few seconds. For a multi-shot story, follow the Storyboard to Short Film example from the xAI cookbook: it breaks the narrative into four shots, generates each separately, with its scene and camera movement, then assembles them in editing. The third, already mentioned, is the absence of described movement, which yields a static shot. The fourth lies in logical conflicts: asking for “a cat flying in water” places the model before a contradiction, and the result is likely to be incoherent. The fifth is the list of negatives—“no cars, no text, no crowd”. The xAI image guide notes that a long list of elements to avoid is easily ignored and fills the prompt with objects that end up appearing: describe what you want to see instead.

As for length, xAI documentation gives no ideal measure for video, and does not quantify the limit beyond which a too-long prompt is refused (invalid_argument error). Its image guide suggests four to six sentences for a single image: the subject and its details, then the decor, light, and framing. For a video shot, start from this structure and add action, camera, and sound.

To practice, write a full prompt for three different destinations: a background video for a restaurant website, a 3-second clip for an Instagram story for a clothing brand, and an abstract animation for a technical presentation. On each, verify that the four layers are present, that camera and sound are described—or the audio track cut—and that duration and ratio match the intended use.

Key takeaways

• A good video prompt contains subject, action, environment, and style, starting with the subject
• xAI also asks you to describe camera movement, light, and sound
• The audio track is generated by default; "generate_audio": false gives a mute video
• Adapt duration (1 to 15 s, 8 default), ratio (nine formats), and resolution (up to 1080p on 1.5 models) to final use
• A multi-shot story is generated shot by shot, then edited
• Describe what you want to see rather than what you do not want