How to Write AI Video Prompts: A Shot-by-Shot Guide (2026)

Most bad video prompts are good image prompts with a verb bolted on. "A woman in a red coat on a rainy Tokyo street, cinematic, moody lighting, walking" describes a lovely frame. It says nothing about the next six seconds: where she walks, what the camera does, whether the shot cuts, or what we hear. The model invents all of that, and its guesses rarely match the clip in your head.

We built the PromptWizz video prompt optimizer, so we spend a lot of time in the official prompting docs for Gemini Omni, Veo, Kling and Runway. They agree on more than they disagree. This guide covers the shared craft first, then what changes per model. Model details were checked against official documentation on November 3, 2026.

1. Write One Shot Before You Write a Story

Google's best-practices guide for Gemini Omni and Veo puts it simply: for short videos, dedicate each prompt to a single focused moment. Most clips run somewhere between a few seconds and fifteen, which is enough time for one clear action, not a plot.

So start with a single shot, in this order:

1. Subject. Who or what stays in view. Be concrete about the details that must stay consistent: the coat, the dog's color, the logo on the mug.
2. Action. One physical thing that happens and can finish inside the clip. "Pours coffee into a mug" works. "Has a realization about her career" doesn't, because a camera can't see it.
3. Setting and light. Where, what time of day, what the light is doing.
4. Camera. One main camera behavior for the shot (more on this in the next section).
5. Style. Film stock, animation style, color grade, genre.

Google's Veo guide lists nearly the same elements, with camera, composition, lens and ambiance as optional extras. You don't need every element every time. You do need the first two.

If you want more than one shot, label them. "Shot 1:" and "Shot 2:" make it clear in plain text that you're asking for a cut, and they push you to keep the subject's description identical across both. (Some tools also have a separate multi-shot setting; see section 6.) And if you want no cuts, say so. The Gemini API docs note that Gemini Omni tends to produce multiple shots unless you ask for "a single continuous shot" or "no scene cuts."

2. Separate What the Camera Does From What the Subject Does

This is where most video prompts fall apart. "Dynamic shot of a skateboarder doing tricks with lots of movement" asks for movement but never says whose. Is the camera tracking the skater, circling, or locked off while the skater flies through frame? The model picks one, often the wrong one.

Give each kind of motion its own phrase:

• Camera: what the viewer's point of view does.
• Subject: what the people, animals or objects do.
• Environment: what the world does around them (rain, traffic, leaves, steam).
• Stillness: what should not move. This line is underrated. "The subject holds still while the background traffic streams past" is far clearer than leaving it to chance.

Google Cloud's video prompt guide lists the camera vocabulary these models are trained to read:

That last row matters. Google's guide calls out that a zoom is different from a dolly because the camera itself doesn't move, so "slow push-in" and "slow zoom" look different. A push-in moves the camera closer, so the relationship between foreground and background shifts as it travels. A zoom magnifies the frame from the same spot, so the subject gets bigger but that relationship stays fixed. Perspective comes from where the camera is, not from the lens setting.

Then add a pacing word. "Slow," "gentle," "sudden" and "steady" do more work than a stack of three camera commands. One camera move per shot, with a speed, is a strong default.

A clear video prompt names one camera move, one subject action, and one thing that stays still.

3. Fit the Action to the Clip Length and Frame

Every tool has a supported duration, and your action has to fit inside it. A dog running across a field, stopping, sniffing a flower, then looking up at the camera is four beats. In a five-second clip, something gets rushed or dropped.

Current ranges, from each vendor's documentation:

• Runway Gen-4.5: 2 to 10 seconds, for both text-to-video and image-to-video (Runway API changelog).
• Kling 4.0 on Runway's API: 3 to 30 seconds, at 720p or 1080p.
• Kling 3.0 on Runway's API (still available): 3 to 15 seconds. Its multi-shot option, a separate API setting, takes up to six segments inside that total.
• Gemini Omni Flash: our optimizer's Omni profile works in a 3 to 10 second range, and Runway's API lists the same range for Omni. Google's own docs describe extending a clip in 10-second increments up to 40 seconds total.

When timing inside the clip matters, say it. The Gemini Omni docs show two styles that work: natural language ("After 3 seconds, a woman enters the scene") or timecodes like [0-3s] A person is walking. Google's prompt guide also flags pacing words such as slow-motion, fast-paced action and time-lapse, and reminds you to keep clip length in mind for events that would normally take longer.

Aspect ratio is the other half. Gemini Omni, Veo and our Runway Gen-4.5 text-to-video profile offer 16:9 and 9:16. Our Runway Gen-4.5 image-to-video profile adds 1:1, 4:3, 3:4 and 21:9, and Kling 4.0 offers 16:9, 9:16, 1:1 and 21:9. Pick it based on where the clip will be watched, then compose for it: a vertical 9:16 frame wants a subject that fills the height, not a tiny figure in a wide landscape.

4. Prompt Sound and Dialogue Deliberately

Several current models generate sound with the picture, which makes the audio part of your prompt rather than an afterthought.

Gemini Omni Flash tries to generate a fitting audio track by default, according to Google's docs, and you steer it in the prompt. Google's examples use a separate line such as "Sound design: gentle breeze, distant bird chirps." Because there's no separate negative-prompt setting for Omni in the Gemini API, exclusions go in the prompt too: "No dialogue" or "No extra sound effects."

Kling 3.0 on Runway's API generates audio by default, with a switch to turn it off. Kling 4.0 on Runway's API always generates audio; there's no off switch, so describe the sound you want rather than hoping for silence. Kling's prompting is strongest when dialogue is anchored to a character. fal's Kling 3.0 guide (a third-party host, so treat it as practical advice rather than spec) recommends giving each character a unique, consistent label, avoiding pronouns, and describing the character's action before their line.

Runway Gen-4.5, in both text-to-video and image-to-video, is different in our optimizer: we treat requested sound as a separate post-production plan, not as part of the Runway prompt.

How to write the dialogue itself depends on the doc you read. The Gemini API's Veo page shows speech in quotation marks. Google Cloud's best-practices page for Omni and Veo recommends a colon after the speaker's action and no quotation marks, so the model doesn't render the words as on-screen text:

The barista slides the cup across the counter and says: Careful, it's hot.

If your clips come back with stray captions, switch to the colon style. Google's guide also suggests giving recurring characters a name and a specific voice style, so the voice stays consistent across clips.

Write sound as its own sentence, and tie every spoken line to a named character doing something.

5. Text-to-Video vs Image-to-Video

In text-to-video, your prompt builds the whole scene: subject, setting, light, style and motion. Everything in sections 1 to 4 applies.

In image-to-video, the starting frame already answers most of those questions. Google's best-practices guide is direct about it: your source image provides the subject, scene and style, so focus the prompt on the motion you want. Runway's help center says the same for Gen-4.5 image-to-video. Re-describing the photo wastes words and can pull the model away from what's actually in the frame.

A good image-to-video prompt reads like a director's note to a camera operator:

The camera slowly pushes toward the mug. Steam rises and curls to the
left. The background stays still. Soft, steady pace.

Two cautions. First, only describe motion the image can support: if it shows a closed door, "the door opens and a dog runs in" asks the model to invent a dog it has never seen. Second, our optimizer can't see your image. With the Runway image-to-video profile you upload the first frame in Runway, so review any suggested motion that depends on something off-frame.

Gemini Omni also accepts reference images and a first-and-last-frame pair, which Google describes as generating continuous video between two keyframes. That's useful for loops and camera orbits, where you know exactly how the clip should start and end.

6. What Changes From Model to Model

The craft above carries across tools. These are the differences worth knowing, as of November 3, 2026:

When you continue a Gemini Omni clip, Google's own tip is to start the extension prompt with "Continue the video" or "Continue the scene," and to end transitions with "One continuous shot, no jump cuts."

7. Three Rough Prompts We Ran Through the Optimizer

Here are three real runs through the PromptWizz video optimizer, one per model profile, each starting from the kind of prompt people actually type. Below each is what the optimizer returned, unedited: the shot prompt, the audio block where the profile includes one, and the suggested settings. The scores come from PromptWizz's own video rubric (stable subjects, ordered action, camera direction, timing, continuity, audio intent and workflow fit). They rate the prompt, not the final video, and we didn't render these in the target tools.

Run 1: Text-to-video for Gemini Omni Flash

Run on November 3, 2026 with the Gemini Omni Flash profile.

a woman walking through a busy night market, cinematic, lots of camera movement, sounds of the market

Shot 1: A single continuous shot in a busy night market. A woman walks forward through the crowd at a steady pace, moving from the left edge of the frame toward the right. Medium-wide framing at eye level. The camera tracks alongside her in a smooth lateral dolly move, keeping her in the center third while vendor stalls, hanging string lights, and passing shoppers slide through the foreground and background. Warm practical lighting from stall lamps and neon signs with soft shadows. Cinematic, shallow depth of field with moderate bokeh in the lights. Pacing is unhurried, with the crowd moving around her.

Audio block returned with it:

SFX: layered market sounds including vendor calls, sizzling food stalls, footsteps, and distant chatter. Ambience: busy open-air night market crowd with occasional music from nearby stalls.

Suggested settings: duration 6 seconds, aspect ratio 16:9.

PromptWizz score: 40 → 80 (free Simple depth, run November 3, 2026)

• Replaced "lots of camera movement" with one move: a smooth lateral dolly that tracks alongside her.
• Stated "a single continuous shot" and gave the woman a direction and pace: left to right, at a steady walk.
• Moved "sounds of the market" into a separate audio block of effects and ambience, and suggested 6 seconds at 16:9.

Against sections 1 and 2, the cuts are settled (a single continuous shot) and there's one camera move at an unhurried pace. Read the two motion lines together, though. She walks from the left edge toward the right while the camera tracks alongside, "keeping her in the center third." Both can be true if she enters on the left and the camera then picks her up, but the prompt doesn't say so, and a model could resolve it either way. If you care, keep one: either she crosses the frame, or the camera keeps her centered. The subject has no fixed details yet. The optimizer's notes say it didn't invent the woman's appearance, and its tips suggest describing her clothing if you want her to look the same from one generation to the next.

Run 2: Image-to-video for Runway Gen-4.5

Run on November 3, 2026 with the Runway Gen-4.5 image-to-video profile.

make my photo of a coffee cup on a cafe table come alive, steam moving and the camera zooming in, cozy vibe

Shot 1: Single continuous shot, image-to-video from the user's starting photo of a coffee cup on a cafe table. The cup, saucer, table, and surrounding background stay in their original positions, colors, and textures. Thin wisps of steam rise gently and drift slowly upward from the surface of the coffee, curling softly in the warm light. The camera begins at the starting framing and performs one slow, steady push-in toward the cup, closing the frame around the rim and steam over the full duration. Pacing is calm and unhurried, with the steam movement and the push-in progressing together. Warm, soft, cozy ambient lighting remains consistent throughout.

Suggested settings: duration 5 seconds, aspect ratio 16:9.

PromptWizz score: 40 → 86 (free Simple depth, run November 3, 2026)

• Replaced "the camera zooming in" with "one slow, steady push-in toward the cup," a camera move rather than a lens zoom.
• Spelled out the steam: thin wisps that rise, drift upward and curl in the warm light.
• Added a continuity line so the cup, saucer, table and background keep their positions, colors and textures.
• Returned no audio block. Its notes say any ambient sound should be planned as a separate post-production step.

This one is a test of section 5. The optimizer never saw the photo, and its notes say so. Read the continuity line closely: it mentions a saucer, which the rough prompt never did. If your photo has no saucer, delete the word before you paste the prompt into Runway. Also note it changed "zooming in" to a push-in. As section 2 explains, that's a camera move with a different feel from a lens zoom, so change it back if a zoom is what you wanted.

Run 3: Dialogue for Kling 3.0

Run on November 3, 2026 with the Kling 3.0 profile, which is still available alongside Kling 4.0.

two friends arguing in a kitchen about who ate the last slice of cake, make it funny, with dialogue

Shot 1: A bright, cluttered home kitchen in the late afternoon. Two friends, Maya (wearing a yellow sweater) and Jordan (wearing a gray hoodie), stand on opposite sides of a kitchen island. On the counter sits a cake with a single slice missing. The camera holds a medium two-shot at eye level, locked off, as Maya points an accusing finger at Jordan. Jordan lifts both hands in mock innocence and leans back slightly. Pacing is brisk with comedic beats, and the cake, kitchen, and both characters' clothing stay consistent throughout.

Shot 2: Tight close-up on Jordan's face as they glance sideways toward the cake, then back to Maya with an exaggerated guilty grin. The camera performs a slow push-in over three seconds to emphasize the reaction. Maya crosses her arms and leans into frame.

Audio block returned with it:

Maya: "[Add exact line here]"
Jordan: "[Add exact line here]"
SFX: light kitchen ambience, faint clink of a fork against a plate.

Suggested settings: duration 8 seconds, aspect ratio 16:9, audio on.

PromptWizz score: 35 → 84 (free Simple depth, run November 3, 2026)

• Split the idea into two labeled shots: a locked medium two-shot, then a close-up with a slow push-in.
• Named the friends (Maya and Jordan) and fixed their clothing, so they stay consistent across shots.
• Left the dialogue as placeholders under each speaker label instead of inventing lines, and suggested 8 seconds with audio on.

Check the premise first. The rough prompt is about who ate the last slice, but Shot 1 describes "a cake with a single slice missing," which is close to the opposite. If the joke depends on an empty plate, fix that line before you generate. On section 4's checks, every line already belongs to a labeled character, but the lines are blank: the optimizer won't write dialogue for you. When you add them, keep them short enough for 8 seconds, and tie each one to the shot where that character is acting, so the action comes before the speech.

A Video Prompt Template You Can Copy

Shot 1 (single continuous shot, no cuts):
Subject: [who or what stays in frame, with fixed details]
Action: [one physical action that finishes within the clip]
Setting: [place, time of day, light]
Camera: [one move + speed, e.g. slow dolly toward the subject]
Still: [what does not move]
Style: [look, color, genre]
Sound: [ambience and effects; dialogue as Name says: line]
Settings: [duration] seconds, [aspect ratio]

For image-to-video, delete Subject, Setting and Style and keep the rest, with one exception: keep the Sound line in the prompt only if your tool generates audio. For Runway Gen-4.5 image-to-video, our optimizer treats sound as a separate post-production plan, so move that line to your edit notes.

Turn a rough idea into a shot plan. Paste your idea into the free AI video prompt generator, pick Gemini Omni Flash, Kling 4.0, Kling 3.0, Runway Gen-4.5 text to video, Runway Gen-4.5 image to video or the generic profile, and get a shot prompt, an audio plan where it applies, and suggested settings. It writes the prompt; you create the video in your own tool.

Keep Reading

• How to Prompt Each AI Differently
• How to Prompt Gemini
• Midjourney Prompt Guide

Sources

• Google, Gemini API: Generate and edit videos with Gemini Omni Flash. https://ai.google.dev/gemini-api/docs/omni
• Google, Gemini API: Veo video generation (model status and prompt guide). https://ai.google.dev/gemini-api/docs/veo
• Google, Gemini API models list. https://ai.google.dev/gemini-api/docs/models
• Google Cloud, Video generation prompt guide (Gemini Omni Flash and Veo). https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/video-gen-prompt-guide
• Google Cloud, Best practices for generating videos. https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/video/best-practice
• Google, Introducing Gemini Omni. https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-omni/
• Google, Gemini Omni 1.1 Flash lets you build with more control (August 27, 2026). https://blog.google/innovation-and-ai/technology/developers-tools/build-with-gemini-omni-1-1-flash/
• Runway API changelog (Gen-4.5, Kling 3.0, Kling 4.0, Gemini Omni Flash). https://docs.dev.runwayml.com/api-details/api_changelog/
• Runway Help Center, Creating with Gen-4.5. https://help.runwayml.com/hc/en-us/articles/46974685288467-Creating-with-Gen-4-5
• Runway Help Center, Text to Video Prompting Guide. https://help.runwayml.com/hc/en-us/articles/42460036199443-Text-to-Video-Prompting-Guide
• fal, Kling 3.0 Prompting Guide (third-party host). https://blog.fal.ai/kling-3-0-prompting-guide
• Sony, Focal length, angle of view and perspective. https://www.sony.com/en-eg/electronics/focal-length-angle-of-view-perspective
• OpenAI Help Center, What to know about the Sora discontinuation. https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation