How to write an AI video prompt: structure and camera moves

The short answer

Write the prompt in English, and build it from five parts in this order: cinematography, subject, action, context, and style and ambiance. This is the structure Google’s official prompting guide for Veo 3.1 suggests, and you can use it in other models too, because it answers everything a video model needs to know. Plus four rules:

• Speech in quotation marks. A line a character says goes inside quotation marks, right after who says it
• Sound on its own. A sound effect goes after the word SFX, and background sound after the words Ambient noise
• One clear action. A clip of a few seconds has room for one moment, not a whole story
• Settings are chosen in the studio. Length, resolution and aspect ratio, such as 9:16 for a vertical clip, are set in the settings, not in the prompt

This is what a prompt built this way looks like. It is an example we wrote ourselves:

Medium shot, slow dolly in. A street musician with a worn acoustic guitar plays and smiles at the people passing by. A narrow old city street on a rainy evening, neon signs reflected on the wet stones. Warm cinematic light, nostalgic mood. SFX: soft rain on the pavement.

The five parts of a good prompt

Google’s guide defines a formula of five parts. Each part answers a different question, and together they give the model a full picture of the shot:

Join the parts in order, and you have the full prompt. You can leave out a part that does not matter to you, and the model will choose it on its own. Without the cinematography part, for example, the model decides by itself where to film from.

The difference in practice

Dialogue and sound in the prompt

In models that create sound together with the picture, the prompt can also describe what you hear. Google’s guide suggests three ways to write it, and these are its examples:

Which models create sound

• Sound created with the clip: the Veo 3.1 family, Seedance 2.5, Wan 3.0, LTX-2.5 Fast and Gemini Omni Flash 1.1. In models with a "Generate audio" switch, check that it is on
• Sound only when you turn it on: in Kling 3.0 sound is off by default, and without turning it on the character will not speak
• No sound: according to Luma, Ray 3.2 makes a silent clip, and for Hailuo 2.3 the maker lists no sound. With them you add voice and music separately

And what about Hebrew speech

No video model here officially supports Hebrew speech. In Kling 3.0, for example, Kling writes that dialogue in a language that is not on its official list is translated into English. So when a character has to speak Hebrew, create the voiceover separately in a voice model whose official language list includes Hebrew, such as Eleven v3, and add it to the clip in editing.

There is another way too: in Wan 2.7 you can upload the voiceover as an audio file, and according to Alibaba the model times the lip movement and the action to it. There is no official confirmation that this is accurate in Hebrew, so check the result. Every option for a Hebrew voiceover is in the guide Hebrew AI voiceover.

A camera move glossary in Hebrew and English

The cinematography part is written in the usual English filming terms. The terms in these tables come from Google’s official prompting guide for Veo 3.1, from the list of camera commands MiniMax documents and from the ready moves you can pick here in the LTX models, each with the Hebrew term and a short note on what you actually see.

Camera moves at a click

In some models you do not have to write the move at all. Seedance 1.5 Pro has a "Fixed camera" switch that keeps the camera in place. In the LTX models, such as LTX-2.5 Fast, you can pick a ready move in the "Camera motion" field: Move in, Move out, Move left, Move right, Rise, Lower, Static or Focus shift.

For Hailuo, MiniMax also documents camera commands inside square brackets, such as [Pan left] or [Push in]. We have not tested that form here, so it is worth writing the move in words as well.

Prompts when you start from an image

When you upload a start image, the model already sees the characters, the place and the colors. So the prompt focuses on what changes: who moves and how, where the camera goes and what you hear. When there is an end image too, describe the move that connects the two images. The full explanation, with examples, is in the guide how to turn an image into a video.

When you work with reference images, it helps to say in the prompt what each image stands for. In the example in Google’s guide, the prompt starts with the words Using the provided images for the detective, the woman, and the office setting, and only then describes the shot. How to keep the same character across several clips this way is explained in the guide how to keep the same character.

A prompt in Hebrew or in English

In English. Most makers do not mention Hebrew at all, and those that publish a prompt language list leave it out:

• For Veo 3.1 and Gemini Omni Flash 1.1, Google writes that English is fully supported, while other languages have not been evaluated and may work with varying results
• For the Wan models, such as Wan 3.0, Alibaba states that prompts are officially supported in Chinese and English
• For the LTX models, Lightricks describes a multilingual text encoder, but without a list of languages

The easy way is to think in Hebrew and send in English: phrase the idea in Hebrew, translate it into English, and check that the translation kept what matters, such as the camera move and what the character does. It also helps to keep the filming terms from the glossary above in English, since that is how they appear in Google’s guide.

The interface language does not change this: you can work in the studio in Hebrew and write the prompt in English. What works in Hebrew in every kind of model, from voiceover to text inside an image, is explained in the guide Hebrew in AI models.

Negative prompt and seed

Some models have a field called "Negative prompt", where you write what should not appear in the result. In that field, list only the things themselves, without "no" or "don’t", for example buildings, roads, text, since the model steers away from everything written there. In the main prompt, on the other hand, Google’s guide recommends describing the scene you want rather than putting what you do not want in general terms: for example a desolate landscape with no buildings or roads, instead of no man-made structures.

Every generation starts from a random starting point, so the same prompt gives a slightly different result each time. In the studio every generation gets a new starting point, so to compare versions keep the same start image or the same reference images and change one detail at a time. If you create through the API, MCP or CLI, you can also fix the starting point with a seed number in models that support it: when you keep the number fixed and change one word in the prompt, it is easier to see what the change did.

Several shots in one prompt

A short clip can include more than one shot. Google’s guide for Veo 3.1 shows how to split the prompt by time, with each segment getting a shot of its own. In its example an 8 second clip is split into four segments of 2 seconds. This is what such a prompt looks like, in an example we wrote with the same structure:

[00:00-00:02] Wide shot of a small bakery at dawn, warm light glowing in the window. [00:02-00:04] Close-up of hands dusting flour over fresh dough. SFX: the soft thud of dough on a wooden table. [00:04-00:06] Medium shot, the baker smiles and pulls a tray of bread out of the oven. [00:06-00:08] Slow push in on the golden loaves as steam rises. Ambient noise: a quiet street waking up.

Wan 2.6 also has a "Multi-shot" switch, and Alibaba describes it as a model that tells a short story in several shots. In Wan 2.7, according to Alibaba, you write the shot structure in words inside the prompt. When you need a longer video, the options are in the guide AI video longer than 8 seconds.

When the model does not do what you wrote

• The character does not speak. Check that the model creates sound and that its switch is on. In Kling 3.0 sound is off by default
• The model ignores part of the prompt. Narrow it down to one clear action, delete instructions that contradict each other, and write in English
• The result is far from what you wrote. Some models expand your prompt before generating, with a switch such as "Prompt enhancer" or "Prompt optimizer". When you want precise control, turn it off
• The camera moves when you did not ask it to. Write Static shot, or pick "Fixed camera" or the "Static" move in models that offer them
• The clip has the wrong shape. The aspect ratio is chosen in the settings, for example 9:16 for a vertical clip or 16:9 for a landscape one
• You need finer control. Kling 3.0 and Kling 2.6 also have a "Prompt adherence" field that sets how closely the model sticks to the prompt

And before the final version, test the prompt cheaply. In Veo 3.1 Lite the price is the same for every length, so test straight at 8 seconds, which cost ₪0.707.03 crVeo 3.1 Lite, 720p · 8s · with sound at 720p right now. In Kling 3.0, which is priced by the second, 3 seconds at 720p cost ₪0.999.85 crKling 3.0, 720p · 3s · no sound. Once the prompt works, run the final length and resolution.

Every model works from the same balance, so you can run the same prompt in two or three models and keep the best result. The exact price shows in the studio before every generation, and that is the amount that leaves your balance.

Questions and answers