How to Write a Text to Video Prompt That Works

A text to video model can only show what you ask for. If your prompt is vague, the clip will be vague too. Each run also costs credits. So it is worth taking a few minutes to get the prompt right before you press the button.

This guide shows a simple way to write prompts for our Text to Video tool. It is based on what the model makers say in their own guides. We link to each source, so you can check it yourself. We do not make up test scores or claims about which model is "best". Where a tip comes from our own site, we say so.

Why the prompt matters so much

A video model turns your words into moving pictures. It has no other clue about what you want. You can't point at the screen and say "no, a bit to the left". The prompt is the whole brief.

That is why the people who build these models all publish prompt guides. Google has one for Veo. Kuaishou has one for Kling. ByteDance and Alibaba show long sample prompts in their launch posts. Their advice is not the same in every detail. But the main ideas line up well. A good prompt says who is in the shot, what they do, where they are, how the camera sees it, and how it should feel.

Start with one scene

Most clips are short. On Everygen AI, the price list shows each model at a set length. Veo models are shown at 8 seconds. Grok Imagine models are shown at 6 seconds. Seedance, Kling and Wan models are shown at 5 seconds. You can change the length when a model allows it. Our site caps a single clip at 30 seconds.

A few seconds is enough for one clear moment. It is not enough for a full story with three parts.

So before you write, pick the one thing that should happen. Here is an example.

• Good: "A barista pours a heart into a latte in a sunny café."
• Too much: "A barista opens the café, serves ten guests, then locks up at night."

The second prompt has three scenes. The model has to squeeze them into a few seconds. The result is often a blur of half-finished ideas. If you need a sequence, make each shot on its own. Then join them in an editor.

Some newer models can do longer clips with more than one shot. ByteDance says Seedance 2.5 can make up to 30 seconds in one pass. Alibaba says Wan 3.0 can also reach 30 seconds. Kuaishou says Kling 3.0 can reach 15 seconds and plan more than one shot. Those are what the makers say their models can do. What you can pick on our site is what the generator shows you. Even with a long clip, one clear idea per shot still helps.

The five parts of a good prompt

Google's Veo 3.1 prompt guide gives a simple formula. It has five parts:

1. Cinematography. This is the camera work and the shot type.
2. Subject. This is the main person or thing.
3. Action. This is what the subject does.
4. Context. This is the place and what is around it.
5. Style and ambiance. This is the look, the mood and the light.

Google lists camera work first. Its guide calls the camera part "the most powerful tool" for tone and feeling. You can put the parts in any order that reads well to you. What matters is that each one is there.

Let's look at each part.

1. Camera and shot

Say how close the camera is. Say if it moves. Here are some common terms:

• Wide shot. Shows the whole place. Good for scale.
• Medium shot. Shows a person from the waist up. Good for talking.
• Close-up. Shows a face or a small object. Good for feeling and detail.
• Static shot. The camera does not move.
• Slow push-in. The camera moves slowly toward the subject.
• Pan. The camera turns left or right.
• Tracking shot. The camera follows the subject as it moves.
• Aerial view. The camera looks down from high up.

Google's guide lists moves like dolly, tracking, crane, aerial, slow pan and POV. Pick one move per shot. "Drone shot, close-up, handheld" pulls the model in three ways at once.

Only add a camera move if it helps. A calm product shot often looks best with a still camera.

2. Subject

Say exactly who or what is on screen. Use details you could check with your eyes.

• Weak: "a bike"
• Strong: "a red vintage bicycle with a wicker basket"

Colour, size, material and age all help. For a person, you can name their age range, clothes and hair. Keep it to one or two subjects at first. Each extra person or object is one more thing the model has to keep steady.

3. Action

Say what happens during the clip. Use one main verb. Some good ones are pours, turns, rises, drifts, opens, waves and smiles.

Make the action small and clear. "She turns and smiles at the camera" is easy to show in five seconds. "She thinks about her past" is hard to show at all, because thinking is not something we can see.

If you want a change over time, say it in order. For example: "The door opens. A cat walks in. It sits down." Short steps are easier for the model to follow than one long, tangled sentence.

4. Place

Say where the scene happens and what is nearby. Add the time of day and the weather if they matter.

• "on a wet stone street at dusk"
• "in a small kitchen with green tiles"
• "on a quiet beach at sunrise"

The place sets the stage. It also helps the model pick the right light.

5. Look, mood and light

Say how the clip should feel. Light is the fastest way to set a mood.

• "soft morning light"
• "warm lamp light, cosy mood"
• "neon signs, rain, night"
• "flat, grey, overcast light"

You can also name a style, such as "documentary", "stop-motion" or "anime". Use one style per prompt.

Putting it together

Here is a full prompt that uses all five parts:

Slow push-in. A red vintage bicycle with a wicker basket leans against a café window. Rain drips from the awning. A wet stone street at dusk. Warm light from the window, calm mood.

Look at how short it is. Each phrase does one job. There is a camera move, a subject, a small action (the rain), a place and a light.

A note on sound

Some makers say their models can make sound as well as pictures. Google says Veo 3.1 can make a soundtrack from your text. Its guide gives three tips:

• Put spoken lines in quotes. For example: A man says, "We're late."
• Name sound effects clearly. For example: "SFX: thunder in the distance."
• Describe the background sound. For example: "Ambient noise: the quiet hum of a fridge."

Kuaishou says Kling 3.0 has native audio in many languages. ByteDance says Seedance 2.0 makes audio and video together. xAI says Grok Imagine video models create clips with audio.

On Everygen AI, sound in clips is still in development. It is coming soon. For now, treat these as claims from the model makers about their own models, not as features of our tool.

Common mistakes to avoid

Piling on praise words. "Beautiful, stunning, amazing, ultra-detailed" adds little. Use those words for real details instead. Name the colour, the material or the time of day.

Mixed camera orders. "Close-up, wide shot, aerial" can't all happen at once. Pick one.

Too many people. Five people in five seconds is a lot to keep steady. Start with one or two.

Asking for text in the frame. Signs, labels and captions often come out wrong. If you need text, add it in an editor later.

Saying what you don't want. "No people" can still pull the model's focus toward people. Say what you do want instead, like "an empty street".

Long, winding sentences. Break them up. Short lines are easier for you to edit and easier for the model to follow.

Pick the frame first

Choose the frame shape before you write the camera line. It changes what fits in the shot.

• Tall (9:16) suits one person or object, for phone feeds.
• Wide (16:9) suits places, groups and two people talking.
• Square (1:1) sits in the middle.

In our generator you pick from the shapes the chosen model supports. There is also an Auto option. Some makers limit shapes. For example, Google's docs for Veo 3.1 Lite list 16:9 and 9:16.

Know the cost before you press go

Every clip uses credits. Our generator shows the exact cost before you send the job. The cost depends on three things.

The model. Each model has its own rate.

The resolution. Higher resolution costs more. For example, on our pricing page, a 5-second Kling 3.0 clip is 15 credits at 720P and 20 credits at 1080P.

The length. We scale credits by length and round up. A longer clip costs more.

A smart habit is to test cheap first. Try a new idea on a low resolution or a low-cost model. When the shot looks right, run the same prompt at the quality you need. For example, Wan 2.2 Ultra Fast is listed at 2 credits for a 5-second clip at 480P. That makes it a cheap way to check the framing and the motion.

If a job fails or the model provider turns it down, we return the credits for that run to your balance on their own. You don't need to ask.

Change one thing at a time

Your first clip will often be close but not quite right. That is normal. When it happens, change one thing and run it again.

• If the motion is wrong, only rewrite the action.
• If the look is wrong, only change the light or the place.
• If the framing is wrong, only change the camera line.

If you change everything at once, you won't know which change helped. Small steps teach you how each model reacts.

Keep a short list of prompts that worked well. Over time you will build your own set of camera moves, light words and subjects that suit the models you use most.

How the model makers write long prompts

It helps to see how the makers write. Here is what stands out in their own examples.

Google's DeepMind prompt page shows very long, rich prompts. They describe the camera, the place, the light and the sound in great detail. They show that Veo can take a lot of detail. You don't have to write that much. But it's a good sign that extra detail, when it is clear, is welcome.

ByteDance's Seedance 2.5 post shows prompts that split a clip into time blocks, like "0–4s" and "4–10s". That is a useful trick for longer clips. It tells the model what should happen and when.

Kuaishou's Kling 3.0 guide talks about multi-shot clips and keeping the same character across shots. If you plan more than one shot, describe each shot in order.

xAI's video docs explain something useful about Grok Imagine 1.5. For text to video, the model first makes a still image from your prompt. Then it animates that image. So for Grok Imagine, a clear picture of the first frame is a good place to start.

Three prompts to try

Copy one of these into the Text to Video tool. Then change it to fit your idea.

Product clip

Static macro shot. A glass perfume bottle on black marble. A single drop of water slides down the side. Soft side light, dark and elegant.

Nature

Slow pan from left to right. A field of tall grass at sunrise. A light breeze moves the grass in waves. Golden light, mist in the distance.

A small human moment

Medium shot. An old fisherman sits on a wooden pier, mending a net. He looks up and smiles at the camera. Grey, soft light, a quiet morning.

Each one has a camera line, a subject, one action, a place and a light. None of them asks for text in the frame. None has more than one person.

When to use an image instead

Sometimes words are not enough. Say you need your own product, your own face or a real place. Text alone can't copy it exactly.

In that case, use Image to Video. Your photo becomes the first frame. Your prompt then only has to describe the motion. Our image to video guide explains how to choose and prepare that photo.

If you don't have the right photo yet, you can make one first. Use Text to Image to create a starting frame. Or use Image to Image to edit a photo you already have. Then animate it.

Which model should you use?

Each model has its own rates, lengths and strengths. We cover this in a separate post: how to choose a video model. It sets out the facts from our price list and the claims from each maker, side by side.

Quick checklist

Before you press generate, run through this list:

• Is there only one scene?
• Did you name the subject clearly?
• Is there one main action?
• Did you say where it happens?
• Is there one camera move, or none?
• Did you set the light and mood?
• Did you pick the frame shape?
• Did you check the credit cost?

If you can say yes to each one, your prompt is ready.

Summary

• Write one scene per clip.
• Cover camera, subject, action, place and light.
• Keep camera orders simple and steady.
• Skip text in the frame. Add it later.
• Check the cost in the generator, and test cheap first.
• Change one thing per try.

Ready to start? Open Text to Video, paste one of the prompts above, and make it your own.

Sources

• Google Cloud: Ultimate prompting guide for Veo 3.1
• Google DeepMind: How to create effective prompts with Veo 3
• Google AI for Developers: Veo 3.1 Lite preview
• Kuaishou: Kling AI launches 3.0 model and Kling VIDEO 3.0 model guide
• ByteDance Seed: Seedance 2.0 official launch and Introducing Seedance 2.5
• Alibaba Cloud: Wan3.0 at general availability
• xAI: Video generation and Video overview
• Everygen AI: Pricing and credit rates