I used to treat text-to-video like a slot machine. I typed an idea, pressed generate, and hoped for something cinematic. Sometimes it worked. More often it gave me a pretty frame with no story. The fix was not a longer prompt. It was a clearer brief.

Now I use text-to-video when I need atmosphere, motion, or an invented scene. I use image-to-video when the subject must stay recognizable. This guide is the process I follow, including the mistakes that taught me where the model needs help.

Start with one sentence, not a paragraph

Before prompting, I write the video idea in one sentence. For example, a paper boat crosses a puddle while rain stops and sunlight appears. That sentence has a subject, an action, and a change. If I cannot write it, the idea is not ready.

Then I decide the duration. A three to five second clip should hold one action. A six to eight second clip can hold a small turn or reveal. Anything longer usually needs several shots edited together.

Build the prompt in five parts

My working structure is simple.

Subject + action + camera + environment + style or timing

Here is a prompt I would actually run.

A paper boat crosses a rain puddle, low camera close to the water, city street after rain, soft sunlight breaks through clouds, shallow depth of field

The subject is clear. The action is visible. The camera is specific. The environment is real. The style is implied through light and depth of field rather than vague adjectives.

One action per clip

This is the rule I break most often and regret every time. A single clip cannot gracefully show a character entering, finding an object, reacting, and leaving. It can usually show one of those moments well.

So I split a story into shots. Shot one shows the environment. Shot two shows the action. Shot three shows the reaction. In the edit, they feel like a complete scene. In the prompt, each one stays manageable.

Use camera language with intention

Cinematic does not tell the model what to do. Slow push-in does. Handheld does. Low tracking shot does. I choose camera movement based on meaning.

Static camera makes the subject carry attention. Slow push-in adds intimacy. Tracking follows motion. A slight handheld feel can make social content feel natural. A wide locked frame works better for atmosphere.

I also specify what should not move when necessary. For product-like scenes, I write static background or camera holds while steam moves. For a character, I might write preserve facial features if the face matters.

Name specific places, light, and materials

I stopped using generic words like beautiful or high quality. Instead, I name the light and place. Overcast morning, golden hour backlight, fluorescent convenience store, wet stone, linen curtain, and dust in sunbeam all give the model something concrete.

This changed the results more than adding more adjectives. The model does not know what premium means to me. It does understand warm morning light or matte ceramic.

Draft cheap, then upgrade

I start with a free or low-cost model to test framing. Agnes Video V2 is useful for this because I only need to know whether the composition and motion idea are right. Once the draft works, I move to a production model.

Seedance 2.5 is good for expressive motion. Kling v3 Omni is more controlled and supports native audio when needed. Veo 3.1 is the premium cinematic option. The right choice depends on whether I need energy, stability, or rich visual quality.

Choose aspect ratio by destination

If the video is for TikTok, Reels, or Shorts, I start in 9:16. For website heroes or YouTube, 16:9 is better. For feeds and marketplaces, 1:1 can be safer. Cropping later is possible, but a prompt that already fits the frame gives better composition.

For vertical, I keep the subject near the center and leave space for captions. For widescreen, I let the environment breathe. This is a small decision that saves a lot of re-cropping.

My failure checklist

When the output is wrong, I check these first. Is there more than one action? Is the subject unclear? Did I ask for text inside the video? Is the camera doing too much? Is the duration too short? Are two prompt instructions contradicting each other?

Then I fix one thing and regenerate. If the subject drifts, I describe it more specifically or switch to image-to-video. If the motion is chaotic, I simplify the action. If the scene feels empty, I add one concrete environmental detail.

When image-to-video is better

Text-to-video invents. Image-to-video preserves. If I need a real product, a specific person, a known pet, or an old photograph, I do not rely on text alone. I use the image as the anchor and describe only the motion.

For invented worlds, product mockups, and atmosphere, text-to-video is liberating. The model can explore scenes a camera crew would never shoot.

Reusable starting prompts

For atmosphere, I use this.

[Subject] [single action], [camera], [place and time], [material detail], [mood]

For a character moment, I use this.

A person with [short visual description] [one subtle action], [camera], [light], natural skin texture, [mood]

For a product-like object, I use this.

[Object] remains centered and unchanged while [one subtle motion], [camera], [environment], commercial lighting

They are not magic. They force the decisions that make generation easier to control.

What I do now

Write one sentence. Choose duration. Break the story into one-action clips. Use specific camera and light. Draft cheap. Upgrade only after the composition works. Check the first second, the subject, and the background. Save the winning prompt.

That process turns text-to-video from a novelty into a tool.

Text-to-Video AI: Complete Guide to Creating Video from Text