For a long time my AI video prompts were wishes, not instructions. I would write cinematic shot, amazing atmosphere, beautiful lighting and then wonder why the output looked like a screensaver. The turning point came when I stopped describing vibes and started describing decisions. What is the subject? What is the one action? Where is the camera? What must stay unchanged?

This is the prompt-writing process I now use for both text-to-video and image-to-video. It is not a pile of magic words. It is a checklist that makes generation easier to debug.

Start with a one-line brief

Before I touch a prompt, I write the brief in plain language. For example, a five-second product close-up where steam rises but the cup stays still. That sentence is the contract. If I cannot say it in one line, the idea is usually too big for one clip.

The brief also tells me whether to use text-to-video or image-to-video. If the subject is invented, text-to-video is fine. If it is a real product, person, or photo, image-to-video is safer because the reference image anchors identity.

Use the five-part prompt

After many failed generations, I settled on this order.

Subject + action + camera + environment + style and timing

Here is a text-to-video example.

A golden retriever runs through tall grass, low tracking camera follows from the side, late afternoon light, gentle depth of field, four-second clip

And here is an image-to-video example.

Keep the person from the reference image unchanged. She slowly turns her head and smiles, soft studio lighting, static camera, gentle and warm mood

The first prompt invents a scene. The second preserves a subject. They need different amounts of freedom. When I mix them up, I get the wrong tool and a disappointing result.

One action is the hard rule

The biggest prompt mistake is trying to make a film in one sentence. A five-second clip cannot hold a scene change, an emotional turn, and a product reveal. It can usually hold one visible change well.

So I split stories into shots. If I need a hand to pick up a mug, steam to rise, and a smile to appear, I generate those as separate shots and edit them. The final video can feel richer. The prompt does not have to do everything at once.

Camera language I actually reuse

Vague camera words waste credits. These are the terms I keep coming back to.

static camera when the subject must carry the motion. slow push-in when I want attention without drama. slow pan when the environment matters. low tracking shot when movement should feel physical. handheld when I want a social, natural feel.

Then I decide what should not move. For product shots, I often write product remains unchanged, static background, or camera holds while steam moves. This is less poetic, but it gives the model boundaries.

Name the light, place, and material

Beautiful lighting is too broad. I write soft window light, golden hour backlight, overcast daylight, or warm practical lamps. For place, I say wooden kitchen table, concrete studio, or wet city pavement. For material, I name matte ceramic, brushed metal, linen, glass, or wet stone.

This changed my results more than any style keyword. The model can render specific materials better than it can guess what premium means.

Keep the subject anchored

For image-to-video, I treat the uploaded image as truth. Then I state the features that must remain stable. If the person has short black hair, I write short black hair. If the product label is blue, I say blue label. If the animal is facing left, I keep that orientation in the prompt.

For text-to-video, I avoid relying on words that must appear exactly. Models often distort typography. If the video needs a name, headline, or legal claim, I add it in the editor.

Test with a draft, not a masterpiece

I now use a free or low-cost model for the first draft. The question is not whether it is good enough to publish. The question is whether the composition and motion idea are right. If the draft shows the wrong framing, no premium model will fix a broken idea.

Once the draft works, I upgrade. For controlled commercial shots, Kling v3 Omni is reliable. For expressive motion, Seedance 2.5 is useful. Veo 3.1 is the premium option for rich cinematic output. But all of them reward a clear prompt and punish an unclear one.

My failure checklist

When output looks wrong, I check in this order. Is there more than one action? Is the subject described clearly? Did I ask for text? Is the camera movement too complex? Is the duration too short for the action? Did I contradict myself by saying static and dynamic camera in the same prompt?

Most fixes are subtraction. Remove an action. Remove a scene change. Remove invented text. Slow the camera. Make the subject larger. Write what should remain unchanged. Then regenerate one variation.

Reusable templates

For a product, I use this.

Use the product from the reference image. Keep label, color, and shape unchanged. Show [one action], [camera movement], [specific environment], clean commercial lighting

For a person, I use this.

Keep the person from the reference image unchanged. They [one subtle action], [camera movement], [specific light], natural skin texture, [mood]

For an atmosphere clip, I use this.

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

These are not perfect. They are starting points that force me to fill in decisions instead of hoping the model reads my mind.

What good prompts produce

A good prompt does not guarantee a perfect video. It guarantees that when the video is wrong, I know what to change. That is the real value. It turns AI video from a lottery ticket into a process.

Write the brief. Lock the subject. Choose one action. Name the camera and light. Test cheap. Upgrade only when the idea is proven.

How to Write AI Video Prompts That Actually Work: A Step-by-Step Guide