How to write AI video prompts that describe what you want

By Vellm · October 19, 2026

Start with what the viewer should see moving. "A dog takes two small steps sideways while the camera stays still" gives a video model a clearer task than "a stunning viral dog video."

How much else you need to describe depends on what you supply. A starting photo already shows the subject and scene. A video made from text needs you to establish them. A remake also has a reference clip to guide its movement.

Choose the mode first

In Vellm Create, Animate a photo begins with an image. Text to video begins with your description. Remake a video takes a clip and a photo, while Mix references lets you combine supported reference files and explain their roles.

The mode changes the job your prompt needs to do. For a photo animation, begin with the action. For text to video, introduce the scene as well. For a remake, explain who should replace the performer and which parts of the reference should guide the result.

Ready-made templates are another route when the routine you want is already available. You upload the requested photos and use the template's configured scene and movement.

Use @ to point to an uploaded reference

In Create, each image or video you add gets a label such as @Image1, @Image2 or @Video1. Use that label in your prompt to explain what the file should contribute: the subject's appearance, a background, or the movement to follow. Images and videos are numbered separately, so you can have both @Image1 and @Video1 in the same project.

For a dog photo and a dance clip, try:

Use the dog in @Image1 as the subject. Follow the body movement and camera framing in @Video1. Keep the dog's coat markings recognizable.

Here, @Image1 identifies the dog photo and @Video1 identifies the movement reference. Add both files first; typing their labels alone does not attach them. Choose Remake a video for this photo-and-clip workflow, or Mix references when you want to explain the roles of several files. The selected model must support the files you add.

To insert a reference:

Add your files and check the labels shown on their cards.
Open the description field. If it is folded, select Add a description.
Type @ at the start of the prompt or after a space. Choose a file from the suggestions. You can also use the arrow keys and press Enter or Tab.
Write what you want that file to supply. For example: "Use the subject in @Image1 and the background in @Image2."

Choose the label from the list to keep its spelling and capitalization correct. If you remove files, the remaining labels can change, so recheck the references in your prompt before generating. If no suggestions appear, check that you have added a file and that the current mode and model offer a description field.

The labels help you distinguish the references. How closely the result follows them still depends on the model and the task.

Build one shot you can picture

Write the action as a short sequence of visible events. Add a camera instruction if framing matters. You can use this outline without treating it as a rigid formula:

[Subject] [performs a specific action].
[Camera position or movement].
[A useful detail about timing, setting or the ending].

For a portrait animation:

The person looks toward the window, then turns back with a small smile. A slow camera push-in ends at a head-and-shoulders view. The movement stays gentle throughout.

For an object photographed on a table:

The ceramic mug remains on the table while a thin curl of steam rises. The camera slides slowly to the right, keeping the mug centered. Soft window light stays consistent.

These examples are starting points for experimentation. They have not been benchmarked across Vellm's models, and a prompt cannot guarantee an exact expression or camera path.

Give text-to-video a setting

Without a photo, the model has to invent what the shot contains. Name the subject, location and appearance details you care about before describing the motion.

A small cream-colored dog stands on a wooden floor in a bright living room. It tilts its head, takes one step toward a red ball, and stops. A low, locked camera shows its whole body. Natural daylight, realistic home-video style.

This will create a new dog. If the dog's identity matters, supply your own photo in a mode that accepts it.

Choose a duration that gives the action room to finish. A head turn fits a short shot; entering a room, greeting someone and changing clothes is a larger sequence. Split a complicated idea into separate clips when the model struggles to complete it.

Be clear about the camera

"Cinematic" leaves the camera largely unspecified. "A locked wide shot" asks for stable framing. "The camera slowly moves closer" asks for a push-in. Use the one that helps the viewer see the action.

For a pet dance, a wide shot gives you room to judge the feet. For a facial reaction, a closer shot makes the expression easier to see. Start with one camera movement. A fast orbit plus a zoom and a cut gives the model several changes to coordinate.

If you have two subjects, identify them by position: "the pet on the left lifts a paw; the pet on the right watches." Keep those labels consistent instead of referring to both as "it."

Change the part that failed

Save your prompt, model and settings with each attempt. If the camera cuts off your dog's paws, try specifying "a locked full-body shot with floor visible below the paws." If the action looks frantic, remove secondary motions or select a longer duration when available.

Change one thing at a time while you are diagnosing a problem. Switching the model, image, prompt and duration together makes it hard to tell what helped. Each completed new generation can use more credits; look at the Generate button before rerunning.

Runway's image-to-video prompting guide also recommends clear motion descriptions and simple initial prompts. Its model-specific advice does not establish identical behavior in every Vellm model. Use these examples as a first attempt, then adjust to the output you actually see.

For a clip whose movement you want to follow, try the remake workflow. For a photo with cropped limbs or a blurry face, fix the starting image before adding more instructions.