Field guide
AI Video Prompts: Examples and How to Improve Them
Write clearer AI video prompts with original examples, a real framing failure, model-specific guidance, and focused revisions you can test.
What an AI video prompt needs to explain
An AI video prompt describes a shot: what is in the frame, what happens, and how the camera observes it. For a homepage hero, a campaign teaser, or a creator intro, start with the moment the viewer should notice. A short clip of fabric catching the light needs different direction from a subject entering a room.
For text-to-video, describe appearance as well as motion. For image-to-video, the image already supplies the visual starting point, so spend more of the prompt on what changes. Runway makes this distinction in its text-to-video guide and image-to-video guide, both written for Gen-4.5.
Write the shot in three passes
First, write the action as if explaining it to someone who cannot see your idea. “A toy boat drifts across a puddle” gives you an event to look for. Add a mood once the action is clear.
Next, add the visual details you actually care about: a red paper boat, a rain-soaked pavement, a low viewpoint. Leave unspecified details open if different interpretations would all work for the project.
Finally, choose the camera's job. Should it stay beside the puddle, follow the boat, or move closer? Our camera movement reference explains those choices without loading this first draft with a list of film terms.
A drafting pattern you can copy is:
[Framing] of [subject] in [setting]. [One observable action]. [Camera behavior]. [Light or style that matters to this shot].
Use this as an editing aid. Reorder or delete any part that adds no useful direction.
Three AI video prompt examples
The following are original writing examples, supplied for adaptation and testing.
A small object with a clear action
A low, close view of a red paper boat in a shallow puddle on gray pavement. The boat drifts slowly from left to right as small ripples spread around it. The camera remains in place beside the puddle. Soft overcast daylight.
What to check: the boat moves across the frame while the pavement stays in place. Keep an eye on the folded tip as it travels.
An environment without a character
For a title backdrop or a quiet transition between portrait shots, give the environment one visible movement:
A wide view of an empty greenhouse at dawn, with rows of terracotta pots on wooden benches. A light breeze lifts the loose curtain beside the open door. The camera slowly moves forward down the central aisle. Pale morning light enters through misted glass.
What to check: the forward move continues along the aisle, the benches retain their arrangement, and the curtain is the intended moving detail. If the camera circles instead, revise that instruction before adding more atmosphere.
The same idea with an input image
If you already have the greenhouse frame, try a motion-focused draft:
The camera moves slowly forward along the central aisle. The loose curtain beside the door lifts gently in the breeze.
Here, the source image carries the pots, benches, glass, and lighting. If the image is tightly cropped or points away from the aisle, rewrite the shot around that composition or choose another source. The image animation walkthrough covers that workflow.
A real result: detailed framing instructions still missed
The brief for this homepage visual was a black horse cantering across a reflective salt flat. The whole horse needed to stay inside the frame so the clip could work in different page layouts.
The prompt included this exact instruction:
Medium-wide composition with the full horse visible and generous negative space. The horizon remains level. The stallion stays inside the central 65 percent so its head, tail and legs remain safe on mobile and desktop crops.
The second clip missed the framing requirement despite that explicit direction. For a hero or portrait crop, inspect the opening, middle, and ending before choosing the take. If the subject is already cut off in the original, a tighter page crop will lose more of it.
Example details
Both clips were generated through the Seedance 2.5 provider API on August 27, 2026: six seconds, 4:3, audio off, and an identical prompt. Only the submitted resolution changed; no seed was recorded. With one run per setting, we cannot attribute the framing difference to resolution or compare prompt-writing methods. The six inspected frames are at 0, 1, 2, 3, 4, and 5.8 seconds.
Playback copies are resized transcodes. The 720p record and 1080p record contain the full requests and original file hashes. These archived provider runs have no retained billing amount or product Credit record. See our editorial method.
Revise a mismatch instead of rewriting everything
Save the prompt and settings before changing them. Write one sentence about the result, such as “the boat stays still while the whole frame pans right.” That observation points to a camera/subject separation problem rather than a need for more style words.
Scroll sideways to compare all columns →
| What you observed | A focused next draft | Keep unchanged for comparison |
|---|---|---|
| The whole frame moves instead of the boat | “The boat travels across the frame; the camera stays beside the puddle.” | Boat, pavement, light, and generation settings |
| The boat moves too quickly | Replace “drifts slowly” with “drifts a short distance across the puddle.” | Camera, setting, and style |
| The scene includes an unwanted cut | Describe one continuous view and remove wording that introduces a new scene | Subject and intended action |
| The boat changes into another object | First document where the change occurs; use the artifact diagnostic guide | Keep the original result for comparison |
Try one revision in Text to Video, keeping the other settings the same. Save both results: generation varies between attempts, so judge the change across several clips before relying on it.
Use model-specific instructions only where they apply
A negative sentence in the main prompt and a separate negative-prompt field are different things. Runway's Gen-4 guide says negative phrasing is unsupported. Google's Veo guide documents negative prompts, while advising against “no” and “don't” wording within them.
Start with a positive description of the desired shot. Use a negative-prompt field when the selected model and interface offer one, following that model’s instructions.
Keep a short record of each attempt
Record the date, model, input mode, full prompt, source image if used, selected duration and resolution, displayed Credits, and result. Add the timestamp of the most important mismatch. A note such as “the boat’s tip changes at 00:03” tells you where to compare the next attempt.
Before submitting another attempt, check its displayed cost; a creative retry is another generation. The Credits guide explains how to plan a retry budget. When your draft describes a shot you can evaluate, open Text to Video and test that first version.
Use the method