A video prompt is a set of instructions for a shot. It works best when you can describe what success looks like before seeing the result. This guide uses hypothetical examples to show how to build and revise a prompt. They are writing exercises rather than measured comparisons between models. Check the options offered by the current video tool, because duration, reference inputs, and camera controls can differ.
Write four short notes: the subject, the action, the setting, and the intended use. A useful card might say: a ceramic mug; steam rising; a quiet window table; a website banner. Add a fifth note for anything that must remain stable, such as a product label or a person's identity. These boundaries make it possible to reject a visually attractive result that does not meet the actual brief.
Avoid beginning with a long list of visual adjectives. Words such as beautiful, cinematic, premium, or stunning do not explain whether the cup should move or the camera should stay still. Describe observable changes first. Style can follow once the subject and action are clear. If your source image already establishes the visual style, repeating every static detail may make the motion instruction harder to find.
Subject movement describes the main actor or object. Environment movement describes secondary elements such as water, steam, fabric, or foliage. Camera movement describes the viewer's position relative to the scene. A clear prompt can use all three, but a first attempt is easier to review when only one layer changes. The other layers can remain explicitly stable.
For an image of a coastal building, try: “Small waves move toward the shore. The building remains still. The camera stays fixed.” This tells you what to inspect: water movement, building stability, and camera position. A request for a sweeping orbit, changing weather, moving boats, and a time transition would introduce several independent failure points. Add those only when they serve the shot's purpose.
Text-to-video asks a model to invent the opening scene as well as its motion. Describe enough of the subject and setting to establish a composition. Image-to-video begins with a supplied composition, so focus on how visible elements should change. A prompt cannot reliably reveal accurate hidden sides of an object that are absent from the source. Large camera moves may therefore be unsuitable when product accuracy matters.
For a text-only brief, “A blue ceramic mug sits on a wooden table beside a window; a thin curl of steam rises; fixed medium shot” establishes a concrete scene. For an existing image of that mug, “A thin curl of steam rises slowly; the mug and camera remain still” may be a more focused instruction. Neither example guarantees a result; both make the intended action understandable.
Consider the prompt “Make this energetic and cinematic with lots of movement.” It does not identify the moving subject or define the camera behavior. A revision could say: “The cyclist pedals forward along the path while the camera follows from behind at a steady distance.” You can then evaluate the cyclist's motion, the camera distance, and whether the path remains coherent.
Another common conflict is “locked camera, smooth orbit around the product.” A locked camera and an orbit describe different shots. Choose the one the project requires. Similarly, “the person remains perfectly still while turning toward the door” needs clarification. You might keep the torso still while allowing a head turn. Describe that distinction directly rather than expecting the model to reconcile incompatible requests.
Stability instructions are useful when they specify something observable: keep the background architecture consistent, maintain the product color, or avoid adding extra people. Very long lists of unrelated exclusions can distract from the shot. If the interface provides a separate negative-prompt field, check the selected model's support before relying on it. Do not assume every model interprets special syntax in the same way.
When exact words must appear, plan to add them as an editing layer. Generated lettering can change during motion even if the source is readable. For an advertisement, the final headline, price, and legal text need independent review. Treat a model's output as visual material to inspect, not as a trusted source of business facts or accurate product specifications.
Save the source image, exact prompt, model label, and settings for each attempt. Record the displayed cost before submission. Add one observation afterward, such as “camera moved despite fixed-camera instruction” or “steam acceptable, cup handle changed.” Use that observation to choose the next edit. A short log gives you a clearer process than repeatedly adding adjectives to the same unsuccessful request.
Change one variable at a time when you want to learn from a comparison. If you change the source, model, duration, and prompt together, you may get a better clip without understanding why. Results can vary even with identical inputs, so avoid presenting one comparison as proof that a wording always works. Your immediate goal is to choose a suitable result within the available budget.
Define an acceptance checklist before the final retry: the requested action is visible, the subject stays recognizable, important geometry remains usable, and the framing fits the destination. Inspect the beginning, middle, and end. If an issue persists, consider changing the source image, simplifying the shot, or completing part of the work in an editor. More attempts are not automatically a better use of time.
Before publishing, verify rights to your inputs and review any misleading implications in the output. Keep your prompt log with the approved file so another person can understand the intended shot. Continue with the complete image-to-video workflow, or use the publishing review guide to inspect the finished clip. These guides support a repeatable process; actual results depend on the material, settings, and model in use.
