You rewrite the prompt for the fifth time. The room is finally correct, the lighting is close, and the camera moves in the right direction. But the product still has the wrong shape, the character’s jacket keeps changing, or the final frame never matches the composition you need.
At that point, more adjectives may not solve the problem. The missing information may be visual rather than verbal.
This guide explains when a prompt is enough, when a reference becomes useful, and how to avoid giving several inputs the same job. It describes a planning workflow, not the result of a controlled product test.What text communicates wellText is useful for information that can be stated as an instruction:

• the subject and setting;
• one main action;
• general lighting and mood;
• camera direction, such as a slow push-in or locked shot;
• the order of simple events;
• broad visual style.

For example, “A ceramic perfume bottle stands on pale stone while the camera slowly moves closer” gives the scene a clear subject and motion. If the exact bottle design does not matter, text may be enough for an early concept.
Prompt rewriting is still worthwhile when the request is vague, contradictory, or overloaded. Removing three simultaneous actions can be more useful than adding another paragraph of description.Signs that the prompt is no longer the main problemConsider changing the input when the same visual error survives several clear prompt revisions.
If the object silhouette must match a real product, words such as “round,” “premium,” or “minimal” leave too much room for interpretation. If a character must keep a specific face, hairstyle, or outfit, a written description may produce a similar person rather than the intended one. If the shot must begin or end with an exact composition, text does not provide the same spatial target as an image.
The useful question is not “How can I make the prompt longer?” It is “Which part of this shot would be easier to show than describe?”Choose a reference for one specific jobUse a starting frame when the opening composition, product placement, character identity, or camera angle matters. It gives the shot a visual state from which movement can begin.
Use an ending frame when the destination matters: a box must finish open, a person must reach a doorway, or the camera must settle on a particular layout. The prompt should describe the transition between the two states rather than inventing a second destination.
Use a reference image when shape, color, clothing, face, or styling would take too many words to describe. A clean image with a clear subject is usually more useful than a crowded collage.
Use a reference video when timing, movement rhythm, or camera behavior is the difficult part. The video should demonstrate that narrow job. It should not also be expected to define an unrelated character, product design, and location.
Use audio reference when sound, pacing, or a spoken beat must guide the sequence. Audio does not remove the need to inspect lip movement, timing, and unwanted fragments.A short exampleImagine a six-second product clip for a blue travel mug. The text prompt repeatedly produces the right kitchen and camera move, but the handle changes shape and the lid color drifts.
Instead of describing the mug in greater detail, use a clean starting image of the actual mug. Give the image responsibility for its silhouette, handle, lid, and color. Give the prompt responsibility for one action: “The camera makes a slow, straight push-in while soft window light moves across the mug. Keep the mug stationary.”
Do not add a hand opening the lid, steam changing direction, a rotating platform, and a background transition in the same test. First find out whether the basic product identity remains stable during the camera move.A practical input workflowThe independent third-party MiniMax H3 tool supports text-based, frame-controlled, and multimodal reference workflows. That makes it suitable for comparing these input strategies within one planning process. It does not remove the need to review generated text, packaging, faces, object details, audio, and physical movement.
Before generating, assign each input a job:

• Prompt: action, camera, lighting, and sequence.
• Starting image: opening composition and visual identity.
• Ending image: final position or state, if required.
• Reference video: motion or camera rhythm, if text is insufficient.
• Reference audio: timing or sound guidance, when relevant.

If two inputs give different instructions about the same feature, simplify them. More references do not automatically create more control.Submission checklistBefore treating the clip as ready, check:

• Does the first frame match the intended starting state?
• Does the subject keep the same shape, colors, and important details?
• Is the requested action visible from beginning to end?
• Does the camera follow one understandable path?
• If an ending frame was supplied, does the transition reach it naturally?
• Are labels, logos, small text, hands, faces, and contact points acceptable?
• Does the exported file match the preview?

Stop rewriting when the unresolved requirement is clearly visual. Add the smallest reference that answers that requirement, give it one job, and test a short version before building a more complicated scene.