Generators read every word literally. These habits cut the number of retries you need.
Be specific about the subject
"A cat" gives you a random cat. "A short-haired ginger cat sitting on a wooden windowsill, morning light, shallow depth of field" gives you something usable on the first try.
Say the style out loud
Prompts benefit from one or two style anchors: documentary photo, flat vector illustration, watercolor, isometric 3D render. Without a style, you get the model's default — usually not what you want.
Set the frame
Mention composition: close-up, wide shot, top-down, centered subject with negative space on the right. Combined with the right aspect ratio, this is how you get a usable hero or thumbnail.
For image-to-image and video-from-image
Describe what changes, not the whole scene. The source image already carries the subject and composition — the prompt should only instruct the edit or motion.
Avoid negatives
"Without text" and "no people" are unreliable. Describe what you do want instead.
Iterate cheap
Try a few image-from-text variants with short prompts before committing to a longer video generation. Images burn far less quota and cost.