Prompting images and video is not prompting an LLM
Prompting a language model is about giving instructions and context. Prompting an image or video generator is closer to art-directing a photoshoot: you're describing a scene, not issuing a command. Effective prompts front-load the subject and action, then layer in specifics that push the model toward a particular visual outcome, things like camera and lens details ('shot on 35mm film, shallow depth of field'), lighting ('golden hour, volumetric fog'), composition ('wide shot, rule of thirds'), and style references ('in the style of a Wes Anderson film'). Vague adjectives like 'beautiful' or 'high quality' do far less work than concrete, visual, specific language, because the model has learned strong associations between specific visual vocabulary and specific pixel patterns, not between abstract quality judgments and pixels.
Negative prompting, explicitly telling the model what to avoid (blurry, extra fingers, watermark, distorted text), is another technique with no real LLM equivalent, since it works by steering the diffusion process away from regions of the learned distribution associated with those unwanted qualities. Iteration also looks different: instead of rephrasing and hoping, effective users change one variable at a time (lighting, then composition, then style) and use img2img or reference images to hold a composition steady while exploring variations, since generation involves real randomness at every run.