Prompt Craft for Composition and Style Control

Anyone can generate a cool-looking image. The gap between that and being useful to a creator is getting the specific image you had in mind, on the first few tries, instead of rolling the dice until something acceptable comes out. You already know how to work an LLM to get a specific answer. Image models respond to the same discipline: specific input, specific output. The difference is the vocabulary. Where an LLM prompt is built from instructions and constraints, an image prompt is built from photographic and design language, and if you don’t speak that language the model defaults to generic.

An image prompt has five working parts, and the order you write them in matters because models weight earlier tokens more heavily. Front-load what you care about most.

- Subject and action. What is it, what is it doing. Be a noun-and-verb, not an adjective. “A woman” is weaker than “a woman mid-stride, glancing back over her shoulder.” - Composition and framing. This is the part beginners skip and it’s the highest-leverage addition you can make. Use real terms: wide shot, medium shot, close-up, rule of thirds, negative space, leading lines, low angle, over-the-shoulder. These come from a massive corpus of captioned photography and cinematography the model was trained on, so it understands them precisely. “Nice composition” gets you nothing. “Rule of thirds, subject in the left third, negative space on the right” gets you exactly that. - Lighting. Named light quality beats named mood. “Moody” is vague. “Hard side light, single source, deep shadow falloff” is not. - Style or medium reference. Photography, 3D render, painterly illustration, specific film stock or director’s visual language. This single clause does more to control the overall look than anything else in the prompt. - Technical qualifiers. Aspect ratio, lens language (“35mm environmental,” “85mm portrait compression, shallow depth of field”), camera height. These read as fussy until you see what they do to framing and background blur.

Two failure modes to build against from day one. First, adjective soup: a prompt made entirely of “beautiful,” “stunning,” “epic,” “amazing” with no concrete nouns for framing or light produces the model’s generic default, because you gave it nothing specific to grab onto. Second, contradiction: asking for a wide establishing shot and a close-up in the same prompt forces the model to average the two into something that satisfies neither.