A prompt does not command a specific output. It changes which outputs are likely. Add the word "sunset" to an image request and warm orange tones become far more probable while cool blue tones become less so. Add "in the style of a technical manual" to a writing request and plain, numbered sentences become more probable than flowing prose.
This is what it means to call a prompt a conditioning signal. The model still samples, and the result is still not fully determined, but the field of plausible outputs has been reshaped. A vague prompt leaves a wide field, so results vary a lot and may drift from what you had in mind. A specific prompt narrows the field, so results cluster closer to your intent.
Specificity works because each added detail is another constraint the output has to satisfy. "A dog" allows almost anything canine. "A small brown dog sitting on a red bicycle in a snowy street at dusk" rules out most of that space. The trade-off is real: more constraints give you more control but also less surprise, and constraints that contradict each other leave the model with no coherent option to favor.