Not sure how much of this is still relevant considering how fast AI develops, but as far as I know:

The AI processes the prompt in chunks of a set size, so if your prompt is too long anything after that gets sent as a new chunk. This can add an uneven amount of priority to elements of the prompt depending on where the break occurs.

The way around this is to manually insert 'BREAK's where they are best for you. e.g. I do like:

pony, score, tags, artstyle, and, location, tags
BREAK character, description, tags
BREAK action, description, tags (if the character description above is long)

This (in theory) makes it so that each component receives an equal amount of focus from the AI.

Stable Diffusion breaks the prompt up into 75 token chunks, then recombines them using... magic. (I don't have a good explanation for how that actually works.) The use of BREAK forces the prompt parser to start a new set of 75. You can think

of this like manually putting a return in when typing text so it doesn't get cut off mid-sentence, separating tokens that you want together.

@Nabonidus yeah if you don't use BREAK and have more than 75 tokens, you'll find keywords that span a token chunk boundary get mangled and sometimes while editing prompts you'll see your image just go completely haywire. That's what's happening there

Agree with everything everyone has written... I use it to break up concepts. You'll notice that the first words/tags have more influence than the later ones. When you use break, that seems to reset that chain so the next block starts important and reduces again normally.