Prompt weighting

It is a scalpel where CFG scale is a dial. Word order and repetition already nudge a model's attention, as the prompt entry covers, but that is an inherent bias of how the text is read, not a control you invoke on demand. Weighting syntax is deliberate: mark one term and only that term's pull on the generation moves, up or down, while every other word keeps the influence it already had.

That is what separates it from CFG scale, and the two are routinely reached for interchangeably when they should not be. Raising CFG scale turns up obedience to the entire prompt at once — the words that were already landing get pushed harder along with the one that was not, which is why the fix for one ignored word so often arrives with a burnt, over-contrasted result attached. Weighting only that one word leaves the rest of the sentence exactly where it was.

The syntax itself did not standardise. Parenthetical weights, a trailing colon-and-number, doubled colons — the punctuation differs by provider, and none of it is portable. A model that never implemented the mechanism does not politely ignore the brackets either; it reads them as literal characters, which is how a weighted prompt copied from one tool to another sometimes comes back with stray punctuation rendered into the scene.