How to Set Prompt Weights: The Real Effect of Bracket Syntax and Word Order

Weighting is, at its core, assigning an influence coefficient to each keyword. A higher coefficient means the model leans toward that direction more strongly during sampling. A lower coefficient means the word is essentially decorative. Same principle as tuning a material parameter — a small numeric change can swing the output wildly, or barely register at all, depending on whether you're touching a variable that actually matters.

Bracket syntax — the most universal one first

One layer of parentheses adds one notch of weight; stacking more layers keeps adding weight, but not linearly — by the third layer you're basically overfitting, and that word starts crowding out everything else in the frame in an ugly way. Square brackets work the other way, lowering weight, and they stack the same way.

Numeric weight syntax is more direct, written as "keyword:1.3" — 1.3 being the explicit coefficient, which is more precise than stacking brackets and gives you finer-grained control. Not every tool understands this syntax, so check whether the platform you're using actually supports this format before relying on it — get the format wrong and the system either throws an error or just treats it as plain text, and your weighting does nothing.

The effect of word order — something a lot of people miss

Most tools apply positional decay to prompts during encoding — words placed earlier carry a higher implicit weight, and this holds true regardless of whether you've manually applied any weighting syntax. So the same set of words, reordered, can produce noticeably different results — this isn't superstition, it's a consequence of the encoding mechanism.

Practical tip: put your most critical subject and style words first, and push modifier or secondary detail words toward the end. If you find a word you weighted manually still isn't having an effect, check whether word order has buried it too far back — weighting syntax and position can end up fighting each other more often than you'd expect.

Differences between tools — don't use the same parameter set everywhere

Some tools fully expose both bracket and numeric weighting interfaces, letting you tune parameters as directly as adjusting a rendering engine. Some tools don't support manual weighting syntax at all — in those cases, the only options are repeating keywords or adjusting the intensity of the wording to simulate the effect of weighting, like swapping "red" for "vivid true red" to substitute descriptive precision for numeric weight.

This is where things most often go wrong when porting prompts between tools — take a prompt with numeric weights and drop it straight into a platform that doesn't support that syntax, and those weight markers either get output as literal characters or simply ignored, leaving the result well off from what you expected — all that tuning wasted.

Tuning approach: change one variable at a time

Same habit as tuning render parameters — don't adjust several words' coefficients at once, because then you can't tell which one caused the change, which is a pure waste of time. Lock in one word, tune its coefficient, generate and compare, confirm the direction is right, then move to the next word. It's slower, but it saves more time than blind trial and error — the forty minutes I once spent hard-grinding out material parameters taught me that lesson, and it applies just as well to prompt weighting.

Weight tuning is especially useful in scenarios that need precise control over the proportion of elements in a frame — e-commerce product shots, for instance, have very specific requirements for how weight is distributed between subject and background. If the subject's weight isn't high enough, the background stealing the spotlight is a common failure — that's where numeric weight syntax works a lot better than piling on adjectives.

By the way, weight tuning pairs well with reverse prompting — reverse prompting gives you the words, weight decides which of those words takes priority. The two work in sequence, not as substitutes for each other.

There's no one-size-fits-all parameter table for weighting — different models have different sampling mechanisms, and the same set of numbers that works well on one model might need a complete retune on another. Don't expect to copy a parameter sheet and use it for the rest of your life — methodology is worth more than any specific number.