Constraints beat adjectives
"Modern", "clean", "professional" mean nothing to a model and slightly less to you. Numbers and prohibitions do the work.
Adjectives feel like direction and carry almost none.
"Modern" often turns into whatever patterns are common in the model’s training data, such as rounded corners, soft shadows and familiar colour choices. "Clean" may become more whitespace and lower contrast. The problem is that lower contrast can quickly become a readability issue.
Constraints resolve to one thing. "Maximum 640 pixels wide." "Body text at 4.5:1 contrast or better." "Two font sizes, 16 and 32." "No shadows." "Stacks to one column below 640."
Prohibitions are the underrated half. "No gradients, no emoji, no icon library, no CSS framework" removes an entire class of default in four words, and models follow explicit prohibitions well.
Use this test on any line of a prompt: could two people reasonably disagree about whether the output satisfied it? If yes, the instruction is still too vague.
Take a prompt you wrote earlier this week and rewrite every adjective in it as a number or a prohibition. Run both. The difference is usually larger than you expect and it is not subtle.
You should now be able to
- Replace an adjective with a constraint
- Say what "clean" resolves to in practice
- Use prohibitions, which are often the strongest instruction available
Loading…