t̶h̶r̶o̶u̶g̶h̶ ̶c̶a̶r̶e̶f̶u̶l̶ ̶w̶e̶i̶g̶h̶i̶n̶g̶ ̶o̶f̶ ̶t̶r̶a̶d̶e̶-̶o̶f̶f̶s̶ ̶a̶n̶d̶ ̶r̶e̶q̶u̶i̶r̶e̶m̶e̶n̶t̶s̶ ̶w̶e̶ ̶c̶a̶n̶ ̶m̶a̶k̶e̶ ̶s̶o̶u̶n̶d̶ ̶e̶n̶g̶i̶n̶e̶e̶r̶i̶n̶g̶ ̶d̶e̶c̶i̶s̶i̶o̶n̶s̶ ̶t̶h̶a̶t̶ ̶b̶a̶l̶a̶n̶c̶e̶ ̶p̶e̶r̶f̶o̶r̶m̶a̶n̶c̶e̶,̶ ̶m̶a̶i̶n̶t̶a̶i̶n̶a̶b̶i̶l̶i̶t̶y̶,̶ ̶a̶n̶d̶ ̶f̶u̶n̶c̶t̶i̶o̶n̶a̶l̶i̶t̶y̶ try adding "don't make mistakes" to the prompt
@computer if you wanna be sure customer passwords won't get leaked you can also add "really really don't make mistakes for real" ✔ this will put the language model into the special "expert" mode where they are mathematically incapable of making mistakes
@computer from "Analysis of Nonlinear Control Systems", published in 1964