It's not actually optimal. Each check should account for all previous feedback, but it may be optimal to make a known-incorrect guess and trade the chance of winning with that guess for additional information.

For example, if your first guess on wordle is BOUND and you learn that the word is _OUND, you know the answer is one of FOUND, HOUND, MOUND, POUND, ROUND, SOUND, WOUND. Satisfying all previous feedback leaves you checking those one at a time and losing with probability 2/7. Or you could give up the 1-in-7 chance of winning in 2 and trade it for certainly winning in either 3 or 4: HARMS checks four of those options, and WHOOP identifies the remaining three.