Words are all you need

Models currently sit in an uncanny valley where they are good enough to produce roughly what you want from very few words but not good enough to produce exactly what you want unless you have all the words.

And unfortunately, words are not trivial. "Having the words" equates to having understanding and having understanding is difficult. Perhaps no amount of plan-mode or spec-driven development or chain-of-thought super-max-colossus reasoning will ever get you around the hard work of learning the words.

From another perspective, the effectiveness of chain-of-thought in the latest reasoning models is obvious: it fills in the words that you could not.

And if you don’t want to learn the words, which is fine, a worthy short cut, though you must recognize it as one, is to loosen your acceptance criteria.