Since LLMs are sometimes wonderfully useful, and sometimes not, I'd suggest effective use involves figuring out in which cases it's likely to succeed, which it's likely to fail.

For example, the mentioned graph has "initial prompt with iterative tweaks", followed by iterations of 'starting from scratch'. -- I don't understand why you'd think "this is an ineffective way of doing things", and then keep doing it.

Describing LLMs as "slot machines" seems like the author has no curiosity about the shape of what LLMs can/can't do.