Prompt engineering as the sole means of getting anything of value and quality is a thing of the past.
But I also see people complaining about frontier models writing too many helper functions, changing more code than necessary, and etc.
These are things that can be addressed through context engineering; AGENTS.md, skills, etc.
Also a lot of use cases require smaller, lower latency models. And these faster, smaller models to this day exhibit similar failure modes to 2-3 years ago and require very purposeful context engineering to get satisfactory results.