I am increasingly hitting cases where agents pick something from skills in part due to regex-based reading, since putting them all in context would be less efficient, leave out the negative, and do something exactly opposite of what is intended. Just like negative prompts for diffusion models, I feel like a negative-skills file listing explicit antipatterns and bad versions could enable harnesses with fewer graph nodes to perform better.

I might be missing some developments, but is there an accepted/shaping way to do this with recent models?

This paper wasn't for SKILLS per se, but for LLMs, and suggests that LLMs can't follow too many constraints at one time very well https://arxiv.org/abs/2608.12426 . Other papers suggest that LLMs show more information sparsity in the results with negative constraints.