“As language models become more capable, they require less explicit direction, fewer constraints, and fewer examples in their system prompts to perform effectively.”