Comment by CuriouslyC

For Gemini 2.5 and ~GPT5.0-5.1, longer prompts with lots of explicit instructions and examples produced better conformance. Seems like heavily second guessing the models started to get counter productive around the end of last year.