5.6 and fable seem markedly better at writing prompts. historically, models have been very bad at writing prompts. it makes sense - historically training data don't contain a lot of good prompt data for obvious reasons. there should be a benchmark for prompt writing!