“A key learning at Anthropic for prompting frontier models is to soften instructions from absolute commands to more nuanced guidance to avoid confusing the model in edge cases.”