“In Yana Welinder's experience, prompts generated by GPT-5 are workable but still perform a step worse than human-written prompts.”