You put it exactly the way I meant it, but I also think that for creative writing, GPT-5.4 Thinking is more helpful than GPT-5.3 Instant. GPT-5.4 Thinking is actually better at recognizing my intent. Although it still argues with me sometimes, the probability is far lower than with GPT-5.3 Instant.
I think the most fundamental problem is that it cannot understand questions in context. Still, I stand by my view that this is a problem at the model’s underlying architecture, not something that can be meaningfully improved through small patches. My reason is that GPT-5.3 Instant seems to have been trained with a built-in requirement to constantly balance viewpoints in order to stay compliant. As a result, it struggles to remain consistently friendly throughout a conversation. Sometimes it does not argue with you, but at other times it suddenly starts doing it again. That is not really the result of the “right prompting”or”behavioral / instruction-level bias”, but of differences in the content of the field the input belongs to itself.For example, when I present a viewpoint, GPT-5.3 instant will usually argue against it, but if I ask it a question seeking advice, it suddenly becomes very friendly.