With the complexity of modern pipelines, there are very few humans that can beat a good optimizing compiler. Considering that with an LLM you're also bloating limited context with unsemantic instructions I can't see how this is anything but an exercise in failure.