The End of One-Sized-Fits-All Prompts: Why LLM Models Are No Longer Interchangeable