This is when you use ML to optimize an embedding vector to serve as your system prompt instead of guessing and writing it out by hand like a caveman.
Don't know why the big cloud LLM providers don't do this.