Why not just give the model a prompt?

Every gain in LLMs is either through increases in compute efficiency, Architecture or Harnesses...

The rest seems like bells and whistles