Good work. Like the author, I usually prefer experimenting with different system prompts, ordering of data in a prompt, etc. using small local open models. I get the impression from talking to people who exclusively use the most expensive and performant models that they don’t spend as much time learning to steer LLMs.