That's very interesting! What sort of direction did you head in with prompt optimization? Was it mostly in shrinking it and then using multi-shot examples? We found that shorter prompts (empirically) perform better than longer prompts.