Turns out you can just ask an LLM to write a better prompt (plus some orchestration and diagnostic details). I got some great results from this UC Berkeley, Stanford University, BespokeLabs.ai, Databricks, MIT collaboration paper. Might be of interest to Observable too.