I find I get great results from explicitly asking an LLM to write tests which profile memory and cpu usage. Then ask it to optimise to improve those metrics.