I recently had success with a problem I was having by basically doing the following:

- Write a correct, pretty implementation

- Beat Claude Code with a stick for 20 minutes until it generated a fragile, unmaintainable mess that still happened to produce the same result but in 300ms rather than 2500ms. (In this step, explicitly prompting it to test rather than just philosophising gets you really far)

- Pull across the concepts and timesaves from Claude's mess into the pretty code.

Seriously, these new models are actually really good at reasoning about performance and knowing alternative solutions or libraries that you might have only just discovered yourself.