> or at least first navigate manually and then based on that, write a playwright typescript script for future tests This has always felt like a natural best us

> or at least first navigate manually and then based on that, write a playwright typescript script for future tests This has always felt like a natural best use for LLMs - let them "figure something out" then write/configure a tool to do the same thing. Throwing the full might of an LLM every time you're trying to do something that could be scriptable is a massive waste of compute, not to mention the inconsistent LLM output.