I was wondering if Whiz was using agent-style planning prompts, but it's much simpler and passes a single `function` parameter named `shell` to the OpenAI API.

I thought that was clever, because shell commands are highly composable, so it's likely that something can be done in "one shot".

It will also be way more efficient with token usage.