Thing is, if I do this in, say, Python, it's not just that it's a single line of code, but it will get it right every single time.

If I do it with a LLM, there is a fairly good chance of an error, particularly if it's a complicated file. This means I have to proofread the result against the original. But it turns out humans are very bad at this task, and proofreading a YAML file vs a JSON one would take a lot longer than simply doing it in a deterministic fashion.