This is awesome. The promise of reliably preserving formatting and placeholders like {{username}} is the absolute killer feature here.

I'm really curious about the prompt engineering involved. How do you instruct the model to translate the surrounding text while ensuring it never touches the placeholders or any embedded HTML, especially since syntax can vary between languages?

That seems like the most difficult part of the problem to solve robustly.

Great work!

You can see it being used in my other project: https://tududi.com