However, we don't parse markdown on the server and convert to HTML. Rather, we just prompt the model to emit HTML directly.

Considering the target audience it probably doesn’t matter but it sounds like this could lead to pretty heavy prompt injections, user intended or not. Have you considered that and are there any safeguards?

The domain is great by the way. Congrats on getting it!