I agree. I've been really impressed with Anthropic. The issue for me comes when I want to take arbitrary user input and ask Claude questions about the user provided input. Claude is very, very, very ethical. Which is great, but it won't provide a response if the user tends to use a lot of curse words.
Do some masking of curse words with sht, ?!, verybad, or similar? Something that Claude will accept. It might work, if users are just generally badmouthed, not actively trying to trigger the model/system.
