I'd really prefer to build AST's for each language, then output a stringified representation of that - rather than just build strings that happen to be valid syntax.
I don't expect the string approach to be weak or short-term (languages do not often get rewritten, with old syntax invalidated), but it feels very wrong to be writing literals with newlines, tabs, brackets and variable names baked-in.
Every generator looks very similar, and all approach the same goal of writing out equivalent code in slightly different syntax. But every one is essentially copy-pasted and modified from the last (with the JS one being the common ancestor).
The only way to avoid this is abstraction - making each generator a type, then defining methods which return the appropriate string for a given type. Unfortunately that makes it hard to read, because the amount of methods would be huge. Every language has slightly different structure, from semicolons to tabs to names of errors to quoting strategies to string concatenation to varible naming. Everything is different, it feels like it'd be more confusing to abstract rather than just copy/paste changes across generators.