Ask HN: Extracting html templates from corpus of rendered html documents

I'm thinking that it might be possible to reverse engineer handlebars-style templates by diffing a corpus of text/html responses from live (esp. legacy) websites, and looking for common structures.

This would be a useful thing for companies with legacy applications - you could record your outbound traffic, and produce a set of templates, and corresponding view model schemas.

Does such a thing exist?

How would one go about making such a thing? I have a few ideas, but it would be interesting to hear your thoughts.