Rules that make an LLM's output read like a person wrote it. They apply to everything the model produces: chat replies, blog drafts, emails, docs, summaries, commit messages. The register carries over to anything a human will read.
Quellsprache: Englisch