The safety practices you describe are basically the right mental model: assume it's fallible, keep writes reversible, review everything, commit often