Inspired by this question where there are differing views on SET NOCOUNT...
Should we use SET NOCOUNT ON for SQL Server? If not, why not?
At the risk of making things more complicated, I encourage a slightly different rule to all those I see above:
NOCOUNT ON
at the top of a proc, before you do any work in the proc, but also always SET NOCOUNT OFF
again, before returning any recordsets from the stored proc. So "generally keep nocount on, except when you are actually returning a resultset". I don't know any ways that this can break any client code, it means client code never needs to know anything about the proc internals, and it isn't particularly onerous.