Remember that there are three things a Turing machine can do, given an input. It can:
- Diverge (infinite-loop)
"Recognizable" means "we can accept on strings in our language, and reject-or-diverge on anything else". "Co-recognizable" means "we can reject-or-diverge on strings in our language, and accept on anything else". Or, equivalently, "we can accept-or-diverge on strings in our language, and reject on anything else".
So the key is, intuitively, to find a language so that a TM always diverges on (some) strings not in the language.
A classic example is the Halting Problem. (To be more specific, the set of encoded Turing machines that halt on empty input, for a given encoding.) You can recognize this language by simulating the machine, then accepting. But you can't recognize its opposite.
(Formally, you have to do a bit more work. You have to prove that the language is undecidable: that no Turing machine can always say "no" correctly, even if it can always say "yes" correctly. Usually this means reduction to the Halting Problem or Rice's Theorem.)