When a word is compiled, it behaves differently from being interpreted.
E.g., consider +:
These two behaviours are known as compilation and interpretation
semantics. For normal words (e.g., +), the compilation semantics
is to append the interpretation semantics to the currently defined word
(foo in the example above). I.e., when foo is executed
later, the interpretation semantics of + (i.e., adding two
numbers) will be performed.
However, there are words with non-default compilation semantics, e.g.,
the control-flow words like if. You can use immediate to
change the compilation semantics of the last defined word to be equal to
the interpretation semantics:
Two conventions to mark words with non-default compilation semantics are names with brackets (more frequently used) and to write them all in upper case (less frequently used).
For some words, such as if, using their interpretation
semantics is usually a mistake, so we mark them as compile-only, and you get a warning when you interpret them.
In this example, first the interpretation semantics of flip is
used (and you get a warning); the second use of flip uses the
compilation semantics (and you get no warning). You can also see in
this example that compile-only is a property that is evaluated at text
interpretation time, not at run-time.
The text interpreter has two states: in interpret state, it performs the interpretation semantics of words it encounters; in compile state, it performs the compilation semantics of these words.
Among other things, : switches into compile state, and ;
switches back to interpret state. They contain the factors ]
(switch to compile state) and [ (switch to interpret state), that
do nothing but switch the state.
These brackets are also the source of the naming convention mentioned above.
Reference: Interpretation and Compilation Semantics.