Once in a while I run into literature that invokes vanishing cycle machinery with a cryptic sentence like, "this follows from a standard vanishing cycle argument." Is there a good way to look at vanishing cycles, nearby cycles, and specialization? I have a decent idea how some of it works for studying the cohomology of a one-parameter flat family of degenerating complex manifolds (see below), but the general sheaf picture still gives me a headache. Any recognition principles (e.g., "this looks like a place where I can use a vanishing cycle argument") would be most welcome.

Say I have a family of complex manifolds where the fibers are smooth over a punctured unit disc, and have some mild singularities over zero (a priori the singularities could be arbitrarily bad, but say we blow up until we have simple normal crossings). The cohomology of the fibers forms a vector bundle on the punctured disc, and it comes equipped with some extra structure, such as a pure Hodge filtration and a Gauss-Manin connection that identifies nearby fibers. When we attempt to extend the vector bundle over the whole disc, the extra structures degenerate - the Hodge structure becomes "mixed", and the connection acquires logarithmic singularities. These structures aren't immediately relevant to this question, but they seem to be interesting.

As far as I can tell, vanishing cycles and nearby cycles arise when we try to relate the cohomology of smooth fibers X_{t} with that of the special fiber X_{0}. Each smooth fiber X_{t} has an inclusion map to the total space X, and X is homotopy equivalent to X_{0} by a fiberwise retraction. The composition yields a map from X_{t} to X_{0}, and the pushforward of a sheaf on X_{t} along this map yields the nearby cycles sheaf. When I start with the constant sheaf on X_{t} this yields a sheaf on X_{0} that computes cohomology of X_{t} for some abstract nonsense reason. So far, I'm okay, but it seems that choosing t is not canonical enough, so one replaces X_{t} with the homotopy equivalent universal fiber X_{oo} over the universal cover of the punctured disc (the upper half plane), and defines nearby cycles by some crazy pullback-pushforward-pullback sequence. Specialization and vanishing cycles seem to be similar - I think there is a nice geometric picture somewhere, but the proliferation of upper and lower stars makes me sad. Is there a good way to see through that thicket?