The venerable make of Unix [Fel79] is target-oriented: one starts with a main goal (target) and applies production rules (with associated actions such as invoking a compiler) until no more rules are applicable. The leaves of the resulting derivation tree1 can be seen as defining the set of source files that are used to make the main target.
CM, on the other hand, is largely source-oriented: Whereas with make one specifies the tree and lets the program derive the leaves, with CM one specifies the leaves and lets the program derive the tree. Thus, the programmer writes down a list of sources, and CM will calculate and then execute a series of steps to make the corresponding program. In make terminology, this resulting program acts as the ``main goal'', but under CM it does not need to be explicitly named. In fact, since there typically is no corresponding single file system object for it, a ``natural'' name does not even exist.
For simple projects it is literally true that all the programmer has to do is tell CM about the set of sources: a description file lists little more than the names of participating ML source files. However, larger projects typically benefit from a hierarchical structuring. This can be achieved by grouping ML source files into separate libraries and library components. Dependencies between such libraries have to be specified explicitly and must form an acyclic directed graph (DAG).
CM's own semantics, particularly its dependency analysis, interact with the ML language in such a way that for any well-formed project there will be exactly one possible interpretation as far as static semantics are concerned. Only well-formed projects are accepted by CM; projects that are not well-formed will cause error messages. (Well-formedness is defined to enforce a unique definition-use relation for ML definitions [Blu99].)