We present a methodology for the systematic realisation of control flow analyses and illustrate it for Concurrent ML. We start with an abstract specification of the analysis that is next proved semantically sound with respect to a traditional small-step operational semantics; this result holds for terminating w well as non-terminating programs. The analysis is defined coinductively and it is shown that all programs have a least analysis result (that is indeed the best one). To realise the analysis we massage the specification in three stages: (i) to explicitly record reachability of subexpressions, (ii) to be defined in a syntax-directed manner, and (iii) to generate a set of constraints that subsequently can be solved by standard techniques. We prove equivalence results between the different versions of the analysis; in particular it follows that the least solution to the constraints generated will be the least analysis result also to the initial specification.