Based on the JDEST corpus data, this paper investigates the distributions of patterns and functions of the characteristic Sentence Stems (CSSs) in academic texts. It is argued that CSSs show strong features of contextual conventionality and are frame expressions to build a skeletal academic text. CSSs fall into seven major structural categories: personal subject, impersonal subject, dummy-it construction, existential construction, as-introduced structure, wh-question structure, and others; they are also closely connected with 26 main categories of academic discourse functions, which consist of 15 text-oriented functions, six participant-oriented functions and five research-oriented functions. It is also found that the use of CSSs depends largely on the local textual environment in which they occur, and they are the place where co-selections of various types are engineered to allow meanings to take shape. CSSs are the co-selection result of form, meaning and context.