The paper addresses the issue of corpus-design in relation to research questions, for under-described languages. It shows how a corpus emerges from the methodology and habitus of its contributors, and how it is shaped by the technical tools used for data organization. It also underlines the ways in which a morphosyntactically-annotated corpus, segmented into intonation units, is amenable to a wide array of searches, both corpus-based and corpus-driven, and both formal and functional. After a presentation of the annotation layout, and the segmentation choices that characterize the two projects, CorpAfroAs and CorTypo, scientific results are illustrated for two languages, Kabyle and Beja, and more marginally for Zaar, Juba Arabic and Modern Hebrew. They exemplify corpus-driven and corpus-based approaches of information structure and grammatical relations. Both types of approaches plead for an integrated view of prosody, closely interacting with syntax, semantics, phonology, information structure, and all levels of human communication and cognition. They also plead for a general endeavour to annotate as much as possible the large array of prosodic cues that are inseparable from speech processing and interaction dynamics.