GRAID (Grammatical Relations and Animacy in Discourse) is a system for annotating natural discourse, designed to be applicable to a maximally large and typologically diverse set of languages. The annotations target referential expressions in texts, noting the form, person value, and animacy values, and are thus primarily of interest for researchers working on issues such as referential density (Bickel 2003), or Preferred Argument Structure (Du Bois 2003). The aim of the annotation is to provide a set of tagging categories and procedures to enable quantitative comparison across corpora from different languages to be made. The Manual sets out the rationale for the system, discusses different theoretical approaches, illustrates the main features and discusses problematic cases. An index with the core inventory of symbols is included at the end of the manual.