A variant of the well‐known recursive descent parsing technique is developed, based on the assumption that each non‐terminal of the language is implemented through a separate module. The modularization principle requires that the recursive descent parser of one non‐terminal must be written knowing only the production rules of that non‐terminal, and not the other parts of the language. This leads to a parsing technique which is ‘lazy’ in the sense that part of the parser is constructed during parsing, according to the needs of a particular input text. It turns out that this technique can be implemented as efficiently as conventional recursive descent parsing, which is one of the fastest known parsing methods. The main benefit of the technique is that it allows a language to be implemented in small pieces which are easy to modify, replace, and reuse.