As software systems evolve, they become more complex and harder to understand and maintain. Certain reverse engineering techniques attempt to reconstruct software models from source code with the help of a parser for the source language. Unfortunately a great deal of effort may be required to build a specialized parser for a legacy programming language or dialect. On the other hand, (i) we typically do not need a complete parser that recognizes all language constructs, and (ii) we have a rich supply of (legacy) examples. We present an approach to use these facts to rapidly and incrementally develop parsers as follows: we specify mappings from source code examples to model elements; we use the mappings to generate a parser; we parse as much code as we can; we use the exceptional cases to develop new example mappings; and we iterate. Experiments with Java and Ruby, two very different languages, suggest that our approach can be a very efficient and effective way to rapidly construct software models from legacy code.