My strategy when faced with doing a DSL of some kind is to write an accepter in bison or similar and knock the bugs out of the grammar and basic readability issues. Then go recursive descent with a table-driven precedence-climbing parser for expressions when ready for the real deal.
This is, I think, a very good strategy for many reasons. Not only is it a cheap way to work out tthe design of the language and a way to verify the correctness of your implementation. It's also a good fallback solution: if your (more expensive, expressive) implementation is not ready in time, you can fall back to integrating the generated implementation to meet an important deadline or whatnot.
