In DSPy, prompt engineering is replaced by foundation model programming. At the lowest level, this means that systems are expressed in modular computer code rather than with fragile hand-crafted prompt templates. To the extent that prompts are written by hand at all, it is just to express the central goals and requirements of the system. The details of formatting, output typing, and message passing are left to DSPy, in much the same way that traditional software compilers handle the translation from high-level programming languages to machine code.
At Bigspin, DSPy is also our primary tool for data-driven optimization. All too often, GenAI developers end up iterating on prompts based on vibes and a few examples. Such iteration is exceedingly unlikely to lead to an optimal system; no one would try to set the weights of a neural network by hand, but even experienced AI developers end up doing something akin to this with their prompts. One of DSPy’s key innovations was to replace these inefficient and uncertain development patterns with proper data-driven optimization: given a set of labeled examples and an initial system specification, DSPy’s optimizers seek to find a prompt that maximizes performance on those examples.
