Comet Lab Atlas

A framework for programming language models with optimizable modules instead of hand-written prompts.

DSPy reframes the work: instead of hand-tuning prompt strings, you declare modules with typed inputs and outputs and let DSPy compile and optimize the prompts against a metric and examples.

It is a genuinely different mental model — closer to programming than prompting — and it shines when you have data to optimize against and care about squeezing out quality systematically.

Where it's ideally used

A fit when prompt quality is critical, you have examples to optimize against, and you want that tuning done by a compiler rather than by hand.

Where it doesn't fit

The learning curve and the need for evaluation data make it heavy for a simple, one-off prompt.