Use this skill when an agent needs to turn a brittle prompt or multi-step LLM workflow into a DSPy program with explicit modules, examples, metrics, and optimizers. The operator prepares the task definition and validation examples, then asks the agent to implement or revise DSPy signatures/modules, run evaluations, inspect scores, and apply DSPy optimizers to improve outputs.

## When to invoke

Invoke this instead of editing prompts manually when a workflow has repeatable inputs, observable outputs, and a metric or rubric that can be scored across examples. It is especially useful for RAG answer generation, extraction chains, classification, routing, and agent subroutines where regression testing matters.

## Scope boundary

This is not a generic listing for the DSPy framework. The skill is bounded to measurable optimization of LLM or agent pipeline behavior: define the DSPy program, build the eval set, run metrics, tune with optimizers, and report before/after quality with failures. Do not use it for one-off creative prompting or for production deployment concerns outside the DSPy optimization loop.

## Source-backed signals

Source: https://github.com/stanfordnlp/dspy. Upstream describes DSPy as a framework for programming rather than prompting language models, with modular AI systems and algorithms for optimizing prompts and weights. The verified repository has 34,308 GitHub stars, recent push activity, MIT license, and documentation at https://dspy.ai/.

## Installation

Install with `pip install dspy` for the published package, or use `pip install git+https://github.com/stanfordnlp/dspy.git` for the latest upstream main branch. Follow https://dspy.ai/ for current setup and optimizer examples.