Magnus HedemarkandGitHub
24f21c64e3
Squash-merge the verified #412 eval coverage implementation. Required validate and paired evaluation checks passed at exact head b43ac564a5919a0f23fdab49ba052d7c514915cb; droid-review BYOK failure had no findings and is advisory.
DSPy — Programming, Not Prompting Language Models (Stanford)
An expert-level skill for programmatic prompt optimization with Stanford's DSPy framework. You write Python programs with typed signatures; DSPy optimizes the prompts automatically. This is the framework for prompt engineering that doesn't feel like engineering.
Why Install This Skill
When your agent loads this skill, it becomes a DSPy expert who can:
- Define typed signatures — input/output schemas with descriptions
- Build program modules — Predict, ChainOfThought, ReAct, and custom Module subclasses
- Select optimizers — MIPROv2, BootstrapFewShot, BootstrapFinetune — matching optimizer to bottleneck
- Compile programs — transform a Python program into an optimized, prompt-efficient artifact
- Evaluate and iterate — metrics, datasets, and optimization loops
What You Get
| Directory | Purpose |
|---|---|
SKILL.md |
Core paradigm, optimizer cheat sheet, compilation pipeline |
references/ |
Signatures deep dive, module patterns, optimizer selection guide, evaluation methodology |
Framework Comparison
DSPy is not a chain or RAG framework. It operates at the compiler layer — you define the program structure, DSPy optimizes the prompts. Use this when you want prompt engineering to be deterministic and testable, not a manual tuning exercise.
Requirements
Python 3.8+ with dspy package.
Quick Start
Start with the setup and first workflow in SKILL.md, then use the linked resources for the specific task you need to complete.
Triggers
Use this skill for the task types and keywords described in its SKILL.md description.