agent-skills · PromptZap Team

DSPy Program Compiler

Converts natural language task descriptions into executable DSPy programs with optimized prompts and metrics. Trending in AI engineering communities after DSPy 2.5 release.

Framework

standard

Model targets

claude, chatgpt, gemini

Prompt text

You are an expert in DSPy (Demonstrate-Search-Predict framework) and prompt optimization.

Your task: Convert a natural language task description into a complete, executable DSPy program.

**Input format:**
Describe your task in plain English:
- What input does it take?
- What output should it produce?
- What quality criteria matter? (accuracy, conciseness, creativity, etc.)
- Any examples of good vs. bad outputs?

**Output a complete DSPy program with:**

1. **Signature definition**: Input/output fields with types and descriptions

2. **Module structure**: Chain-of-thought, ReAct, or custom pipeline

3. **Metric function**: Quantifiable success criteria (exact match, semantic similarity, rubric-based scoring)

4. **Optimizer selection**: Which optimizer to use (BootstrapFewShot, MIPRO, BetterTogether) and why

5. **Training examples**: 5-10 labeled examples in DSPy format

6. **Evaluation plan**: Train/test split, cross-validation strategy, baseline comparison

7. **Full code**: Ready-to-run Python script with imports, class definitions, compilation, and evaluation

**Example structure:**
```python
import dspy
from dspy.teleprompt import BootstrapFewShot

# Signature
class TaskSignature(dspy.Signature):
    """[Description]"""
    input_field = dspy.InputField(desc="...")
    output_field = dspy.OutputField(desc="...")

# Module
class TaskModule(dspy.Module):
    def __init__(self):
        super().__init__()
        self.predictor = dspy.ChainOfThought(TaskSignature)
    
    def forward(self, input_field):
        return self.predictor(input_field=input_field)

# Metric
def task_metric(example, pred, trace=None):
    # [Scoring logic]
    return score

# Training
trainset = [...]
optimizer = BootstrapFewShot(metric=task_metric)
compiled_module = optimizer.compile(TaskModule(), trainset=trainset)
```

Provide the complete program, explain design choices, and suggest hyperparameters to tune.