A prompt is the instructions and input sent to a language model to elicit a response. Its wording and structure matter: a small change can produce a different answer, and the best arrangement can vary by task and model. Manually improving a prompt usually means adding, removing, reordering, or rewording its sections. section 2.1 section 1.0
There are two practical ways to work with a prompt:
- Write it directly, pass system instructions and user input to an LM API, then test revisions against examples. section 4.4
- Use DSPy to define the LM, input and output fields in a signature, and a module that chooses the prompting technique. The module generates the prompt and returns a structured prediction; the generated system and user prompts can be inspected with lm.inspect_history(). section 2.9 section 3.8
Treat prompt quality as something to measure, not guess. Optimizers can try different demonstrations or instructions; compare their resulting prompts on the task rather than assuming one optimizer always wins. A custom module can also append an instruction, although repeated hand-tuning may be better left to an optimizer. section 5.6 section 7.4.1
A manual classifier baseline
An intent classifier using a hardcoded system prompt through the OpenAI API scored 0.88 accuracy on a 50-item test set—about the same as the unoptimized DSPy-generated prompt. The small test set does not establish a meaningful difference, but the comparison gives optimization a baseline. section 4.4
Listing 7.11 Custom module to add a closing section
from dspy.primitives.module import Module
from dspy.signatures.signature import Signature, ensure_signature
from typing import Any
class Closer_Example(Module):
def __init__(self, signature: str, **config: dict[str, Any]):
super().__init__()
signature = ensure_signature(signature)
signature.instructions += " Answer very carefully."
self.predict = dspy.Predict(signature, **config) #1
def forward(self, **kwargs):
return self.predict(**kwargs)
async def aforward(self, **kwargs):
return await self.predict.acall(**kwargs)
prog = Closer_Example("question -> answer")
prog(question="What is the capital of Spain?")
print(lm.inspect_history(n=1))
The custom-module listing shows how appending “Answer very carefully.” to signature instructions changes the generated prompt, which can then be inspected. section 7.4.1
