Prompting Foundations for Reliable Systems

Prompting at production scale isn't about clever phrasing — it's about reducing variance. The goal is a prompt that behaves the same way across thousands of inputs.

Structure beats prose

Give the model a clear contract: role, task, constraints, and an explicit output format.

You are a {role}.
Task: {task}
Constraints:
- {constraint_1}
- {constraint_2}
Output: respond with valid JSON matching {schema}.

Make outputs machine-checkable

If you can validate the output programmatically, you can build an eval. If you can build an eval, you can improve systematically.

from typing import Literal
from pydantic import BaseModel
 
 
class Result(BaseModel):
    sentiment: Literal["pos", "neg", "neu"]
    score: float
 
 
parsed = Result.model_validate_json(model_output)

Evaluate, then iterate

Keep a fixed test set of inputs with expected behaviors. Change one variable at a time and measure. Prompt engineering without evals is guessing.