Advanced Prompt Engineering

Overview

Advanced prompting moves beyond single instructions toward controllable, testable behavior. Key levers are role and system messages, few-shot exemplars, explicit output schemas, and reasoning scaffolds like chain-of-thought or step decomposition. The goal is to reduce variance so the same prompt yields reliable results across inputs, which matters far more in production than clever one-off wording.

Syntax / Usage

Separate a stable system message (rules, format, persona) from the dynamic user message (the task). For machine-readable output, request structured JSON rather than parsing free text. Modern APIs support schema-constrained responses.

from openai import OpenAI

client = OpenAI()

SYSTEM = (
    "You are a support classifier. Return JSON only. "
    "Never invent categories outside the allowed set."
)

resp = client.chat.completions.create(
    model="gpt-4o",
    messages=[
        {"role": "system", "content": SYSTEM},
        {"role": "user", "content": "My invoice charged me twice this month."},
    ],
    response_format={
        "type": "json_schema",
        "json_schema": {
            "name": "ticket",
            "schema": {
                "type": "object",
                "properties": {
                    "category": {"enum": ["billing", "bug", "account", "other"]},
                    "urgency": {"enum": ["low", "medium", "high"]},
                },
                "required": ["category", "urgency"],
                "additionalProperties": False,
            },
        },
    },
)
print(resp.choices[0].message.content)  # valid JSON matching schema

Examples

Few-shot exemplars anchor tone and edge-case handling better than adjectives. Keep examples short and representative:

FEWSHOT = [
    {"role": "user", "content": "reset link expired"},
    {"role": "assistant", "content": '{"category":"account","urgency":"medium"}'},
    {"role": "user", "content": "site is completely down"},
    {"role": "assistant", "content": '{"category":"bug","urgency":"high"}'},
]
messages = [{"role": "system", "content": SYSTEM}, *FEWSHOT,
            {"role": "user", "content": "how do I export my data?"}]

For hard reasoning tasks, ask the model to think stepwise but return only the final answer, keeping the chain out of user-facing output:

prompt = (
    "Solve step by step internally, then output ONLY the final number.\n"
    "A cart has 3 items at $12 and 2 at $7.50. Total?"
)

Common Mistakes

• Cramming rules, data, and format into one giant user message
• Relying on prose parsing instead of JSON schema or function calling
• Over-long few-shot sets that waste tokens and bias toward examples
• Vague adjectives ("be concise") without concrete constraints or examples
• Never running a prompt against a fixed eval set before shipping changes

See Also

prompt-engineering large-language-models ai-evaluation