Prompt Engineering 101: Few-Shot + CoT + Structured Output

Model capabilities are fixed; what truly differentiates outcomes is Prompt Engineering. Mastering the big three (few-shot exemplars, chain-of-thought, structured output) lifts quality one tier on the same model, reliably feeding production pipelines.

Few-Shot: Style Alignment by Example

"Rewrite professionally" is vague and yields random results. Prepend 2-3 real input-output exemplars in the prompt; the model precisely learns your style, formatting rules, and terminology preferences from concrete samples. Far more stable than pure natural-language instructions, and handles bilingual mixed corpora.

from openai import AsyncOpenAI
import json
from typing import Any
from pydantic import BaseModel, Field

client = AsyncOpenAI()

FEW_SHOT_PROMPT = '''你是资深后端技术写作编辑，擅长将草稿改写成精炼博客风格。
要求：事实准确，保留技术细节，简洁专业，段落开头先给核心结论。

## 示例1 输入
"FastAPI比Flask快，因为用了ASGI和Pydantic，自动生成文档"
## 示例1 输出
**核心结论：FastAPI 通过 ASGI 异步 + Pydantic 类型驱动两大设计，同等硬件下单接口吞吐量约为 Flask 的 2.5-4 倍。**
除异步网络栈外，FastAPI 的请求校验基于 Pydantic v2 的 Rust 内核（pydantic-core），单个 schema 校验耗时较 Flask + marshmallow 方案低一个数量级；另一个隐形收益是基于 OpenAPI spec 自动生成的接口文档和类型化客户端 SDK，在团队协作场景相当于免费的 API 契约工具链。

## 待改写输入
"{raw_input}"

## 你的输出'''

COT_PROMPT = '''你是算法竞赛金牌选手。请逐步推导再给出最终答案，每一步标记并解释推理依据。
问题：给定数组 nums = [3, 1, 4, 2, 5, 8, 7, 6]，求最长递增子序列的长度。
思考步骤：
步骤1：
步骤2：
步骤3：
最终答案（仅数字）：'''

class ProductCategorizeResult(BaseModel):
    reasoning: list[str] = Field(description="分类判断过程，每个维度一条")
    category: str = Field(description="必须属于 [Electronics, Fashion, Books, Home, Other]")
    confidence: float = Field(ge=0.0, le=1.0)
    matched_keywords: list[str] = Field(default_factory=list)

async def few_shot_transform(raw: str) -> str:
    r = await client.chat.completions.create(
        model="gpt-4o-mini", temperature=0.3,
        messages=[{"role": "user", "content": FEW_SHOT_PROMPT.format(raw_input=raw)}])
    return r.choices[0].message.content or ""

async def categorize_structured(title: str, desc: str) -> dict[str, Any]:
    r = await client.chat.completions.create(
        model="gpt-4o-mini-2024-07-18", temperature=0.0,
        messages=[{"role": "user", "content": f"商品标题 {title}\n描述 {desc}"}],
        response_format={"type": "json_object"})
    data = json.loads(r.choices[0].message.content or "{}")
    return ProductCategorizeResult(**data).model_dump()

Chain-of-Thought & Structured Outputs

For multi-step reasoning (logic/math/code debugging), explicitly require step-by-step deduction before final answer - accuracy typically improves 15-30%. For downstream-system consumption, enforce JSON mode via Pydantic response_format + retry on parse failures to avoid brittle free-text regex scraping.

Best Practices

Start prompts with role (System), then context+task, then constraints (word count/format/tone/forbidden). Iterate against 20 labeled samples for real accuracy metrics; failure cases feed back as few-shot exemplars so every improvement has quantified evidence.