Prompt Engineering Playbook | Prompt 模式速查手册

A quick-reference playbook of battle-tested prompt patterns. Use it when you (or the user) need to write a prompt that reliably produces good output — or to diagnose why a prompt is underperforming.

一套经过实战检验的 prompt 模式速查手册。当你需要写一个「稳定产出高质量结果」的 prompt，或想诊断 某个 prompt 为什么效果不好时，直接按本手册查模式、对样例。

When to use | 适用场景

• The user asks: 「帮我优化这个 prompt」「怎么写 prompt 才能…」
• Writing or auditing prompts for classification, extraction, generation, coding
• Choosing a prompt strategy before jumping into a long prompt

The patterns | 核心模式

1. Few-shot | 少样本示例

Give 2–5 input→output examples before the real task. Best when the expected output style matters more than the rules.

给 2–5 组「输入→输出」示例再进入正式任务。当输出风格比规则更重要时最有效。

Task: tag the sentiment of a review.
Review: "Love this app, fast and clean!" -> positive
Review: "Crash on every launch." -> negative
Review: "It's okay, nothing special." -> neutral
Review: "Worst update ever, lost my data." -> ?

2. Chain-of-thought | 思考链

Ask the model to reason step by step before answering. Best for math, logic, multi-step reasoning. Keep the reasoning in a scratchpad and give the final answer separately.

让模型先分步推理再给结论。适合数学、逻辑、多步骤推理。推理放在草稿区，结论单独给出。

Think step by step, then answer in the format: "Answer: <final>"

3. Structured output | 结构化输出

Ask for JSON / markdown / table with an explicit schema. Best when the output will be parsed by code.

要求 JSON / 表格等明确结构。适合需要被程序解析的输出。

Return JSON only, with keys: {"model": string, "price_per_million_input": number, "notes": string}

4. Role prompting | 角色设定

Assign a role + constraints up front. Best for tone, domain knowledge framing, and consistency.

开头设定角色与约束。适合语气、领域视角与一致性。

You are a senior ML engineer doing a code review. Be concise and cite specific lines.

5. Self-consistency | 多路验证

Generate several candidate answers (e.g. via temperature or by re-asking) and pick the most common / most coherent one. Best for high-stakes single answers.

多次生成候选答案并取最一致者。适合「答错代价高」的单点问题。

6. Anti-patterns to avoid | 常见反面模式

• ❌ 过长的任务清单，缺少优先级 → 把任务分步骤并编号
• ❌ 负向指令含糊（「不要写太多」）→ 改为正向且可度量（「≤150 字」）
• ❌ 没有给输出格式 → 永远给出期望的格式与示例
• ❌ 一次问太多问题 → 拆成多个 prompt 或明确「逐条回答」

Workflow | 工作流程

1. Ask what the prompt is for / what keeps failing.
2. Pick the pattern(s) above that match the failure mode.
3. Rewrite the prompt with: clear goal → context/constraints → few examples → explicit output format.
4. If the output is still off, adjust iteratively: change examples first, then format, then constraints.

Tips | 提示

• Examples beat explanations almost every time. | 示例几乎总是比解释更有用。
• State the desired output format explicitly and give one example of that format.
• When output is parsed downstream, ask for JSON and validate it in the same turn.
• Use this playbook as a checklist, not dogma — combine patterns when the task needs it.