Five-module prompt structure
Role, Task, Context, Output Spec and Interaction Guide compose as independent, single-responsibility blocks — merge or omit for simple tasks.
Engineering Guide · Prompt Design
Prompt-Design-Guide
Battle-tested prompt patterns and iteration workflows for AI agents, code generation, and content processing.
License MITFormat Markdown · PDFLanguages EN · CNScenarios Agent · Code · Content
Features
What it does
A practical, engineering-oriented guide for designing prompts in AI Agent, code generation, and content processing scenarios. It collects battle-tested patterns, a standardized iteration workflow, and measurable evaluation metrics — no subjective aesthetics, only practices that reproduce across tasks. Shipped as Markdown and PDF, in English and Chinese.
Principles that hold up
Positive instructions over negative constraints, explicit reasoning paths, few-shot example rules, and anchor reset for conversations beyond 4000 tokens.
Six-step iteration workflow
Baseline, smoke test, root-cause analysis, targeted fix, regression, release — with every change logged and one module modified at a time.
Web Q&A playbooks
A context-injection template, a High/Medium/Low confidence self-check, and multi-turn evidence summaries keep answers inside the given material.
Large-project agent strategies
Project-level instructions, module-level task decomposition, change prompts with impact analysis, and debug prompts for OpenCode, Cursor and peers.
Four hard evaluation metrics
Format compliance, key-element coverage, factual hallucination rate, and instruction-following deviation — each with an acceptable threshold.
Prefix tag system
[Hard Constraint], [Important], [Optional], [Background] and [Counter-example] prefixes help models parse long prompts reliably.
Details
The five modules
| Module | Purpose |
|---|---|
| Role | Define expertise and perspective |
| Task | State what to accomplish explicitly |
| Context | Provide background, constraints, or prior logs |
| Output Spec | Specify format, length, structure, style |
| Interaction Guide | Request step-by-step reasoning or intermediate outputs |
Quick Start
How to use
- Clone the repository — the whole guide is plain files, no build step.
- Read Prompt Design Guide.md (renders on GitHub) or open the PDF edition.
- Start with Prompt Structure, then apply the Design Principles to your own prompts.
- Tune with the Iteration Workflow and score outputs against the Evaluation Metrics.
shell
git clone https://github.com/AkiroMusic/Prompt-Design-Guide.git
“Prompt quality directly determines output accuracy, consistency, and safety.”
工程化指南 · 提示词设计
项目简介
一本面向 AI 智能体、代码生成和内容处理场景的工程化提示词设计实战指南。汇集经过实战验证的模式、标准化迭代工作流与可度量的评估指标——不谈主观审美，只讲跨任务可复现的方法。提供 Markdown 与 PDF 两种格式，中英双语。
五模块提示词结构
角色、任务、上下文、输出规范与交互指引作为独立、单一职责的模块自由组合——简单任务可合并或省略。
经得起推敲的设计原则
正面指令优于负面约束、明确的推理路径、少样本示例规则，以及针对 4000 token 以上长对话的锚点重置。
六步迭代工作流
基线、冒烟测试、根因分析、定向修复、回归测试、发布——记录每一次变更，一次只修改一个模块。
Web 问答场景方案
上下文注入模板、高/中/低置信度自检与多轮证据摘要，让回答严格限定在给定材料之内。
大型项目智能体策略
面向 OpenCode、Cursor 等工具的项目级系统指令、模块级任务拆解、含影响分析的变更提示词与调试提示词。
四项硬评估指标
格式合规率、关键要素覆盖率、事实幻觉率与指令遵循偏差——每项均给出可接受阈值。
前缀标签体系
[Hard Constraint]、[Important]、[Optional]、[Background]、[Counter-example] 等前缀帮助模型更可靠地解析长提示词。
明细
五模块结构
| 模块 | 作用 |
|---|---|
| 角色 (Role) | 定义专业领域与视角 |
| 任务 (Task) | 明确陈述要完成的目标 |
| 上下文 (Context) | 提供背景、约束或历史日志 |
| 输出规范 (Output Spec) | 规定格式、长度、结构、风格 |
| 交互指引 (Interaction Guide) | 要求逐步推理或输出中间结果 |
- 克隆仓库——整个指南就是纯文本文件，无需构建。
- 阅读 Prompt Design Guide.md（GitHub 可直接渲染）或打开 PDF 版。
- 从提示词结构入手，再将设计原则应用到自己的提示词上。
- 用迭代工作流调优，并按评估指标为输出打分。
shell
git clone https://github.com/AkiroMusic/Prompt-Design-Guide.git
“提示词的质量直接决定输出的准确性、一致性与安全性。”