During my adventures of designing class structures, I’ve noticed that it’s very helpful to start with writing some pseudo-code. That means: I start with “writing” some general fragments of application’s code on a highest level, play with it, and discover the elements that are appearing – in fact, the elements that I – as a programmer – would like to use. It’s a very good starting point for designing general structure of modules and their interactions. After few iterations the whole structure starts to look more like a full system of classes. It’s a very flexible way to design parts of code. You can call it a programmer-oriented design.

老实说，最好是回头看看流程图和序列图。有大量的好网站告诉你如何做到这一点。当我考虑如何将程序分解为类时，我发现这是非常宝贵的，因为我确切地知道程序需要输入、计算和输出什么，并且每一步都可以分解为程序的一个部分。

我用于初始设计(得到类图)的步骤是:

Requirements gathering. Talk to the client and factor out the use cases to define what functionality the software should have. Compose a narrative of the individual use cases. Go through the narrative and highlight nouns (person, place, thing), as candidate classes and verbs (actions), as methods / behaviors. Discard duplicate nouns and factor out common functionality. Create a class diagram. If you're a Java developer, NetBeans 6.7 from Sun has a UML module that allows for diagramming as well as round-trip engineering and it's FREE. Eclipse (an open source Java IDE), also has a modeling framework, but I have no experience with it. You may also want to try out ArgoUML, an open source tool. Apply OOD principles to organize your classes (factor out common functionality, build hierarchies, etc.)

大型项目的问题在于您无法监督组件之间的所有交互。因此，降低项目的复杂性是很重要的。类图和序列图对于这个设计阶段来说太详细了。

首先试着从更高的抽象层次来思考。考虑主要组件及其职责(它们与其他组件的接口)，查看一些架构模式以获得灵感(不，不是设计模式，这些太低级了!MVC和多层是架构模式的例子)。 对于相当大的项目，这样的视图应该有大约3-5个组件。

只有这样，你才能放大到某个组件，并尝试设计它。现在我们到了设计模式和类图的层次。试着把注意力集中在项目的这一部分，如果你发现你需要向其他组件之一添加一个职责，就把它添加到你的文档/待办事项列表中。不要在这个时候浪费时间思考它们的含义，因为它们的变化太快了，当设计更加可靠时再进行检查。

此时您不需要完整地设计每个组件，尽管使用一段代码来实现未实现的组件接口并生成简单但有用的响应可能是明智的。这样，您就可以一次开始开发(和设计)一个组件，并对其进行合理的测试。

当然，当新组件完成后，您应该在继续之前测试它们如何(以及是否)相互集成。

简而言之: 采用面向对象和信息隐藏原理，并将其提升到另一个层次!

PS: 在设计的时候画很多草图，就像真正的建筑一样!

PPS:试着从不同的角度来看待这个问题，跳出固有的思维模式(尽管固有的思维模式可能是正确的)，与同行讨论可能会非常有用……午餐时你们有话可聊。

我在实际项目中使用的成功的技术是责任驱动设计，灵感来自Wirfs-Brock的书。

从最顶层的用户故事开始，与同事一起，在白板上勾勒出它们所暗示的高级交互。这让你对大模块有了初步的了解;重复一两次高级CRC-card(游戏邦注:如play)，你应该已经稳定了一个主要组件列表，它们的作用以及它们如何相互作用。

然后，如果任何职责很大或很复杂，那么细化这些模块，直到有足够小且简单的东西成为对象，方法是在模块内执行由更高级别交互确定的每个主要操作的交互。

知道什么时候该停下来是一个判断问题(只有经验才能决定)。

恐怕人们不愿意听到这样的回答。无论如何，让我陈述我的观点。

OOP should be viewed as one of the paradigms, not as the superior paradigm. OOP is good for solving certain kind of problems, like developing a GUI library. It also fits into the style of software development usually followed by large software companies - an elite team of designers or architects lays down the software design in UML diagrams or some other similar medium and a less enlightened team of developers translate that design into source code. OOP offer little benefit if you are working alone or with a small team of highly talented programmers. Then, it is better to use a language that supports multiple paradigms and will help you to come up with a prototype fast. Python, Ruby, Lisp/Scheme etc are good choices. The prototype is your design. Then you improve on that. Use the paradigm that is best to solve the problem at hand. If needed, optimize hot spots with extensions written in C or some other systems language. By using one of these languages, you also get extensibility for free, not just at the programmer level but also at the user level. Languages like Lisp can dynamically generate and execute code, which means your users can extend the application by writing small code snippets, in the language that the software itself is coded! Or if you choose to write the program in C or C++, consider embedding an interpreter for a small language like Lua. Expose functionalities as plugins written in that language.

我认为，大多数时候OOP和OOD所创造的软件都是过度设计的受害者。

总之，我喜欢的写软件的方式是:

使用动态语言。 用这种语言编写设计(原型)。 如果有必要，使用C/ c++优化某些领域。 通过实现语言本身的解释器提供可扩展性。

最后一个特性使软件能够轻松地适应特定用户(包括我自己!)的需求。