Design – Should I always start with UML as a good practice when designing classes

When I have to return to a section of code I've not worked on for a while, I have difficulty remembering how it worked. I spend a lot of time reviewing the header files for the relevant classes and reading the comments I placed along the way in the source files.

Imagine how the poor guy who comes after you is going to feel -- he doesn't even have the benefit of having once known how your code works. Instead of trying to decipher your code, you should be reviewing the documentation that you wrote for the module in question. That documentation should offer a reasonably accurate view of what the module does any why. It's not "the module starts by initializing three arrays using a triple for loop...", but instead: "this module retrieves the data collected by the Fabulotron's main sensor, rearranges it into standard Neopolitan (chocolate, vanilla, strawberry) format, and delivers it to the Analysis module."

In a perfect world, you'd have a design document that sets out the various modules in the system and describes their respective responsibilities, and each of your modules could just refer back to that document to explain what they do: "This module provides the Fabulotron data collection service as detailed in section 4.8 of the design document: http://fabulotron.org/design/section4-8.html." If you don't have something like that, start writing down an overview of each module as you work on it. You don't need to write a book -- a few paragraphs are often enough to get you oriented.

When I introduce changes, sometimes I realize half-way that what I'm trying to do will break things somewhere else (or worse, it shows up only at runtime as a surprise). I revert and start doing it differently, only to find out I neglected the influence on some other component.

That might be an indication that your modules are too interconnected. The more independent you can make your modules/classes/units, the less likely it will be that you'll run into this kind of problem. Try to make the interfaces between modules as explicit as possible, and try to limit them to just what needs to be there. An interface is a contract -- if you know that some module lives up to its obligations as specified in the interface, you don't need to know anything else about it. That prevents changes in the module you're working on from affecting other modules.

I wish there was some "architecture diagram" where I could see how things get done

Using standard parts can help in this respect. C++ provides standard parts in the form of the Standard Template Library, and using those parts where appropriate lets you work at a higher level of abstraction. If you've written your own code to manage data structures like lists, you and those that follow you will constantly have to read the source to figure out what's going on. If you instead use standard parts provided by the STL, any C++ programmer will quickly be able to tell what your code is doing without digging into your data management routines.

Another kind of standard part comes from design patterns. They're standard concepts that can be used as a shorthand to explain how the relationship between two objects works.

There is no fixed order. It is all based on what you think is most useful to you/your team.

Initially, there should be some collection of requirements. If you start with SW architecture once someone tells you to do a software for X, then you already missed the most important step there.

During the requirements engineering phase, not all UML diagrams make sense yet. But some do make a lot of sense, in particular use-case diagrams. I have often seen use-case diagrams as the basis for discussions from which the requirements are then determined.

Similarly, state machines are already popular on the requirements level as well.

Once you have a set of requirements and actually start thinking about your architecture and later the design, more and more diagrams will start to make sense. Obviously, the class diagram comes in handy on lower levels of detail, but others, like sequence diagrams, may also make sense if you have complex sequences in your architecture.

Are you developing a distributed system, or something where components have to communicate a lot with each other? Try out the communication diagram.

Unless your process somehow requires you to use a specific UML diagram, there is no one forcing you to do so. Always remember that you do not model your software in UML for its own sake. If you don't feel like it'll help you, then simply don't do it.

Obviously, it takes some experience to decide this. I suggest you at least invest time to try out state machine, class, activity, and sequence diagrams, as those are the most popular. If you have enough time, take a look at the others as well.