Overall, you wouldn't use protocols, records, and custom types very often. Using normal maps with regular functions will do for most application domain specific logic. Still, I'll try and explain when you'd want to reach for them.
Use deftype when you need to implement custom data containers which need mutability under the hood, or some form of data encapsulation. This is mostly for very primitive constructs like data-structures or reference types.
If you have a function which is called with different things at runtime, and what it does must be specific to the things it was called with. Then you want to use protocols or multi-methods. If you can manage with choosing what to do based on the type of the first argument, use protocol, if not, use multi-methods.
If you have a function that should do different things based on the type of a Map. So say you need to tag a Map as representing something, like a Person, a Car, a User, etc. And you have functions which will be called with Maps of more than one type, and you want them to choose what to do based on that type, then use defrecord.
If you have a function which needs to do something based on the number of arguments it was called with, use multi-arity.
Finally, if you find yourself writing many functions that start with the same name, but end with a kind of discriminator which relates to the type of what you call it with. Say:
add-item, this might be a good indicator that you can model this instead using protocols or multi-methods and a single
I'll start with deftype. You would reach for a deftype the least often. A deftype allows you to implement a new abstract data type (ADT). It is useful for creating new types of data containers. Basically, anything which needs to encapsulate data (mostly mutable data), and provide access to the data only through a safe interface which enforces all invariant of the underlying data in whatever way makes sense for that type.
For example, if you wanted to add a new data-structure, say you needed to implement a doubly-linked list, deftype would be what you'd use.
Now, you almost never will need it, since most of the useful data containers have already been implemented for you either in core, in interop, or as a library. For example, Java provides a doubly linked list already: https://docs.oracle.com/javase/8/docs/api/java/util/LinkedList.html
To give a better intuition, deftype allows data encapsulation. In general, you don't need data encapsulation in Clojure, because the default data-structures and bindings are immutable, so there is no danger in exposing the data to outsider as read-only. That said, for implementing the data containers themselves, like the immutable data-structures or various reference types like Atom, you're going to need mutation in order to provide a space and time efficient implementation. In such case, you shouldn't let outsiders be able to touch the data freely, because they could easily break the required invariant. So instead, you want to provide an abstract interface, such as for a doubly-linked list, you might have add-first, add-last, remove, next, prev, etc. So again, deftype is useful for adding any form of data container to the language, which you'll almost never need to do yourself.
You would reach for defrecord when you need to create a custom type, in order to add semantic meaning to some Map. A record is really just a Map, but with the type for it replaced by the record type name. So instead of having type Map, it will have the type as the record name.
Say you wanted a Map of type Person, you'd use defrecord for that.
This is only really useful when you need a piece of code to at runtime do something different based on the type. So if there is a piece of code which will receive Maps of different types, and you want it to do something different for each type AS WELL as you want better performance.
I'm saying as well, because there's actually many ways to provide this dynamic functionality. One is with using native types, and is provided by records and protocols. The other is with manually modeled types and the use of defmulti. The latter is more powerful, because you can represent the type as anything you want, and the type can even be inferred from the structure or the value of the data itself. On the other hand, it will be slower.
defrecord can only allow functions to do different things based on the type of their first argument, and the structure and values of the record can not dictate its type. This is often called nominal typing. Your record is of a certain type because you named it that way explicitly. On the other hand, defmulti allows for duck typing, in that a Map can be of a certain type because of its inherent structure and/or the values it contains.
You reach for defprotocol when you want functions which do different things based on the native type of their first argument. Such as say a function which will be called with different records, and should do something different for each different record type.
As I explained when talking about records briefly. If you want a function to do something different not just based on the type of the first argument, but also because of the type of the other arguments, or because of the structure of the data in any/all arguments, or because of the values of the any/all arguments, you would use defmulti instead of defprotocol.
Also, if you just want a function to do something different based on the number of arguments it was called with, you can just use multi-arity functions.