Brief look into how an app file structure with 45k LoC and 436 files evolved. Without a pre-meditated plan, our file structure grew as needed, the only exception was to make sure global, popular modules ( lots of imports ) got broken up to avoid compile time hell.
How we structure
|-- Main.elm -- obligatory. Fun Fact: Ours is 802 lines and counting  |-- Login.elm -- simple subsystems are root level modules  |-- Notification.elm -- another simple subsystems |-- Alfred.elm -- OBSOLETE: all helper functions  |-- Alfred -- helper folder, so named to prevent namespace clashing |-- Components -- stateless/stateful/hybrid components  |-- Membership -- large subsystems group into a folder | |-- Api.elm -- OBSOLETE: all http helper functions  | |-- Forms.elm --  | |-- Forms |-- Membership.elm -- large subsystems generally have a top level TEA module to distribute msg |-- Model -- aka Repository, business logic for get/set of data  |-- Person -- another large subsystem | |-- Forms | |-- Forms.elm |-- Person.elm |-- Search -- YALS [YASD?] |-- Search.elm |-- Ports.elm -- all our ports in one place |-- Types -- non domain specific, non module specific, custom types | |-- Api -- all endpoint types |-- Types.elm -- global Type, wait... I can explain  |-- Validators -- common validation helpers |-- View -- | |-- Components -- UI components | |-- Fields -- domain specific UI components to make building forms easier
Why we structure it so
Some commentary on our file structuring.
We do not care about file size. We care a lot about responsibility. When considering what goes into a file, the only consideration is: Does this module fulfil one responsibility. When a module does more (or less) than this, coupling occurs. Every. Single. Time. Having said that, Main.elm is a bit of a mixed bag though, it handles routing, page loading, executing jobs and generally delegates everything.
There is a bunch of files at root level, they are typically directly imported by Main or they are global like Types.Elm or Routes.elm. The responsibility of these are clear enough to be self contained in one file.
The elm-guide says don't use reusable components but doesn't define what it is. In our experience, certain patterns of components have emerged. We discuss these in Components. Overall, these are very useful in keeping modules decoupled with single responsibilities.
So we used to group helper functions in one module. However, as the avid reader would know from having already read Compile Time, this is a really bad idea in Elm because a helper module by definition is a toolbox for the rest of your app.
touch Alfred.elm elm-make ... Compiling 105 modules (Have a nice day!)
Ah compiler, always so friendly.
Instead, we now have a folder called Alfred and try to break our helpers up into modules that are commonly used together.
Forms, what a mess.
We don't use opaque types and it's already bitten me twice. As a team grows and introduces more junior/non-domain experts, It's definitely worth making domain models opaque. This folder contains all domain specific helpers (e.g Membership.getDependants will sort by age). Unfortunately, domain logic is also tied in page updates, in the page views and validators.
So Type.elm contains Id tags. The kind that looks like this:
type alias PersonId = Int type QuoteId = Person PersonId | Member MemberId
The compile time on this is tolerable because these IDs are defined by the backend and rarely change.