Four projects, two where it clearly helped and two where we were mostly writing adapters for a CRUD application.

Where it earned its keep

A domain with real rules — pricing, eligibility, scheduling — benefits enormously from being testable without a database or an HTTP server. Those tests run in milliseconds, express business language, and survive infrastructure changes untouched. That is the whole promise and it delivers.

Where it did not

An application whose logic is "validate, save, publish an event" gains nothing from a port and an adapter. You have added two interfaces and a mapping layer to protect a domain that does not exist yet.

My rule now: start with the layering that lets you test the interesting part in isolation, and add ports when a second implementation actually arrives. Building the abstraction for an imagined future adapter is the most reliable way to get it wrong.