What Terraform brings to the table for us is the capability of calculating the delta of "I want those resources" and "these resources are actually there" by having a separate state stored e.g. as JSON in S3 to compare your code, the world as it should be and how it actually is. That takes away reimplementing that.

Why just not writing idempotent resource creation? Terraform also uses this to calculate a "plan" that shows the diff of your changes with reality, which really helps to figure out what happens to your RDS when executing, especially when more abstraction (in form of Terraform modules) is involved.

We used Terraform also in a situation where writing custom code would be "prettier" but would have required to write this actual vs desired state code ourselves and could save us the work of doing so.

The DSL of Terraform is sometimes quite cumbersome though as it's derived JSON and not some actual programming language.