In my experience, the value doesn’t really shift from “coding” to “architecture” — it shifts toward understanding constraints.
On platforms like macOS, a lot of architectural decisions aren’t about clean abstractions but about what the system actually allows you to do (WindowServer rules, accessibility APIs, focus behavior, sandboxing). LLMs can generate plausible code, but they don’t feel where the platform pushes back.
The hard part is still knowing which designs will collapse once they meet real OS behavior, not writing the glue code itself.
On platforms like macOS, a lot of architectural decisions aren’t about clean abstractions but about what the system actually allows you to do (WindowServer rules, accessibility APIs, focus behavior, sandboxing). LLMs can generate plausible code, but they don’t feel where the platform pushes back.
The hard part is still knowing which designs will collapse once they meet real OS behavior, not writing the glue code itself.