I have been using AI to write "snippets" of code for a while now. I started with Github Copilot but now use Claude more and more. Generating code snippets is useful but it's hardly revolutionary... it's just a like better form of intellisense as I see it. I find tools that are able to generate 90-100% of the code much more interesting. Although some human rework may been needed (or several iterations with the LLM) this approach obviously has huge potential. A typical project for me might be to build a modern REST API in front of one ore more legacy systems, adding stuff like security and observability to the API layer and simplifying future integrations. Sometimes an API might also bridge an enterprise data model and the data models used in legacy applications. Building a "standalone" fullstack application does not require knowledge of obscure legacy system APIs or specific enterprise data models, and I guess that makes them easier to work with, especially if you are basically just doing CRUD for a simple data model. In cases like this, there is little context to deal with apart from a simple specification that can be written by the developer, and LLMs have been trained on tons of similar code. I used to think that the amount context needed for the projects I work on makes them unsuitable for large scale code generation - but I'm not so sure anymore. I think LLMs like Claude Sonnet 3.5 might be advanced enough to produce some fairly decent API code if I feed it documentation for the legacy system APIs and enterprise data models, and provide the right instructions. Yet I have never heard about anyone doing something like this... it always standalone fullstack applications being built from scratch. Why is that?