As you'll likely have found out, or will, this is difficult to scale, keep reliable, and ensure correctness.<p>I've written a tool to go from declarative config to code: <a href="https://github.com/hofstadter-io/hof">https://github.com/hofstadter-io/hof</a><p>We are using LLMs to write the JSON or Yaml part to be the input to hof, which can now interface with OpenAI and Google models directly now. This opens up the possibility to have modules for code gen'n anything with an LLM interface with all the prompts packaged along side it<p>There are interesting intersections of LLM and declarative code gen for sure!