Code generation is a dead end for review: you cannot diff intent. Spec generation is different. A generated spec is a small, structured artefact you can read, validate, and reject before it ever reaches a renderer.
In dfs the block builder already emits a spec. Pointing an AI at the same schema means the model outputs something the runtime already understands, no glue code, no prompt-to-component shim.
The work ahead is constraints: teaching the model the catalog, the allowed root nodes, and the prop types so its specs compile on the first try.