On the topic of procedural, one thing I experiment with is having the llm part of the procedural loop.

Sort of writing a narrative on top live.

Unfortunately, local models are still a bit slow and weak but was interesting to see what it came up with nonetheless.