[spaced-out]

No coding required...yet.

I've been playing around with generating stories with ChatGPT for a while and...English (or any natural language) is really bad at being specific. I've made progress by learning some specific words to describe the type of scene I want and how much of it I want ChatGPT to generate (such as a scene for just that evening verses a few paragraphs describing weeks of traveling). I've also started getting some intuition for when I've given ChatGPT too much info (it'll cram all the facts in in weird ways) and too little info (it'll get really random and start inserting new characters and stuff).

Having a way to manage the meta aspects of story generation would be a big help.

[truculent]

I’ve wondered if it would be helpful to teach the model a micro-DSL for the given task at hand, and use that for greater control and precision.

[antibasilisk]

It's helpful for doing things like mathematical reasoning, since the LLM can't really do it directly. A DSL would allow you to offload that reasoning.

[ly3xqhl8g9]

A node editor would be interesting [1], although it might be a recipe for spaghetti prompt [2].

[1] https://en.wikipedia.org/wiki/Node_graph_architecture

[2] https://blueprintsfromhell.tumblr.com