I've been working on a new LLM prompt format and accompanying dataset for finetuning. The idea is that the "main character" being prompted always has to perfor

I've been working on a new LLM prompt format and accompanying dataset for finetuning. The idea is that the "main character" being prompted always has to perform an action/function. So even "saying" something to the other participants of the chat is a deliberate action. Actions can be something like: ### actions: 1 recollect(query) -> {memory_id: id[]} 2 think(thought) 3 say(message) 4 memorize(category, subject, memory, replace?: id) The LLM then has to respond, like so: {"action":2,"thought":"Some internal thought","queue_action":3} The JSON overhead is pretty negligible.