Something like: Prompt GPT4 with basic rules (in several batches) and have it create JSONL prompt-completion pairs to use as fine-tuning for GPT3. Freeze that model then continue fine-tuning it with a new set of prompt-completion pairs for each adventure module.