Bounding the Capabilities of Large Language Models in Open Text Generation with Prompt Constraints