Eliciting Detailed Responses from LLMs
Introduction
Today, we explore the strategy of eliciting longer, more detailed responses from Large Language Models (LLMs). In contrast to seeking concise outputs, there are scenarios where depth, detail, and expansiveness enrich the interaction, such as storytelling, in-depth analysis, or comprehensive explanations. Our focus is on mastering the art of crafting prompts that encourage LLMs to "open up" and provide us with the wealth of information they're capable of generating.
The Challenge of Eliciting Detailed Responses
While LLMs are almost never succinct unless specifically asked to be, they are also almost never very thorough unless, again, they are specifically asked to be.
Here is a simplistic example prompt:
This would give you a response that's roughly 300-450 words long (depending on the LLM being used). Considering that the actual story is close to 25,000 words, that's not much detail. What if we wanted the LLM to go deeper?
Simple Approach: Requesting Detailed Responses Directly
A direct method to obtain more detailed information is to explicitly ask for it within the prompt's constraints. This strategy relies on clearly communicating your expectations for a detailed or lengthy response.
This will start to push the response close to 700-800 words, which is a 50% improvement on what we had so far, not bad.
Complex Approach: Iterative Enhancement
For even greater depth and structure, we can employ an iterative enhancement strategy. This involves first asking for an outline or skeleton of the desired content, then requesting the LLM to elaborate on each part of the outline, and finally asking for further enhancements or details for each section.
Step 1: Request an Outline
Step 2: Request Elaboration on Each Part
This iterative approach not only encourages an initially structured and detailed response but also builds upon it to achieve an even greater level of depth and elaboration. You can, of course, add a Step 3 here and repeat the same approach again, but you'll start running out of context windows at that point, especially with this specific topic.
