Part 3/8:
Shapiro explains that traditional recursive prompting often involves fixed placeholders: the output of one prompt replaces a placeholder in a template to generate the next input. However, this approach is limited because the structure remains static. Instead, he suggests a more flexible system where the entire output becomes the new input—making the recursion more seamless and adaptable.
The goal, he notes, is to work backwards from the desired end state. For example, if the goal is to brainstorm ideas to solve climate change, what initial prompt would lead GPT-3 to produce that very instruction? Asking GPT-3 itself this kind of meta-question helps in designing prompts that can auto-generate others, creating a chain of prompts that evolve or expand upon specific topics.