Prompt Structure in Conversations With Generative AI
Metadata
- Author: Nielsen Norman Group
- Full Title: Prompt Structure in Conversations With Generative AI
Highlights
We identified four components that were often (but not always) part of the AI prompts used in our study:
• Request
• References to either previous bot answers or to external sources
• Format
• Framing
Analysis of an AI (View Highlight)
ChatGPT has a global setting that allows users to enter format specifications for all ChatGPT’s answers. Such a setting is unlikely to be helpful because many people use AI bots for both work and personal purposes, and they will likely need different kinds of formats for these settings. (View Highlight)
Support multiple prompt types when building apps around AI
Some framing prompts followed the roleplaying model, popularized in news and social media. Word phrases used to introduce such prompts included Be a, Imagine you are, Pretend you are, Act as a. The intent behind these phrases was to reduce ambiguity by setting upfront the domain of the query. (View Highlight)
Prompts like Best of or Give me more are very inefficient — if you find yourself using them, it’s a sign that you may be better off with a search engine. (View Highlight)
Search engines are better than AI at showing you the “Best of” something
In particular, designers should: • Help users add framing to their prompt by proactively asking relevant questions and providing suggested followups that better define the problem instead of launching into a general-purpose response. (View Highlight)
Great apps built on top of AI should help their users add framing to their prompts
Add UI elements that allow users to select among certain possible formats (e.g., list, table, visual) for the bot response. Allow them to also add additional attributes such as the number of words or the length of the response. (View Highlight)
Make it obvious the different formatting options that are available with AI
Be sensitive to filler prompts and Can you prompts to determine what kind of response they require and whether they are meant to finish a conversation or, respectively, understand the bot’s capabilities. (View Highlight)