I recently came across this great article about writing effective prompts for generative AI tools.
We’ve already touched on the topic of AI prompts, and more is on the way, but I thought I’d share two interesting insights here:
1. Formatting for accurate results
Structure your prompt into clearly defined sections. These can specify the context, type of desired output, basic instructions, style, constraints etc. Do this using clear headings and punctuation.
This is particularly useful to avoid a problem where certain prompts “scope out” of their intended context. Here’s an example:
“In section 2 of your answer, include one list of 5 bullet points”
To us this instruction is clear. But chatbots like ChatGPT have a tendency to let the second part of that instruction run wild, placing bullet points all over the place.
You can get increase the chance that your instruction stays in the right place with formatting like this:
“###In section 2 of your answer, include one list of 5 bullet points###”
This example also leads into the next point, that chatbots often interpret the logic of language very literally. So, where a human would understand “Include one X in your output” to mean “one AND ONLY one”, a chatbot interprets it as “AT LEAST one”, meaning you might get a lot more.
That brings us to the next point…
2. Understanding logic
Think carefully about the underlying logic of your phrasing. When humans communicate, we often use a lot of indirect or vague language. One annoying example: “CAN you do X?” usually means “I am asking you to do X” not “Is it possible for you to do X?”.
But chatbots can be pedantic.
For example, when we say “Do X or Y” we usually mean that you have to choose one: Either X, or Y, but not both. In other words, the choice is exclusive. However, in most contexts, an OR leaves open the possibility of choosing both options, and a chatbot might just be literal enough to take that option!
So very often, if you give a chatbot a prompt like “Do X, then Y or Z”, chances are you’re getting X, Y, AND Z.
So once again: be specific.
Understanding how to effectively handle AI prompts is why human expertise still has a vital role to play in AI content creation.
Anyone experienced any frustrating misunderstandings or unforeseen results from a prompt?