If your prompt fits in a tweet, it is too short
A prompt that fits in a tweet is at best a jumping-off point
If you ask simple questions to ChatGPT, it gives answers that can best be described as meh or bleh. ChatGPT answers like an average internet human, and hence, this output is usually a lump of boring text.
But, ChatGPT is very versatile, and good at following directions if you can give it good directions. So, create prompts with lots of specific suggestions so ChatGPT will give a better answer.
Here’s an example from Wharton MBA School Professor, Ethan Mollick. Prompt #1 is what you would normally do. Compare that with prompt #5.
To drive the point home, here’s another example from Ashish Kulkarni. Ashish has a great example of a “detailed prompt” in Put Me Out of a Job – 1. This is his prompt for how a student can use ChatGPT to learn the principles of economics, along with an explanation of the various components of the prompt, and why it is so detailed:
Remember, LLM’s work best when you give really detailed prompts. Note the following:
I began by giving some information about myself – my limitations as regards economics, where in the world I come from, and what my interests/hobbies/passions are.
I specified what I’m looking to learn from the LLM.
I specified the quantum of output required (thirty classes).
I specified how broad the output should be.
I specified how I would like the answer to be customized for me
I would like to learn about economics by relating it to what I like to read about in any case (use examples from the Mahabharata)
I would like to learn about economics by relating it to real life situations.
It is amazing to me, regardless of how many times I experience it, that it “gets” what I really mean in spite of having phrased my question using really bad grammar.
The specific examples aren’t the point, the idea is the point. Learn calculus by relating it to mandala art, for example. Learn history by relating it to dance forms. Learn geography by relating it to food from different parts of the world. A teacher in a classroom cannot possibly do this for all the students, because of the size of the class, and because a teacher cannot possibly know your hobby in as much detail as you can. Make good use of AI!
Should the examples from the Mahabharata be chosen for how prominent the examples were in the text, or should they be chosen for their relevance to economics? My preference is for the latter, and I made sure the LLM knows this. Ditto for the real life examples.
I ended with a meta-prompt, that will stay true for the next thirty (or more questions) – ask if I need to learn more, and only then proceed with the next class.
I now regularly use long prompts for serious ChatGPT use. I still use short prompts for for quick answers to simple questions. After a little bit of experience you’ll get a feel for when you can use a short prompt and when you need to be elaborate1.
This is going to be a running theme of this newsletter: Using ChatGPT effectively is an art, not a science, and after some practice, you start getting an instinct for what kind of questions it is good for and how best to ask those questions to get the best answers out of it.
Love the contrast between two examples. Where Ethan's prompt shows the example of someone experienced: should know the babbel hypothesis to ask for it, Ashish's prompt shows the example of someone just starting: intent replaces experience.
I often use ChatGPT for proofreading my writeups. I have noticed that it sometimes offers to correct a mistake and the corrected version is the same as the original version. E.g. following is one of the suggestions given by ChatGPT for my write up.
In the paragraph discussing hedgehogs and foxes, there's a small grammatical error: "And it does this rolling-up-in-a-ball trick very well" could be improved for clarity and consistency: "And it does this rolling-up-in-a-ball trick very well."