This lines up with my experience. I will add one tip. If you know you're going to be exploring a topic. Ask for ChatGPT to outline a book, speech, or presentation on the subject. It may come up with some ideas you weren't expecting and help you eliminate certain paths you don't want to explore. Then you can provide a prompt. One of the prompts I've found most useful is variations of "write an informative and entertaining TED talk in the voice of [author or famous person] about [topic]." This always gives me new avenues to explore in my writing on the topic.
This is the best summary of how to approach prompt crafting that I've read, on the basis of my own experiments – thank you! Your warning about facts is important: GPT-3/ChatGPT will happily tell you things that are correct, partially correct or completely wrong in exactly the same tone. I have managed to get it to correct itself to some degree, which has helped – but it's still not perfect, and therefore users need to be very wary. Certainly using it well is all about refinement and iteration!
As a songwriter, I used ChatGPT to create song lyrics in the style of Harry Chapin about my father and I playing chess with the title "I Challenge You" (which was an expression that my father used to use to invite me to play with him). What came back were two verses and a chorus that I extended, put my own spin on and voiced with a wistful, nostalgic melody. What I found is that the ChatGPT AI offers a songwriter somewhat universal language that helps audiences quickly relate to a storyline so that, when I add my personal style to it, it decorates a structure that already feels familiar. If someone would like to see the original AI response, the finished lyric and hear the finished song ("Challenge"), then let me know.
The idea of seeing it more as prompt engineering like with visual generative AI than a text conversation is quite useful and makes a lot of sense. Thanks Ethan, that's an unlock for me.
One way to combat ChatGPT amnesia is to ask it to summarize what it's done so far before asking it to continue. The summary takes fewer "tokens" but usually contains enough info for it to stay on track (although it does sometimes repeat the same specific examples).
When trying to get ChatGPT to write in my voice I took an article of mine and asked it to analyze the style of my writing. It didn't do a perfect job, but it was a good starting point for my own list of style points.
I gave ChatGPT a story summary that I have in mind, then asked (asked not told because I am being polite 8-) ) it to create a full outline in the Hero's Journey template. After that I told it create a 3-chapter outline for each step in the Hero's journey. I deleted the repetitions in its outline and now have a working outline for my novel idea.
Great piece. I've spent the last week or so buried in learning more and playing with ChatGPT. Used it to assist in preparing an Abstract for an academic conference presentation. Proposing to present on Bullshit in Higher Education.
As other commenters mention, exploring ChatGPT came up with some excellent suggestions I hadn't considered. Pretty good abstract that I edited and fiddled with. And a very good outline for a 40 min. presentation.
Without exaggerating, this may be a time that we look back on in coming years. Remember when it was ChatGPT 4....
I used it to write a scene between two Dickens characters and found that starting with a simplified but specific prompt and then adding lots of edits and continuations got the best result, like directing a play almost.
I ran some cognitive neuroscience by ChatGPT the other day, asked it to cite references, and to my surprise, the two it provided were valid and topical.
Rizzolatti, G., & Craighero, L. (2004). The mirror-neuron system. Annual Review of Neuroscience, 27(1), 169-192.
Merleau-Ponty, M. (1962). Phenomenology of Perception. London: Routledge.
I am absolutely LOVING this revolution... so glad I'm alive to see this momentous shift in the way the world works. "Civilization-altering" may be understatement.
I didn't realise that you can prompt it with many sentences at once. Very useful to know.
I have corrected it when given incorrect replies and it happily updated it's response, complete with an apology! I think it's memory is the best attribute since you can really construct a conversation and it keeps to the topic (mostly).
One real concern I have in its current form is the confidence with which it replies. On certain questions it clearly uses a general trend to make up the answer, like asking for assistance to get the field ID of an MS Forrm, it assumed that selecting the entity in question will bring up a properties window with the ID listed there. This is probably typical of how AI is trained to generate answers it clearly doesn't have from answers that it does. My take on this is to know that you can get confident sounding wrong answers and to watch out what you take from it.
For me, prompt engineering makes a ton of sense in the context of generative art. But I haven't gotten great results by asking more specific queries in ChatGPT.
Instead, I actually find it more useful as a scaffolding tool. It's nice to find high level talking points that I might not have considered when writing organically. Then I might disregard most of the writing and just use it as a general framework.
It connects ChatGPT with the Wolfram Alpha knowledge base so that you actually can use ChatGPT to do math and to report actual facts in many areas of science, geography, etc.
This lines up with my experience. I will add one tip. If you know you're going to be exploring a topic. Ask for ChatGPT to outline a book, speech, or presentation on the subject. It may come up with some ideas you weren't expecting and help you eliminate certain paths you don't want to explore. Then you can provide a prompt. One of the prompts I've found most useful is variations of "write an informative and entertaining TED talk in the voice of [author or famous person] about [topic]." This always gives me new avenues to explore in my writing on the topic.
After a few iterations here is a prompt I came up with. It is a bit over the top, but I love throwing everything at it and seeing what sticks.
```
Write an essay on Organizational structure and design.
Use a creative writing style that uses:
- a strong opening
- Use symbolism
- Use foreshadowing with a twist at the end
- an anecdote
- use vivid imagery and descriptive language
- powerful emotion
- use metaphor, simile or analogy
- ask questions to engage the reader and encourage them to think critically
- output in markdown format with subheadings
```
This is the best summary of how to approach prompt crafting that I've read, on the basis of my own experiments – thank you! Your warning about facts is important: GPT-3/ChatGPT will happily tell you things that are correct, partially correct or completely wrong in exactly the same tone. I have managed to get it to correct itself to some degree, which has helped – but it's still not perfect, and therefore users need to be very wary. Certainly using it well is all about refinement and iteration!
As a songwriter, I used ChatGPT to create song lyrics in the style of Harry Chapin about my father and I playing chess with the title "I Challenge You" (which was an expression that my father used to use to invite me to play with him). What came back were two verses and a chorus that I extended, put my own spin on and voiced with a wistful, nostalgic melody. What I found is that the ChatGPT AI offers a songwriter somewhat universal language that helps audiences quickly relate to a storyline so that, when I add my personal style to it, it decorates a structure that already feels familiar. If someone would like to see the original AI response, the finished lyric and hear the finished song ("Challenge"), then let me know.
The idea of seeing it more as prompt engineering like with visual generative AI than a text conversation is quite useful and makes a lot of sense. Thanks Ethan, that's an unlock for me.
One way to combat ChatGPT amnesia is to ask it to summarize what it's done so far before asking it to continue. The summary takes fewer "tokens" but usually contains enough info for it to stay on track (although it does sometimes repeat the same specific examples).
When trying to get ChatGPT to write in my voice I took an article of mine and asked it to analyze the style of my writing. It didn't do a perfect job, but it was a good starting point for my own list of style points.
I gave ChatGPT a story summary that I have in mind, then asked (asked not told because I am being polite 8-) ) it to create a full outline in the Hero's Journey template. After that I told it create a 3-chapter outline for each step in the Hero's journey. I deleted the repetitions in its outline and now have a working outline for my novel idea.
Great piece. I've spent the last week or so buried in learning more and playing with ChatGPT. Used it to assist in preparing an Abstract for an academic conference presentation. Proposing to present on Bullshit in Higher Education.
As other commenters mention, exploring ChatGPT came up with some excellent suggestions I hadn't considered. Pretty good abstract that I edited and fiddled with. And a very good outline for a 40 min. presentation.
Without exaggerating, this may be a time that we look back on in coming years. Remember when it was ChatGPT 4....
I used it to write a scene between two Dickens characters and found that starting with a simplified but specific prompt and then adding lots of edits and continuations got the best result, like directing a play almost.
What people don't know if that this article is in fact written by ChatGPT
I ran some cognitive neuroscience by ChatGPT the other day, asked it to cite references, and to my surprise, the two it provided were valid and topical.
Rizzolatti, G., & Craighero, L. (2004). The mirror-neuron system. Annual Review of Neuroscience, 27(1), 169-192.
Merleau-Ponty, M. (1962). Phenomenology of Perception. London: Routledge.
I am absolutely LOVING this revolution... so glad I'm alive to see this momentous shift in the way the world works. "Civilization-altering" may be understatement.
I didn't realise that you can prompt it with many sentences at once. Very useful to know.
I have corrected it when given incorrect replies and it happily updated it's response, complete with an apology! I think it's memory is the best attribute since you can really construct a conversation and it keeps to the topic (mostly).
One real concern I have in its current form is the confidence with which it replies. On certain questions it clearly uses a general trend to make up the answer, like asking for assistance to get the field ID of an MS Forrm, it assumed that selecting the entity in question will bring up a properties window with the ID listed there. This is probably typical of how AI is trained to generate answers it clearly doesn't have from answers that it does. My take on this is to know that you can get confident sounding wrong answers and to watch out what you take from it.
For me, prompt engineering makes a ton of sense in the context of generative art. But I haven't gotten great results by asking more specific queries in ChatGPT.
Instead, I actually find it more useful as a scaffolding tool. It's nice to find high level talking points that I might not have considered when writing organically. Then I might disregard most of the writing and just use it as a general framework.
I don't know if you've seen this: https://writings.stephenwolfram.com/2023/03/chatgpt-gets-its-wolfram-superpowers/
It connects ChatGPT with the Wolfram Alpha knowledge base so that you actually can use ChatGPT to do math and to report actual facts in many areas of science, geography, etc.