I heard a great tip the other day on how to get good prompts: "This is what I want you to do, but I am not sure how to go about doing this, so you need to ask me the questions to find out". I've tested it, and in many cases it works really well!
Ethan, much gratitude for this. I led a session for almost a thousand people in the Americas of a global company last week with the promise that the event be about digital experience transformation and productivity. They are fortunate to have a supportive leadership team and I was fortunate enough to have had your experience to lean on. Your new prompt worksheet is amazing. I will add pedagogy and personalization to my work going forward. I’m surprised that I’m surprised that the key to me personally getting better results seems to often fall back on me giving better, more authentic context. It’s like I’m discovering more about my own mind by working with this alien mind as co-pilot! All the best. You are making a huge difference.
Have a look at https://structuredprompt.com/ - a free tool to help create and easily share structured prompts. It provides prompt framework templates such as TRACI - and has an AI assist mode to use AI to help you create better structured prompts. Provides an easy way to create a prompt library and make it very easy to share these with others.
Also - it can save prompts in JSON, XML, etc (you don't need to know anything about these - the tool takes care of the formatting) - GPT4 seems to like prompts being supplied in JSON. Claude's own prompt guidelines suggest using XML tags can help the LLM better understand what you want.
Appreciate that long term, you won't need to go to all this trouble. But right now, this is a good way to get your head around how to approach building a prompt to give you a superior output - IMHO.
This is great advice and something I've figured out over the past few months. I have lots and lots of structured prompts, several from Ethan's papers and examples, but more recently I just use it the way Ethan describes. If time is the trade off and one of the reasons AI is so useful is to be more efficient, spending an inordinate amount of time on prompting often defeats the purpose, especially if you can get 80% of the way there without it. One thing I would love to see is a comparison of output using customized instructions - I have several versions of these but often forget to change them up so I don't always remember if I'm getting the best results. If anyone has a link to differing outputs for similar prompts when you use a detailed customized instruction vs. without one, it would be useful to see.
I like the way you think Ethan. "Just do it.", "Just try it." "Test it." "Experiment!" "Play with it for a while.", was the drum I beat incessantly at my prior job, trying to get people to use the various chatbots to assist in their jobs. So many lightbulb moments happen with just the act of playing and experimenting with something. But there is quite a natural resistance in many people to try new things. It can be quite frustrating at times, when you can see how they will benefit so profoundly with just a little effort. And the current "fearful projections" don't help either... I think your posts help considerably to continue to circle back again and again to the ease of use and the benefits of doing so. So, thank you.
Looking at your examples, I also think that currently midjourney provides users with better images.
And nice to see that you run in the same issue that I do: when I ask him to change something in an image, he changes many details of it (the Firebird is now red, but the car on the left was also changed a little bit).
This is an excellent piece. I think the description of the "center" of LLMs vs the "Edges" is very helpful. I have found that some Chatboxes (Claude) respond well to gentle critiques and suggestions that some of the cited sources may be biased . An AI Chatbox user has to see him or herself as a coach who encourages the LLM to be more creative and dig deeper into sources. Easy answers are just that,easy answers.
I know Ethan knows this but, for the benefit of others, GPT-4 also powers Bing Chat’s Precise Mode. This post only referenced Creative Mode as being a free way to use GPT-4 and that is incorrect. Precise Mode gives more succinct, concise answers than Creative so it is a good option for research-oriented tasks vs creative ones
I get extremely good results from Claude2 and typically have ChatGPT 4 open in a separate tab to receive the exact same prompt. I like them both and often end up using the Claude response at least 2/3rds of the time. Why are you somewhat dismissive of Claude2?
FYI looks like cutoff dates now 4/2023 for modes: Default (+ App Voice Conversation), Advanced Data Analysis and DALLE-3. Remains 1/2022 for modes: Browsing and Plugins. https://twitter.com/MagnetsOh/status/1720663950017331601
The emotional prompt that is dramatically better than the others with ChatGPT is EP01: "Write your answer and give me a confidence score between 0-1 for your answer."
I heard a great tip the other day on how to get good prompts: "This is what I want you to do, but I am not sure how to go about doing this, so you need to ask me the questions to find out". I've tested it, and in many cases it works really well!
Ethan, much gratitude for this. I led a session for almost a thousand people in the Americas of a global company last week with the promise that the event be about digital experience transformation and productivity. They are fortunate to have a supportive leadership team and I was fortunate enough to have had your experience to lean on. Your new prompt worksheet is amazing. I will add pedagogy and personalization to my work going forward. I’m surprised that I’m surprised that the key to me personally getting better results seems to often fall back on me giving better, more authentic context. It’s like I’m discovering more about my own mind by working with this alien mind as co-pilot! All the best. You are making a huge difference.
Have a look at https://structuredprompt.com/ - a free tool to help create and easily share structured prompts. It provides prompt framework templates such as TRACI - and has an AI assist mode to use AI to help you create better structured prompts. Provides an easy way to create a prompt library and make it very easy to share these with others.
Also - it can save prompts in JSON, XML, etc (you don't need to know anything about these - the tool takes care of the formatting) - GPT4 seems to like prompts being supplied in JSON. Claude's own prompt guidelines suggest using XML tags can help the LLM better understand what you want.
Appreciate that long term, you won't need to go to all this trouble. But right now, this is a good way to get your head around how to approach building a prompt to give you a superior output - IMHO.
This is great advice and something I've figured out over the past few months. I have lots and lots of structured prompts, several from Ethan's papers and examples, but more recently I just use it the way Ethan describes. If time is the trade off and one of the reasons AI is so useful is to be more efficient, spending an inordinate amount of time on prompting often defeats the purpose, especially if you can get 80% of the way there without it. One thing I would love to see is a comparison of output using customized instructions - I have several versions of these but often forget to change them up so I don't always remember if I'm getting the best results. If anyone has a link to differing outputs for similar prompts when you use a detailed customized instruction vs. without one, it would be useful to see.
Please don't forget that this was a test using consultants from BCG. They have higher than average intelligence.
I like the way you think Ethan. "Just do it.", "Just try it." "Test it." "Experiment!" "Play with it for a while.", was the drum I beat incessantly at my prior job, trying to get people to use the various chatbots to assist in their jobs. So many lightbulb moments happen with just the act of playing and experimenting with something. But there is quite a natural resistance in many people to try new things. It can be quite frustrating at times, when you can see how they will benefit so profoundly with just a little effort. And the current "fearful projections" don't help either... I think your posts help considerably to continue to circle back again and again to the ease of use and the benefits of doing so. So, thank you.
Thank you for this article.
Looking at your examples, I also think that currently midjourney provides users with better images.
And nice to see that you run in the same issue that I do: when I ask him to change something in an image, he changes many details of it (the Firebird is now red, but the car on the left was also changed a little bit).
This is an excellent piece. I think the description of the "center" of LLMs vs the "Edges" is very helpful. I have found that some Chatboxes (Claude) respond well to gentle critiques and suggestions that some of the cited sources may be biased . An AI Chatbox user has to see him or herself as a coach who encourages the LLM to be more creative and dig deeper into sources. Easy answers are just that,easy answers.
Thanks a lot for this master class.
Do you know good repositories of prompts to look for examples (beyond the educational ones you shared). I am manily interested in content creation.
Thanks in advance
https://www.linkedin.com/posts/atheermahir_working-with-ai-two-paths-to-prompting-activity-7156048745418137600-JfHe?utm_source=share&utm_medium=member_desktop
I know Ethan knows this but, for the benefit of others, GPT-4 also powers Bing Chat’s Precise Mode. This post only referenced Creative Mode as being a free way to use GPT-4 and that is incorrect. Precise Mode gives more succinct, concise answers than Creative so it is a good option for research-oriented tasks vs creative ones
I get extremely good results from Claude2 and typically have ChatGPT 4 open in a separate tab to receive the exact same prompt. I like them both and often end up using the Claude response at least 2/3rds of the time. Why are you somewhat dismissive of Claude2?
FYI looks like cutoff dates now 4/2023 for modes: Default (+ App Voice Conversation), Advanced Data Analysis and DALLE-3. Remains 1/2022 for modes: Browsing and Plugins. https://twitter.com/MagnetsOh/status/1720663950017331601
Fantastic and useful
I have read on many sources that using negatives in prompts is ineffective.
Thanks so much as always. FYI From the paper page 11. https://arxiv.org/pdf/2307.11760.pdf
The emotional prompt that is dramatically better than the others with ChatGPT is EP01: "Write your answer and give me a confidence score between 0-1 for your answer."