I’ve been playing around with Playground / GPT4 to generate bulk personalized sales emails for my brother’s start up and when I let the A.I. do a lot of work and make decisions I find the output to be vary wildly.
I’ve started being much more prescriptive and getting better results. Here is one example:
Based on a JSON object of a person’s LinkedIn and the top 3 restaurants in their city, create a sales email subject line and icebreaker. Only include these two parts in your response. Do not include any other information in your response.
1. Determine current city
2. Determine top restaurants
3. Enter information in the exact format below
*Use this exact syntax, do not add any sentences or words: *
*Subject line format: *
[restaurant 1] or [restaurant 2]?
*Ice breaker format *
Hey [first name],
*I can offer a $125 gift card to one of the best restaurants in city for a meeting: *
- [restaurant 1]
- [restaurant 2]
- [restaurant 3]
Have you all found this to be true? I don’t like when the AI has too much freedom.
Depends on the task. If it is a creative task and you aren’t really sure what you need, it is better to let the AI have a little bit more freedom with less instructions. If you know the task and have a clear output or outcome that you want, then it is better to be as descriptive as possible.
Personally, to avoid being flagged as spam and an attempt to connect more deeply with the cold call, I’d scrape their profile info and add a small personalization note and let the AI play around with the wording a little bit. Then just do some testing to ensure you are getting good results.
If it is a creative task and you aren’t really sure what you need, it is better to let the AI have a little bit more freedom with less instructions. If you know the task and have a clear output or outcome that you want, then it is better to be as descriptive as possible.
Yeah I think this is very true. If you have some type of expertise and know exactly what you want, better to tell the AI. But if you’re unsure, let the AI do it’s thing.
I like you’re approach as mentioned at the bottom too.
You’re right. It’s like compressing a lot of information into something concise so that the AI doesn’t get confused. But do you feel that AI doesn’t handle these matters?
The other responses to the question are largely correct, ie if you want some creativity than it’s better to not be so detailed with your prompt. However, what I have found personally is that with gpt-3.5-turbo, the output is much better when I include detailed descriptions and instructions into the prompt. But with gpt-4 I have found it is actually better to not be so detailed in your prompt.
It’s true, we call this Structured Prompts and we have writting prompts in this style for last few months, we also design a few prompt templates and meta prompt helper to make things easier. I have this methods open-sourced in GitHub - yzfly/LangGPT: LangGPT: Empowering everyone to become a prompt expert!🚀 Structured Prompt，结构化提示词，结构化Prompt and it earned 2k stars until now!