Social Media and Structured Prompting – tips for ChatGPT prompting – Australia
A brief tutorial on how to do ChatGPT structured prompting for social media managers and marketers.
SOCIAL MEDIA and STRUCTURED PROMPTING for ChatGPT: in this quick tutorial I go through structured prompting for #socialmediamanagers. First the timestamps then the video!
- ✅ 00:20 Custom Instructions for Social Media Managers (setting them up)
- ✅ 01:14 Starting 4 prompts the basics ie provide context and the query then explain the output (format and style)
- ✅ 02:50 Few Shot Prompting is leading with examples Please provide a list of 15 xyzs here are 2 examples
- ✅ 03:50 Chain of Thoughts to explain step by step social media decisions so the #artificialintelligence doesn’t just give output with no explanation!
- ✅ 04:54 Tree of Thoughts for roleplay in social media content for example roleplay an unhappy customer leaving reviews and then evaluate my (social media manager) answer and give me tips for better answers. ChatGPT can give the response but we don’t learn by copying and pasting do we? #AIIsMyMentor
- ✅ 05:47 Chain of Density for social media extracts and short succinct informative pieces.
- ✅ 07:19 AI is My Mentor (social media mentor) more info
- ✅ 08:22 Catastrophic Hallucinations in social media prompting = ask for summaries!
- ✅ 09:20 Advanced Parameters: Temperature to reduce fibbing
- ✅ 10:01 Advanced Parameters: delimiters to upload a 100 reviews.
- ✅ 10:34 ask me questions, critique my prompting, summarise our chat so far, etc.
I hope you found this #chatGPT tutorial on structured prompting useful.
I have a 2 day Artificial Intelligence and Marketing course coming up in Sydney (specifically for marketers) in Feb 2024. https://laurelpapworth.com/workshop/marketing-ai-course/
Transcript of Structured Prompting for ChatGPT for Social Media Managers
Hello, my name is Laurel Papworth, and I’m an AI lecturer in algorithms, data sets, and artificial intelligence. I also teach social media. Where to start? Structured prompting for social media. You may end up with what’s called a “series of prompts” that’s around 15 prompts. Always make sure your custom instructions say that (you’re in the first box) that you’re a subject matter expert in social media and that you have a number of clients, and then list the industries. This will give the AI an idea that it needs to be versatile in its answers specifically to particular industries. Now, if you’re a social media manager in, for instance, hairdressing, then you would just say hairdressing. But if you’ve got a hairdressing client, accountant, a small business selling car parts, then list those businesses in the first box. In the custom instructions, the second box, say that you need a variety of styles and tones. And also, I like to say that I want Australian grammar and spelling. It doesn’t always work, but I ask. [00:01:14.850]
When you get into the prompts, rather than just saying, I need a social media marketing plan, you give, the first prompt is a construct of four. I always give context, query, style, output, and then a question. The first one is, I need to write a social media strategy for—and strategy, by the way, there is output, for a hairdressing salon. There’s context. Can you help me write this strategy? There’s your query. The strategy should be easy to read, dense, step-by-step, something along those lines. I’ll go through what those keywords mean in another video. Once you have those four and then you ask the question, What else do you need to know? I always start by throwing it back onto ChatGPT. If I march in and say, List me 10 things, or do this, I might miss something. So if I ask what you need to know? ChatGPT will tell me. Usually, it’s along the ideas of audiences, products and services, any analytics you have, things like that. It’ll ask you for them. Those four things: context plus query move you to style and output. Output are things like I need it in a table, I need dot points, and then style is in a friendly voice, in a professional voice, in a lively, entertaining voice, something like that. Once those four have been done, I move into few-shot prompting. [00:02:55.420]
Few-shot prompting is giving the AI an example. Here’s an example of a script that we did last week on how to put a pool into the garden. Please write a script in the same style on how to build a retaining wall. That way, if you’re doing YouTube, educational videos, or something like that on a particular topic, the AI will read the original script or scripts if you have a few, and then be able to extrapolate out what another script would look like. In my case, the AI sometimes comes back and says things like, Would you like me to include your jokes? Yes. Are they actually funny? I don’t always use them, but I do find them funny. So few shot or end shot prompting means to give an example. [00:03:50.610]
Chain of thoughts for social media would be along the lines of, I have a prompt, I have an output, give me the steps in between. I always want the steps because I want to understand why it’s come to that decision. So if I say, I’m looking after this company, here’s their information on their numbers, on their target audiences, on the advertising spend that we have, all that thing. Should I focus more on LinkedIn, X, Instagram threads, or Facebook, or something else? I don’t just want an answer that says Facebook. So that prompt, which one should I focus on? Please give them to me in a list prioritised, is the output, and then explain to me step by step your decisions. That way I get a full answer of your audience is not really on LinkedIn, or you have very visual products, so therefore, Instagram would be good for you. I need to know what and why the AI is thinking the way it is.
[00:04:54.690]
With Chain of Thoughts, we can then move into Tree of Thoughts. Tree of Thoughts is simply asking the AI to role play as multiple members of a panel or a team. I might say, I want you to role play three social media experts and argue which is the best platform for us to be on. You’ll get a deeper answer than choose a platform or give me a list. If I don’t want three of the same type, I can also use tree of thoughts or please role play as three different people. I want you to roleplay the CEO, CFO, and the chief marketing officer. Or I want you to roleplay the client, the sponsor, and the event social media manager, something along those lines. Then say, Tell me what a meeting on Monday between them would look like if we were talking about X, Y, Z topic.
[00:05:48.200]
We have few shot prompting. We have chain of thoughts, step by step. We have tree of thoughts role play different people. Another one is called chain of density. This is really good for densifying information, particularly if you’re a social media manager that has a lot of text, for instance, for academia. Maybe you work at a university or college or higher education. Maybe you’re working with faith-based churches and organisations, and there’s a lot of text or a lot of information to pass on. Fashion, you can show a jacket, put the price up, and click here to buy. You might need for some audiences and some products to work in a different way. Chain of density is I want you to read this proposal from the government or this news release or something, and then give me a summary in 90 words. Then I want you to give me five more summaries, densifying each summary, adding in another core concept called an entity. It’s how transformers, GPT transformers work. I’ll go and make each of the summaries denser, but keep them readable because they can get pretty unreadable pretty quickly. I usually don’t go with the fifth one, but if I need a nice succinct paragraph for a Facebook post, I might go with the third or fourth one.
[00:07:20.080]
I think one of the challenges with AI is to ensure that it’s not the pilot, but the copilot, meaning it’s your mentor. The best thing to always do for me is to ask more questions, to have it ask me questions. Sometimes I’ll put in something like a negative review that’s on social media and say, You’re the customer. You left this review. I’m going to give you a response. I want you to critique my response. Every time I’ve done that, the AIs come back with, It’s a good response, but I would also include this, this, and this. That helps me get better on the fly with responding in crisis com situations. I could, of course, put the review up and ask the AI to write the response, but sometimes that mentoring part is worth more than simply not knowing whether it’s a good answer or a bad answer.
[00:08:17.550]
By having it mentor you through all the answers, you get better at your job as well. Finally, with my structured prompting, particularly because I’m concerned about catastrophic failures because we’re getting them a lot at the moment. I think with the new version with 125,000 tokens versus 3,000 or 10,000 tokens, we should be okay. We should be better. But catastrophic failure is where it loses the thread of the conversation. It runs out of RAM, it runs out of memory. It doesn’t really run out of memory, but it runs out of tokens. So I ask it to summarise our conversation so far, and it gives me a really nice succinct summary, which if I do have catastrophic failure, I can take to a new thread and start again. You’ll notice, for instance, if you’re doing large volume work like, I need you to write me 10, 400-word articles, and then we’re going to edit them and we’re going to fix them, and you do a whole lot of work with them, then you may want to shift into asking for summaries each time. A more advanced option, two more advanced options. One is to set the temperature.
[00:09:27.860]
I type in on the first prompt at the beginning of the thread, nowhere else. Just the very first one. Ai has a temperature parameter setting. Please set the temperature to 0.9, which is hot. And that will give you a more creative answer. Ai has a temperature parameter setting. Please set the temperature to 0.2. I find that one best for math and semantic AI, search, retrieve, sight. Another one is to put delimiters in. I’ll say bracket comment one in bracket, type the comment in, and then bracket slash comment one bracket, slash, comment 1, bracket. Do that for 50 comments. Now I can further down the thread reference back. I want you to go back to comment number four and give me another response based on this information so that you can continue to progress through your prompts without having to type the whole copy and paste the whole review or the whole comment or the whole tweet or whatever it is in each time. Very last. Once I’ve done these and I might be working through about 14, 15 different prompts in one thread to get to what I want, I never get it to perfect.
[00:10:45.570]
I always just take what I have and massage it a little bit and then keep going. The very last thing is I say, Please review our chat discussion so far and give me tips on better prompting and let me know if there’s anything that I missed that I should have asked you. I find every time I do that, I learn something new. At the beginning, you get a lot of tips because you’re probably only doing two or three prompts in one thread. As time goes on, the prompts are better. That’s how ChatGPT will work with you, mentor you with your social media, give you advice on social media, help you create content on social media, analyse reviews on social media. You could even upload data analytics and have it review your social media engagement and virality and things like that. I hope you found it interesting and I’ll see you in the next video. My name is Laurel Papworth and thank you.
Resources on Structured Prompting
- Zero Shot (no examples, just heuristics) Large Language Models are Zero-Shot Reasoners https://arxiv.org/abs/2205.11916
- True Few-Shot Learning with Language Models https://ar5iv.labs.arxiv.org/html/2105.11447 (Perez et al)
- CoT Chain-of-Thought Prompting Elicits Reasoning in Large Language Models https://arxiv.org/pdf/2201.11903.pdf Google Research Brain Team
- ToT Tree of Thoughts: Deliberate Problem Solving with Large Language Models https://ar5iv.labs.arxiv.org/html/2305.10601
- CoD From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting https://arxiv.org/abs/2309.04269