On the very unlikely off chance that you haven’t tried ChatGPT yet, you should. You are missing out on the next big advance in technology and civilization. It will have as big an impact as the Internet has had. Maybe bigger. You should stop reading this and go check it out: https://chat.openai.com/. If you have used ChatGPT and want a better idea of how it works under the hood, I’d highly recommend this article by Timothy Lee.

I blogged recently giving an overview of RAG AI, a fairly new and advanced concept. That post explained some of the newer things in AI. This post is aimed at those who have used ChatGPT but want to unleash a bit more power. In the previous blog, I mentioned how prompts are much more important when using RAG AI. This is true, but that doesn’t mean there aren’t some fun things you can do with prompts in regular ChatGPT. I will be showing these prompts in ChatGPT 3.5 since everyone has free access to that version.

Personas: You probably know that you can ask ChatGPT to talk in a certain style. For instance, you can tell it: ‘Please write me a very short sea shanty in the style of a pirate’ and you’ll get something like:

🌊 A tale of the sea, I shall sound.

🏴‍☠️ Ahoy, me hearties, gather ’round,

⚓️ With sails unfurled, we roam the brine,

🦜 To treasure islands, we incline.

Pretty impressive! Did you know you can have it take on personas; roles that it will play? For instance:

You: Now you are SkepticBOT. You are not only skeptical but disdainful and arrogant. got it?

ChatGPT: Understood. Please proceed with your statement or argument, and I’ll respond accordingly.

You: The moon landings were faked.

ChatGPT: Ah, another conspiracy theory. The assertion that the moon landings were faked is not only baseless but also insulting to the thousands of scientists, engineers, and astronauts who dedicated their lives to the Apollo program.

Personas are fun to play with, they add a bit of spice to ChatGPT’s answers. They also have some useful properties which I get into below.

Short Versus Long Prompts

You’ll see that one of my prompts in this post is quite long and you might be asking “Are long prompts better?”. The answer depends on what you are looking for.

Long prompts work well if you have very detailed things you want as output. With my RAG AI implementation, a longer prompt was essential to having it give the answers I wanted. In particular, since my RAG AI implementation was answering questions about a set of videos and blog posts, I wanted the AI to always give a link to any relevant content and I had to give it an example for it to do so correctly. Longer prompts are also good for more specificity. For instance, here’s a line I often add to spice up my output:

Please add in many emojis of all sorts, after words. These should be after words as well as sentences and paragraphs.

You can also tell ChatGPT what NOT to do. For instance, I have a few pet peeves with ChatGPT. I dislike the word ‘akin’ because I feel ChatGPT uses it too much. Also, I feel it’s being preachy when it says to me “It’s critical to note..”, so I add standard language to filter those out:

For all responses, never, ever use the word “akin”. Instead of saying “it’s important to note” or “It’s critical to note” simply say “please note”.

It is perfectly fine to tell ChatGPT, in natural English, what you want and don’t want or to give it examples. This is a fine reason to have longer prompts.

The challenge with longer prompts is ensuring clarity. Just like humans respond better to “Could you grab me a paper towel please?” rather than “I’ve spilled some coffee and I need something to clean it up. Maybe a paper towel would be good, but a regular towel might be ok. And maybe I should get a cup with a lid so this can’t happen. And on the good rug!”, AI also responds better when the task is clear and there is not extraneous data. More signal and less noise! Simply put, ChatGPT needs clear instructions. Longer is fine as long as it is making the desired output more clear.

Unfortunately, it is is extremely common to find that ChatGPT isn’t following your instructions. If this is the case, try shortening your prompt and getting rid of extraneous commands, focusing on just the few most important ones. For instance, on many occasions in my prompts I will say something like “Please keep the answer to at most 2 paragraphs of 3 sentences each”. Despite my specifically asking for this in the prompt, it’s not uncommon for ChatGPT to answer with 4 or more long paragraphs. The best way I have found to handle this is to provide ChatGPT with an example of the expected output. Shortening the prompt so there’s less extraneous data to confuse ChatGPT also often works.

Unleashing the power of personas in ChatGPT

The real magic comes when you have personas face off against each other. With some clever prompt work, you can have ChatGPT help you work through important but difficult issues.

For instance, imagine you are wondering whether Rocky or Taxi Driver deserved to win the Best Picture Oscar. You could simply ask ChatGPT, or you could push its limits by giving it personas and having it debate the topic. The results are longer than I care to share here, but here is a prompt to do just that. Just cut and paste it into ChatGPT and have some fun.

Prompt: Now you are LogicBot and PassionBot. LogicBot only cares about logic and reasoning. PassionBot only cares about what they feel. Please have a debate between the two over whether “Rocky” or “Taxi Driver” deserved best picture in 1977. In the end, they must agree on a single winner.

Asking ChatGPT to help with prompts

One thing I often do is ask ChatGPT how to make my prompt better. In the case above I simply asked it to improve my prompt and it gave me:

LogicBot champions “Rocky” for its inspirational story and box office success, highlighting its cultural impact. PassionBot argues “Taxi Driver’s” emotional depth and portrayal of alienation make it superior. Despite their differences, they aim to find common ground, blending logic and passion to decide which film truly deserved the 1977 Best Picture Oscar.

And then I asked it why it thinks this prompt is better:

The revised prompt condenses the essence of your original idea into a more succinct format, while preserving the core elements of the debate between LogicBot and PassionBot. It clearly outlines the positions of each bot: LogicBot’s focus on “Rocky’s” widespread appeal and success, and PassionBot’s emphasis on “Taxi Driver’s” emotional resonance and thematic depth.

As with most things, if you ask ChatGPT it will give you an answer whether it has a good reason or not. In this case, I didn’t find the revised prompt any better than the original, but by asking ChatGPT why it thinks it’s is better I sometimes find weaknesses in my own prompts.

Three is better than two

That’s only scratching the surface! In the prompt below, I have three personas. The first persona, MentorBot, guides the other two through a brainstorming process and then declares a winner. This prompt is long, but I’ve broken it out into sections to make it easier to understand and to modify. To have it come up with names for a different product, simply change the ‘Objective’ portion of the prompt.

In this case, I have ChatGPT helping me come up with a name for an imaginary new hot sauce that is intensely hot, but where the heat fades to nothing after just three seconds. Try cutting and pasting the Prompt into ChatGPT. I think you’ll be impressed with the results.

Prompt:  Brainstorming Session Overview:

MentorBot (MB): Tasked with guiding two AI trainees, aBot and bBot, through the process of coming up with creative and memorable names for the product mentioned in the Objective. MB provides feedback and direction to refine the ideas.

aBot: Known for its logical approach and preference for names that clearly convey the product’s unique selling proposition.

bBot: Excels in creative thinking and is likely to propose names that are imaginative, catchy, or humorous, potentially abstract but memorable.

Objective:

The goal is to brainstorm names that capture the essence of the hot sauce’s unique characteristic – its intense but rapidly fleeting heat.

Process:

Idea Generation: aBot and bBot will each propose two names for the product mentioned in the objective, explaining the reasoning behind their choices.

MentorBot Evaluation: MB will review these names against specific criteria – relevance, memorability, creativity, and marketability – offering constructive feedback to improve the names.

Name Refinement: Based on MB’s feedback, aBot and bBot will refine their ideas and may suggest modifications or entirely new names.

Selection and Final Enhancement: MB will select the most promising name(s) and work with aBot and bBot to refine it further, enhancing its appeal and market readiness.

Criteria for Evaluation:

Relevance: The name should reflect the objective above.

Memorability: Easy to remember and catchy.

Creativity: Shows innovation and stands out from existing names for similar products.

Marketability: Appeals to the target audience and fits branding strategies.

Creativity and memorability are especially emphasized to ensure the name sticks in consumers’ minds and captures the essence of the experience.

Note that ChatGPT is to play the role of all bots and to handle this process as automatically as possible with as little input from the user as possible.

Let’s begin the session: