In our discussions with many different people, how to get started has come up quite a bit. Many of these people aren’t working in technology and haven’t programmed before. When they try to use AI it doesn’t work for them and the results aren’t all that helpful. This gave us the idea of putting together a short guide for anyone who identifies with these problems and would like some helpful pointers to get started.Getting SetupI think the easiest place to start is by getting an account setup to experiment with Ch...
In our discussions with many different people, how to get started has come up quite a bit. Many of these people aren’t working in technology and haven’t programmed before. When they try to use AI it doesn’t work for them and the results aren’t all that helpful. This gave us the idea of putting together a short guide for anyone who identifies with these problems and would like some helpful pointers to get started.
Getting Setup
I think the easiest place to start is by getting an account setup to experiment with ChatGPT. You can go on their site, create an account, and get ready in just a few minutes. They also have a mobile app that you can download which may be easier if you don’t usually use a computer or laptop.
These same ideas will carry through to other systems like Midjourney, DALL·E, or Bard. We just find ChatGPT is one of the easier interfaces to use when getting started.
For our examples we’re going to get ChatGPT to write a letter to our friend that updates them on the latest news in our life and breaks the bad news to them that we won’t be able to visit this summer in as kindly as possible. I highly recommend you follow along on ChatGPT.
Don’t Look for A Single Right Answer
The most common challenge we find people struggle with is looking for the “right answer” to using AI. If you did well in school or took standardized tests you might be geared toward finding one correct answer to how you should use AI. Unfortunately there isn’t just one correct answer. Many different questions can get you the same result. It’s similar to asking your friend a question, they’ll likely give you the same response even if you ask your friend a question in a few different ways.
Instead, using AI is a lot like trying to ask a question that the AI can understand best to answer your question. If you ask someone “What is the capital?”, they might not know what you’re talking about. The AI is the same, you’ll need to be a bit more specific.
The examples we give below are some good ways to get started but you might find that for what you’re trying to do there’s another better way that works for you. Don’t get too attached and try to have fun finding new ways to get what you want back from the AI.
Initial Prompt
Prompts are the questions or queries that we send to a service like ChatGPT. For our initial prompt, we’re going to go ahead and get started writing a letter to our friend using the following:
“Write a letter to my friend, Paul, telling him I got a promotion to Manager at my company, my dog died, and I won’t be able to visit them this summer
If you put this into ChatGPT you’ll notice that it does a great job getting a letter created but it likely isn’t at all what you imagined. It probably doesn’t sound like you, it likely includes details that we didn’t give, and it might not be very interesting to read.
You might be tempted to then just say:
“Make it better”
The result is a letter that likely uses a more complex vocabulary, might have fixes some grammatical issues, and might be a little longer but doesn’t seem to address any of our issues. It’s at this point most people give up and call it a day. But if we push just a little further I think ChatGPT will surprise you!
You can find my chat example here
Adding Specifics
So we can immediately solve some of the problems we had just by providing a few more details. Let’s start a new chat and add some details that can be used to fix some of the challenges we’re having:
“Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.”
Testing this out we’ll immediately see much better results. Describing the tone of letter (casual) helps a lot and can definitely help ChatGPT know what kind of language I want to use. The other specifics help to fill the letter with details that make the letter’s contents a lot closer to what we were hoping for.
You can find this chat example here
If you imagine talking to your friend and asking them to write a letter for you, these are the same details you’d need to provide them. It helps a bit to think of an AI as just an assistant that is trying to figure out what you want and can only use what you give it to help you.
Adding specifics to your prompt might be just enough to get what you’re looking for especially if the letter isn’t very personal. However, this letter still doesn’t have the personal touch you might like it to have. Let’s look at some other ideas for how to make our examples just a little better.
Adding an Example
If we were to hire someone to write letters for us, something we might do is provide them with some good examples of how to do it. That would help them understand what we’re looking for and when we gave them feedback they could continue to learn with those examples as a foundation. For this improvement, we’re going to give the AI the same thing: an example of a letter for them to use.
Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.
Use the following example as a reference for my writing style:
Hey!
It’s been a crazy and amazing week. I’m at the Seville airport, getting ready to board flight TP1105 to Lisboa before TP259 to Toronto. It’s a really small airport here and security took me 5 minutes!
My favourite memory is this dinner of the amazing view of Ronda with local wine, salmon tartar, croquettes, artichokes and Iberian ham. Amazing views, delicious pairings, and incredible company.
Looking forward to being home now. 9.5 hours in the air and I’m there.
Love you,
Ben
Adding this example, we see the style of the result change to be more inline with the way that we talk and we see our formatting style in the resulting letter. However, you still might not be too happy with the result. Don’t worry, there are a couple other ideas we can try out!
You can see the result of adding an example here.
There are a number of other ways to provide an example that might provide even better results in your testing. One prompt engineering technique is called a **One-Shot Prompt** where we give it an example of a prompt and a resulting letter. You can see an example of this here.
Adding Rules
One thing AI are great at is following rules that we provide. They can factor in a large number of factors that are hard for us to track. Because of this, rules are a great way for us to control their behavior and get better results. Let’s add some rules that the AI should follow when they help us write our letter.
Write a casual letter to my friend, Paul. Tell him I got a promotion to Manager at my company, Nintendi, because of my hard work over the last year. My dog, Mari, died from old age and my family is still grieving. I won’t be able to visit him this summer due to the new responsibilities at work.
Follow these rules when writing the letter:
* Don’t be too emotive and stay somewhat reserved
* Try to start and end with good news
* Don’t add any details I didn’t include
* End with “wishing you the best”
Use the following example as a reference for my writing style:
Hey!
It’s been a crazy and amazing week. I’m at the Seville airport, getting ready to board flight TP1105 to Lisboa before TP259 to Toronto. It’s a really small airport here and security took me 5 minutes!
My favourite memory is this dinner of the amazing view of Ronda with local wine, salmon tartar, croquettes, artichokes and Iberian ham. Amazing views, delicious pairings, and incredible company.
Looking forward to being home now. 9.5 hours in the air and I’m there.
Love you,
Ben
These rules help to reinforce our tone, style, and help keep the AI from adding details that we don’t want added to our letter. Rules are easy to add and remove to see different results in the letter’s style and offer great direction for the AI. Sometimes you can jump right to adding rules for a great result, but it very much depends on the task you’re trying to do.
While it may not be perfect, this really gets us quite close to a letter that we’d be happy to send. After a few small tweaks and edits you have something that you can send to Paul and a great template you might use for future letters!
You can see the final result here!
Turning it off and on again
Similar to many technologies, AI sometimes get a little bit overwhelmed and need to be restarted. Consider opening a new chat conversation with ChatGPT after a while.
You can keep a conversation going for a long time in ChatGPT but you will start to lose control of the conversation and the AI might start behaving strangely. This is especially true If you’re trying to do a lot of different things in a single chat. Previous responses may leak into other questions. For example if I’m writing another letter to a friend in the same chat that we were chatting with Paul, it may bring up my promotion even through I didn’t say that it should just because the AI knew about it from earlier in the conversation.
Use a more powerful model
Another great way to get better results is to use a more advanced version of the AI. If you’re struggling to get GPT-3.5 to return something, using GPT-4 could be all you need for the AI to understand what you want. Later models are a little slower but they’re smarter and able to better understand complex requests. The GPT-4 announcement gives some great examples for how much the model improved over time.
Takeaways
Prompting can be a frustrating experience as we try to get an AI to understand just what we want. When struggling, refer back to these improvements that can help AI better understand what to do and how to produce something a little closer to what you’re looking for. AI act a lot like assistants and we need to provide them with the details that help them to do the jobs that we give them. Without that feedback, they’ll continue to struggle to understand how to make their results any better.
Thanks for reading Stable Discussion! Subscribe for free to receive new posts and support our work.
Get full access to Stable Discussion at blog.stablediscussion.com/subscribe
