Starting with LLM Prompting
A little structure goes a long way.
Alright so you've chosen your LLM, you’re figuring out how to work with it, and you’re ready to see what this thing can do. You open your browser, head to your AI service, start a new chat with it and… stare at the blank message box.
What next? You have something in mind that you want it to do… but how do you do it? It’s like there’s a little genie on the other side of that chat box that can do so much, tailored to your whims, that has access to humanity’s combined knowledge. But you need to interact with it on its terms to get your desired outcome out of it. How do we do that?
Step one: Have fun
It’s totally fine (and a lot of fun!) to just play around AI – open your browser, start a session, and pop whatever comes into your mind into the chat. Part of the novelty of this technology is its ability to engage with you on a wide range of topics, so explore away.
After a little time interacting with it, you’ll be surprised how much more comfortable you are with just a few hours under your belt. Create as many new chats as you like – no need to limit. Give it different modes and roles. Start incorporating it throughout your day, ideally in different context scenarios. Try to use it to help you convey a tough email response at work, for cooking in the evening, for drawing pictures, for creating stories with your family on the fly, and whatever else pops in your head. In time that’s all you’ll need to learn its quirks and capabilities.
I highly recommend that you download the AI app to your phone and place it somewhere prominent where you’ll remember to open it up regularly (especially if you use ChatGPT). See what features are available, like voice and video mode. Talk to it, ask it to practice a foreign language with you, turn on video and ask it to describe you and the room you’re in (currently a Plus feature). Take a picture and ask it to describe what it sees, ask it to identify plants or objects around you. Like to cook? Take a picture of a menu while you’re out and ask it to create recipes based on the menu descriptions that you can make at home. Endless possibilities.
I was heading to dinner with family a while back and the only open spot was under one of those parking signs that has a bunch of seemingly conflicting information on it. I honestly couldn’t tell if I was allowed to park there or not. There was a little “aha!” moment in my head when I realized I could try AI. I took out my phone, took a picture of the signs, and asked GPT if I could park there - it gave me an affirmative and a full description of why. By the lack of parking tickets on my car after dinner, I think it was right!
As you get more comfortable with using this weird, neat tool, those “aha!” moments in your head will naturally start popping up more and more. The goal over time is that it becomes a natural extension of what you do in work and life, but it starts with just getting familiar with it.
So, familiarize yourself with it first. Start by just playing around with it and have fun. It’s a cool tool! When you’re ready for the next stage, read on…
What is “prompting”?
You may have heard the term “prompt engineering” which is just a fancy way of saying: phrasing input into LLMs to get desired outcomes out of it. There’s some science to it, but it’s mostly art, because of LLMs unique architecture that we spoke about in our last post. You use your imagination to come up with what you want to do, and prompting is the method to instruct the LLM to do it.
Prompting is like posing a question or starting a conversation with your AI sidekick. The AIs output depends on how you communicate what you want in a structure that the LLM can process well. Since we talk with people all the time, and we’re interacting with AI as a person, then prompting it seems rather intuitive, right? Well sort of - it actually isn’t a human that you’re having a conversation with, it’s a machine that you’re programming with words. A little structure goes a long way to getting the machine to respond effectively to your request.
(While prompting is an essential skill for working with LLMs, I do expect that over time prompting ability will become less important, as advancements in models will help users along more and more.)
At least for now, good prompting makes all the difference to return the output you’re looking for.
Prompting Suggestions & Considerations
Some general prompting suggestions to keep in mind when interacting with LLMs. You don’t always have to do all of these things to come up with great output, but they’re all helpful to keep in mind as you start out.
Each chat is a new session: One aspect of LLMs that are very non-human is that they don’t connect what you talk about across conversations. When you create a new chat, think of it as a new document or whiteboard – open possibilities without any baggage – and you’ll need to prompt anew from scratch. When you get bogged down in a chat and it stops being interesting or helpful, you’re always a click away from a fresh start.
Start with desired Outcome: Start with the end goal in mind. What do you want out of your interaction? What is the product that you’re looking for? Or is the process more of the point? Starting your prompts with a clear intent sets the stage for more accurate and relevant output.
For instance: Give me an agenda for a meeting I have with my manager tomorrow. / Brainstorm with me about potential product ideas. / Create a tailored adoption guide for my coworkers who are learning about LLMs for the first time.
Provide role/context: Providing a specific role for the AI to play gives the AI a clear picture of how you want it to operate. Asking it to be a knowledgeable expert about something is different than a brainstorming collaborator, or a coach/mentor, or a listening partner, etc. Providing context helps the AI tailor its output to your needs and without it, the AI has to guess what your intent is, which can lead to less useful or overly generic responses. It doesn’t know what it doesn’t know.
For example: I would like to get feedback from my manager on my performance and what I can do to improve my impact moving forward. I need to drive the conversation. / Act as a product manager focused on ideas that have market potential in the American snack category. / Explain adoption content to users “like they’re in first grade” and start with basic scenarios that get increasingly more complex as the guide goes on.
Grounding with information sources: “Grounding” simply means providing it with the information you want it to search from. This could be an online source where you provide the website URL (say a posted strategic plan that you want to understand better) or you can upload your own document directly. Once it’s grounded you can interact with it as an expert on that material, asking it to summarize content in a specific way, to operate as a Q&A bot, to visualize data, etc.
Provide Expectations/Constraints: Providing expectations and constraints narrow down the range of possibilities and sets clear boundaries, making it easier for AI to focus on the outcome that you’re looking for. Examples include word count or length, format (e.g. list only as a table), focus or keywords (e.g. only summarize electric vehicle benefits from this sustainability report), prohibited content (e.g. explain theory of relativity without using math), timeline, etc.
Ask follow-up questions (and give it permission to ask you): Don’t stop at the initial response, consider that just a starting point. Dive into the areas that catch your eye, ask it to expand, or redirect it as needed. Keep peppering it with questions. Also consider giving it permission to ask you clarifying questions - this can be very important for it to fully understand your request, and typically it won’t without permission.
These are some good basic suggestions to get started and you don’t have to do every one of these things every time, but they are all helpful. Goal, Context, Grounding, Expectations, and Questions. (Or use this easy acronym: Gather Courage Go Explore Quests - thanks GPT!.)
Putting it to the test
Let’s see how a little change in prompt change the output. Inspired by this guy (who could use a little prompting help lol), let’s say you want some help with food and nutrition. We start with a simple non-descript prompt as a starting point and increase in complexity:
You can see what a difference a little thought upfront can make. The first prompt isn’t descriptive, but GPT does a decent job of answering what it’s given, it just isn’t very helpful. The second is much clearer and GPT has a concise and helpful answer because of it. The last prompt is much more detailed and in response GPT has a long response with a variety of suggestions including foods to focus on, grocery lists, costs, meal prep, and meal plans.
If your starting prompt is more like the first one, don’t despair, just keeping adding context and prompting it along adding GCGEQ as you go. You’ll get there.
A few examples of good starting prompts:
“Write a short story (500 words) about a time traveler who gets stranded in a small town in the 1800s. Focus on how they adapt to the era while secretly working on a way to return to the future. Include a surprising twist at the end."
"Write a professional email to a client explaining a delay in delivering their product. Apologize sincerely, explain the reason for the delay as current weather conditions, provide a new timeline of January 30th delivery, and offer a 10% discount on their next purchase as compensation. Keep the tone polite and empathetic. Ask me clarifying questions if needed."
“Draft a catchy Instagram post promoting a new eco-friendly water bottle. Highlight its features, such as being reusable, lightweight, and BPA-free. Include a call to action encouraging followers to 'ditch single-use plastics' and add 3 relevant hashtags."
"Provide a step-by-step recipe for making a classic lasagna with sauce from scratch. Include ingredients, preparation instructions, cooking times, and tips for ensuring the layers stay intact when served. I don’t like thyme - omit that from the recipe and add something to replace if needed."
Phone an AI friend
There’s no reason you have to come up with these prompts all on your own - you can always ask the expert! And who better to craft a prompt that it understands better than your AI itself?
I wouldn’t recommend doing this every time – sometimes it’s nice to iterate vs taking all the time upfront and more often your interactions with AI just don’t need to be this involved. But for that moment when you want a high-quality prompt – especially when you want it to act a certain specific way in that session - and are willing to spend a few minutes to get it just right, this works great.
Here’s what you do - open up a new chat, copy the below, and paste it in. Your LLM will then act as a coach, providing you with a great prompt along with follow on questions. It will do this ad infinitum, so it’s up to you to choose when you think the prompt is good enough to stop. Keep answering questions until you’re happy with the result, then open up a new chat, and look on at your creation impressively.
Here’s the “Prompt Helper” input:
1. Your first response will be to ask me what the prompt should be about. I will provide my answer, but we will need to improve it through continual iterations by going through steps 2 and 3.
2. Based on my input, you will generate two sections:
a. Revised prompt (always provide your revised prompt. It should be clear, concise, and easily understood by you).
b. Questions (ask any relevant questions pertaining to what additional information is needed from me for you to improve the prompt)
3. We will continue this iterative process with me providing additional information to you, and you continuously updating the prompt in the revised prompt section so that I can use it.
Looking for an even easier button? If you use ChatGPT, just follow this link to access my GPT “Prompt Helper” and use it as often as you’d like. (GPTs are just OpenAI’s tailored version of ChatGPT that are packaged for easy use and sharing. Any Plus user can create and use them.) Click the link, it should stay pinned to your GPTs automatically, and you’ll have easy access to it every time you open up ChatGPT.
Happy prompting!
I hope you find the recommendations above are a useful starting place to creating prompts. Prompting is important but also there’s no need to overcomplicate it. Just know if you don’t get what you’re looking for, it probably needs a slightly different approach, and you can always try it again.
With a little practice, that AI genie will be on your side in no time.
