Transcription
I took Google's Prompt Engineering course for you, so here's the cliffnotes version to save you to 9 hours. But it's not enough just to listen to me talk about stuff, so I've also included a little assessment at the end of the video to help you remember everything that you learned. Because research shows that immediately reviewing information after you learn it is the best way of retaining that information.
All right, let's go. Let's first go over the structure of this course. Prompting Essentials has four modules. Module one is "Start Writing Prompts Like a Pro." This is where they introduce some really helpful frameworks for how to craft prompts. Module two is "Design Prompts for Everyday Work Tasks." This will include prompts for emailing, brainstorming, building tables, and summarizing documents. Module 3 specifically focused on using AI for data analysis and for building PowerPoint presentations. And finally, module four, "Use AI as a Creative or Expert Partner." This is where Google really packs it in. Like, I am genuinely super impressed by this module. We talk about advanced prompting techniques like prompt chaining, Chain of Thought, Tree of Thought, and a framework for creating agents.
All right, module one. Let's do the fundamentals. Let's first, the prompting. Prompting is the process of providing specific instructions through a Gen tool to receive new information or to achieve a desired outcome on a task. This could be text, images, video, sound, or even code. The course provides a five-step framework for how to design a prompt: Task, Context, References, Evaluate, and Iterate.
The task is what you want the AI to do. For example, if your friend's birthday is coming up and they're really into anime, you can say, "Suggest a gift related to anime for my friend's birthday." Now, that prompt in itself is okay, but you can elevate this and get a result that's more unique and specific by incorporating two additional things. The first one is a persona, which is a role that you want the AI to embody. For example, you can update the prompt to, "Act as an anime expert to suggest an anime gift for my friend's birthday." You notice that the results are a lot more specific, and it's actually split into different genres.
The second thing you can add is the format of the output. The default here is just the list and bullet points, but maybe you want something that's more structured. So you can say, "Organize that data into a table."
The second component of the framework is context. The general rule of thumb is that the more context you can provide, the better the output will be. In your birthday gift example, you can specify something like, "Your friend is turning 29 years old. Her favorite animes are Shangri-La Frontier, Solo Leveling, and Naruto, etc., etc." You'll see that the output is much more targeted.
Third part of the framework is references. This is where you can provide examples to the AI. Sometimes when you're trying to explain what you want, it's kind of hard to describe it in words, but providing examples can really clear things up. And AI is especially good at incorporating examples. Maybe you can provide past birthday presents that this person has enjoyed.
Step number four is evaluate. This is after you get the output. Just ask yourself, "Is this output what I want it to be?" And if it's not exactly what you wanted to be, then the last step is iterate. Prompting is rarely a one-and-done kind of thing. It's much more of a circular process in which you're refining the prompt to get the results that you want, just like what we did earlier. Often times, you might just start with a simple task like suggestions for a birthday present. Then you want to get better results, and you start iterating on that and adding things like a persona, context, and references to finally get to a result that you're happy with. As the course calls it, ABI: Always Be Iterating.
Speaking of mnemonics, the course does have one for this five-step framework, which I actually find really difficult to remember. Um, I think it's "Thoughtfully Craft Really Excellent." I don't know what the "I" stands for. I'll put it on screen, but I do have one that I made, which I can remember a lot better. So I don't know, maybe this will help you as well. Uh, which is "Tiny crabs ride enormous iguanas." A lot more memorable, in my opinion. Anyways, whatever it is that you need to do, uh, just figure out some way to remember this framework because everything else in the course is based on this.
The rest of module one, which also includes interviewing different people, I think is interesting but not super necessary. The only other really useful, important thing that they presented is the four iteration methods. So by following date, I'm just going to call it the "Tiny crabs ride enormous iguanas" framework, um, will get you like 80% of the way. But sometimes you're just not quite there. So to iterate and get the 20% there are four different methods that you can try.
The first one is just to revisit the prompting framework. Maybe you can give more references, more examples, provide more context, or, you know, add a persona if you haven't already.
Number two is to separate your prompt into shorter sentences. It's helpful to think about AI like how you would talk to a normal person. Like, if you just word vomit to someone about whatever it is that you want, they'll probably be like, "Overwhelmed," and there's just like a lot of stuff going on, right? So the same thing can happen for AI. And an easy solution for this is just to break your prompt into simpler sentences and feeding it to the AI slowly. So it's less like, "Blah blah blah blah blah blah blah," I'm more like, "Blah blah," and "Blah." Much more organized.
Number three is trying different phrasing or switching to an analogous task. Say you want the AI to help you write a marketing plan, but the results are just kind of like boring and bland. What you can do is that marketing is really just telling a compelling story. So instead, you can ask it to, "Write a story about how this product fits into the lives of our target customer demographics." This is an analog task, but the results are much more lively and interesting.
And the fourth iteration method is to introduce constraints. Just like when you tell someone that they can do anything, or like if you ask people, "What does everybody want to eat for lunch?" and they're just like, "Oh, anything." This actually makes it a lot harder for you to get a result that you're happy with. So instead, you can introduce constraints to narrow the focus down. Say you want to generate a playlist for a road trip, and the AI generates your playlist, but it's just like, not very interesting. You can add different constraints like, "Only specific to a certain region," "Only at this specific tempo," or "Only songs about heartbreak." Deion, for example. I don't know, maybe you like feeling sad.
So with these four iteration methods, um, with the help of AI, I also came up with a mnemonic to remember it better, which is "Rahen saves tragic idiots."
So let's talk about multimodal prompting. The most classical way of interacting with a large language model is just by, like, typing stuff. I like having a conversation, but you can actually interact with many AI models, like Gemini, um, with different modalities as well, including pictures, audio, video, and even code. It's able to take different types of modalities as the input and is able to output using different modalities as well. This doesn't change anything in terms of how you think about prompts. It's still going to be "Tiny crabs ride enormous iguanas," but you just might need to be a little bit more careful about specifying what kind of input or output you're looking for and the kind of context that you're providing.
For example, if you designed a new nail art collection and you want to market it on social media, you could input something like, "Write a social media post featuring this image," and then attach your nail art collection as a reference. "The post should be fun, short, and focus on the fact that it's a collection of new designs I'm selling."
Some other examples of multimodality usage would be asking a Gen tool to suggest recipes based on the photo of the ingredients in your fridge, inputting your brand's logos and colors, and then creating a digital teaser to promote an event, or if you're working on a short story and you get really inspired by a musical piece, you try inputting that music piece and tell it to kind of follow those vibes for the atmosphere and details of the story.
Regardless of the modality that you're prompting in, there are two major issues with using AI tools. The first one is hallucinations. A hallucination is when a Gen AI tool provides outputs that are inconsistent, incorrect, or even nonsensical. A really famous example is that if you ask an LLM how many "R"s are in "strawberry," it tells you that there are two "R"s in "strawberry." The second is biases. Unfortunately, LLMs, being trained on human content, also incorporate human biases, things like gender and race.
So to minimize these sorts of problems, the course recommends that we take a human-in-the-loop approach, which is making sure that you're always checking your outputs and verifying whatever it is that the Gen tool gives you. In the end, it is your responsibility of making sure that whatever is being produced is in fact accurate.
Here is a checklist. Feel free just to take a screenshot for some considerations when you're thinking about using AI responsibly. You know, compared to other Google courses I've taken, especially the AI Essentials course, which you can check out over here, um, this course is a lot more dense, which is a lot better bang for your buck. So pay attention.
Moving on to module two. Module two is called "Design Prompts for Everyday Work Tasks." It's essentially just providing examples of use cases based on the "Tiny crabs ride enormous iguanas" framework and the "Rahen saves tragic idiots" framework too. That's why I'm going to go through this module relatively quickly. I'll highlight some of the examples that I think are really important, and for the rest of them, I'll actually just put on screen so you can take screenshots of it if you want and build out your own prompt library where you can kind of like store the prompts that you want to use.
One of the biggest use cases that most people have when using Gen tools is by using it to produce content, for example, like writing emails. Here's an example of a situation, um, when you want to write an email to your staff about a new schedule change for your gym. "I'm a gym manager and we have a new gym schedule. Write an email informing our staff of the new schedule. Highlight the fact that the MWF (Monday, Wednesday, Friday) Cardio Blast class changed from 7:00 a.m. to 6:00 a.m. Make the email professional and friendly and short so that the reader can skim it quickly. Here's the new schedule," and you can actually attach the link that contains the new schedule.
This sort of email probably takes you like 10 minutes to write, but by using a Gen tool, you can do it in like a minute. Most of us do send quite a lot of little emails here and there throughout the week, um, so the time savings do add up. For this kind of email, you probably aren't super picky, but what happens if you need to write an email that is a lot more important, or you're writing other things like an essay, an article, or a newsletter? You would care a lot more about the tone and the word choice that's being used. Instead of using general terms like "write a casual summary," try to use more specific phrases like, "Write a summary in a friendly, easy-to-understand tone, like explaining to a curious friend." You can also provide references for context, other emails, or articles, or whatever that you've written in the past, and tell the AI to match the tone.
I'm going to now include a few other prompts on screen related to generating text or content, which you can take a screenshot to add to your prompt library if you want.
If you're currently in university, thinking about going to university, or going back to university to maybe get an additional degree, you should check out Straighterline. Straighterline is a credible online education platform where you can take high-quality online courses designed by academics from leading universities and recommended by accredited educational institutions. You can basically take a course at a fraction of the cost and still transfer the credit to your college or university. More than 3,000 colleges and universities have accepted credit from Straighterline courses. So not only are Straighterline courses much more affordable, they also come with the flexibility of being able to take a class when you want and at your own pace. They even have a variety of support services like tutors and resources to help you succeed. If you're still on the fence, just check out their course catalog and you can start a free trial today at this link over here, also linked in description. Thank you, Straighterline, for sponsoring this portion of the video. Now back to the video.
Module 3 is pretty similar to module 2. It's just more example use cases, very helpful example use cases though, specifically data analysis and presentations. The big word of caution here is that be careful about what data that you're inputting into your AI model. Uh, if you're working for, you know, a company, you do not want to be violating any privacy policies, and you probably don't want to be putting some sensitive data into some large language model as well.
The example from the course: If you have a data set for a grocery store chain with information about the store, the area, items that they have available, the daily customer count, and store sales. If you're not very good at Google Sheets or Excel, you might input a prompt like this: "Attached is a Google Sheet of store data. How can I create a new column in Sheets that calculates the average sales per customer for each store?" It can help teach you specific things like this, and it can actually do more. Say, if you're interested in the trends in the data set, you could then add on, "Give me insights into the relationship between daily customer count, items available, and sales based on the given data." Gemini is able to show some interesting trends, including the fact that there's no clear correlation between items available and store sales. If you find this interesting, you can continue prompting it, digging into this, and maybe coming up with ways to figure out why that's the case.
I'm going to put on screen now a couple other prompts related to spreadsheets and data analysis that you might find helpful. The second part of the module is about building presentations, and I'm going to put on screen a couple of prompts related to presentations that could be helpful.
And finally, we are at module 4. We're almost done, guys. Module 4 is titled "Use AI as a Creative or Expert Partner." This is an extremely important module and what made me very impressed about this course. So first, we're going to cover some advanced prompting techniques, starting off with prompt chaining.
Prompt chaining is a way to guide Gen tools through a series of interconnected prompts, adding new layers of complexity along the way. For example, you're an author and you wrote a wonderful novel, and now you want to market and sell your novel, so you need to come up with a marketing plan. The course recommends you use Google AI Studio for this because it has a much longer context window because you're going to be attaching your entire manuscript. The first thing you might want to do is to generate some summaries of your manuscript: "Generate three options for a one-sentence summary of this novel manuscript. The summary should be similar in voice and tone to the manuscript, but more catchy and engaging." So Gemini was able to give some decent options, but you want to focus on a more specific theme. That's where the prompt chaining comes in. Taking the output from the previous prompt and then asking, "Create a tagline that is a combination of the previous three options with a special focus on the exciting plot twist and mystery of the book. Find the catchiest and most impactful combination. The tagline should be concise and leave the reader hooked and wanting to read more." And great, it comes up with, "The desert whisper secret: A young weaver seeks a city of singing stones, but the greatest journey unfolds into whispers of her own heart." Anyways, you can keep refining things if you want, and finally, maybe even asking Gemini to generate a six-week promotional plan for the book tour, including the locations and the channels to promote each stop on the tour.
So that was prompt chaining. There's two other advanced techniques in this module: Chain of Thought prompting and Tree of Thought prompting. As a note, a lot of these AI terminologies and techniques sound super fancy, but they're actually not. Like, Chain of Thought prompting is about asking the AI to explain its reasoning as a step-by-step process. It's similar to how your math teacher might ask you to explain your work so he or she is able to identify like the steps that you're taking and where you could be going wrong. All you have to do throughout your prompting sequence is to tag on the line, "Explain your thought process." This helps you understand the AI's reasoning for things, and you can help improve its decision-making.
Tree of Thought prompting, as its name suggests, is sort of like a tree. It allows you to explore multiple reasoning paths where, like, branches simultaneously. This is really helpful for abstract or complex problems, like developing novel plots with new characters or creating outlines and drafting sections for lengthy documents. You can work with the AI tool to explore different options and evaluate them to finally come up with the best output.
As an example, maybe you're creating an online course and you want to create a cool image on the landing page. You can use Tree of Thought prompting to brainstorm different options. A potential prompt may be: "Imagine three different designers are pitching their designs to me. All designers will write down one step of their thinking, then share it with the group. Then all experts will go on to the next step, etc. If any expert realizes they're wrong at any point, then they leave. The question is: Generate an image that's visually energetic and features images of art supplies and computers. Show me three suggestions in very different styles, from simple to detailed and complex." And here's the output Gemini came up with.
Now, looking at this output, you might be like, "Hmm, I kind of like the vibes of one where one could be going." So you can tell the AI that you like the first one and you'd like to expand the idea a little bit more and perhaps generate three different color schemes for that concept. And you can just keep doing that until you end up with something that you like.
A pro tip is that you can combine Chain of Thought and Tree of Thought prompting by asking the AI to explain its reasoning at each iteration, so you can provide feedback. Another pro tip, as you're prompting along, is if you ever get stuck and you don't really know what prompt to use, you can actually use AI to help you come up with a prompt. This is called meta-prompting.
All right, the last section of the course is on agents, and I have actually not seen a single course be able to cover agents as well as this one. So first, definitions. What is an AI agent? An AI agent is like an expert designed to help with tasks and answer questions. You can have all sorts of agents. You can have coding agents that help you with coding, marketing agents that come up with marketing plans with you, a golf agent that can correct you on your golf swings, or maybe just a friend agent that can be your friend.
The course covers two types of agents. The first one is a simulation agent called AgentSim. AgentSim can simulate scenarios with you, like conduct interviews or do role-playing. For example, if you work in an HR department, you might be tasked with coming up with a training program to help interns improve their interviewing skills for that final job assessment. For AI agents, you want to focus a lot on the persona and the context. The persona here is, "Act as a career development training simulator." The task is, "Your task is to help interns master interview skills and conduct conversations with potential managers." Then you have the context: "You need to support the following types of conversations: articulating strengths and skills, communicating professionally and confidently, and discussing future career development goals." Once an intern has picked a conversation topic, provide details about the situation and the interviewer's role, then act as the interviewer and allow the intern to participate as the employee. Make sure to guide the conversation in a way that will allow the intern to exercise their interview skills. Finally, you want to include a stop role where you can tell the agent that you're done with the simulation. "Continue to roleplay until the intern replies with jazz hands." After the intern gives the stop role, "Jazz hands," provide them with key takeaways from the simulation and skills they can work on.
Now that is set up, you can start doing a simulation, maybe by inputting the chart analysis that I did for my intern project. AgentSim will ask you more questions about the analysis, and you keep responding to them, and at the end, you can insert "jazz hands," and then AgentSim will provide feedback for you.
The second kind of agent is an agent for expert feedback called AgentX. AgentX is able to give you feedback on any topic of your choosing, sort of like a personalized tutor or consultant. Here's an example prompt to create an AgentX that can provide you feedback about a pitch for a potential client. First, the persona: "You're my potential client, the VP of advertising at a world-famous sports car company known for its innovation, performance, and engineering excellence." Now, the context: "You're considering hiring a creative agency to develop a new campaign that will attract a younger generation of buyers. You're in a meeting with me, the design director of a creative agency that's pitching a new campaign for your company." And now, the task: "Act as my potential client. When I provide answers, critique the answers if needed. Ask follow-up questions. Continue the conversation until I give the stop role. Break, then give me a summary of the whole conversation, highlighting ways I can improve my pitch." You also want to include additional material references for your agent. "I've included the brief the car company provided me that has all the relevant information for this project. Use the information from this brief to inform your answers."
AI agents can be super powerful if you can design them correctly, and these are only two examples. I really like how the course also provides a guideline for how to create any AI agent. First, you need to assign a persona that you want the AI agent to take on. For example, "Act like a successful personal fitness trainer and talented nutritionist." Step two is that you want to give as much context and detail as you can about the scenario and the conversation. For example, "I'm looking to improve my overall fitness and adopt a healthier lifestyle." Step three is to specify the type of conversations or the kind of interactions that you want to have with the AI agent. You might also want to set some rules to follow, like, "Ask me about my workout routines and meal planning and give me feedback." Step four is to provide a stop phrase in order to stop the conversation. This can literally be anything you want, so go wild. An example they give is, "No pain, no gain." And finally, step five, as a tool to provide feedback or areas of improvement. "At the end of our conversation, provide a summary of the advice you provided."
And that is it, my friends. You have now completed the Google Prompting Essentials course and saved 9 hours of your time. But as promised, to make sure that you actually have retained this information, I will now put on screen the questions for the little assessment. Please answer these questions to actually retain the information you've just learned. You can like say it in your head, you can say it to your friend, your dog, your cat, whatever. But for proof, you should write it in the comments. Thank you all so much for watching, and I will see you guys in the next video. We livestream.