Learn PROMPT ENGINEERING from Scratch – Full FREE Course (Generative AI, any LLM)

Hi, and welcome. I'm Andrew Privetera, and I'm very glad that you're here.

A bit about me, I've spent about 20-plus years in business analysis, which means my job was always about understanding the root cause of a problem before jumping into devising a solution. Okay, that same approach applies to working with generative AI. Effective prompts aren't just about memorizing scripts or copying what someone else did. They're about thinking through what you actually need so that the AI can give you something useful.

I run a business focused on AI education and digital skills. And if you've seen me on YouTube, LinkedIn, or TikTok, you know I focus on making AI practical and accessible. Why this course? Well, a lot of people approach generative AI in the wrong way. They grab a pre-written prompt, paste it in, and hope for the best. But prompt engineering is actually a skill. And like any skill, it works best when you understand the why behind it, not just the what. That's why I've created this course and why I teach the GRASP framework: a simple framework that helps you build any prompt from scratch in a way that's clear, structured, and adaptable to the situation. So instead of relying on trial and error, you'll be able to guide AI intentionally and get better, more relevant results quicker.

What to expect from this course? This is a hands-on, practical course. You won't just learn about prompt engineering; you'll practice it in ways that make sense for real-world applications. By the end, you'll be able to write effective prompts confidently from scratch, no matter what tool you're using. In order to get started, this course works with free versions of the popular large language model tools like ChatGPT, Bard, Perplexity, amongst others. So, no special paid subscriptions are needed to complete the lessons or the exercises in this course.

If you're here to improve your AI skills, streamline your workflow, or just stop getting useless responses from AI tools, you're in the right place. So, let's get started.

[Music]

The Future of Jobs Report 2025, which came out in January by the World Economic Forum, highlights the rising demand for large language model or LLM skills, positioning prompt engineering as an absolutely critical capability for future careers. According to the report, there's high demand for prompt engineering skills. So, as generative AI tools experience explosive growth, as we've seen, and become mainstream, employees increasingly value advanced prompt writing skills for tasks like content creation, summarization, and multilingual translations. The growing market for generative AI training, individuals are focusing on foundational topics including prompt engineering, whilst organizations prioritize practical, productivity-driven AI applications that free up people's time from repetitive tasks in order to do more analytical and creative tasks. AI is being used as a collaborative tool. So, large language models are seen as tools that enhance human skills, not necessarily replace them, making prompt engineering an essential interface skill for guiding AI outputs effectively. And this is an important point because every AI tool that's around will require an initial prompt by a human. So understanding how to craft one from scratch is absolutely essential.

And workforce reskilling priorities: About 77% of employers plan to reskill their workforce in AI collaboration, with about 69% specifically seeking talent skilled in AI tool design and prompt engineering. So, as an overall career insight, it's plain to see these trends signal an absolute golden opportunity for everybody to master prompt engineering. As industries rapidly integrate large language models like ChatGPT, professionals with strong large language model and AI collaboration skills, in particular, will be in very high demand.

[Music]

What is artificial intelligence, or AI for short? At its core, artificial intelligence, or AI, is about creating machines that can mimic human intelligence. Okay? Whether that means recognizing speech, detecting patterns, solving problems, or even generating creative content, AI is not a single technology, but actually an umbrella term for a set of techniques that allow machines to learn, reason, and make decisions.

Artificial intelligence is often described in three phases: ANI, or artificial narrow intelligence, which is where we are today. AI systems like ChatGPT, Siri, or facial recognition tools are excellent at specific tasks but can't think or reason outside their training. AGI, or artificial general intelligence, is a still emerging form of AI that could perform any intellectual task a human can. AGI would be flexible, adaptable, and capable of reasoning across a wide range of topics without being task-specific. ASI, on the other hand, or artificial super intelligence, is actually a hypothetical future in which AI surpasses human intelligence. ASI could think faster, process more information, and actually make decisions beyond human capacity.

Today's tools, including ChatGPT and other large language models, are examples of artificial narrow intelligence, or ANI. While artificial general intelligence is still in development, its potential to mimic human reasoning at scale underscores the importance of mastering prompt engineering today. The foundational skills you're building now will actually serve you well as AI capabilities continue to expand.

Here's a breakdown of some major AI subfields and what they do:

Machine learning. This is where AI learns from data to improve over time. And some examples of this might be Netflix recommendations, which you would probably be familiar with, and fraud detection.

Deep learning. This is an advanced type of machine learning that uses neural networks to recognize patterns in large data sets. So, it's mimicking the human brain here. And examples of this might be self-driving cars and facial recognition.

Statistical learning is a type of machine learning that uses mathematics and patterns in data to make predictions or decisions, often with simpler, more explainable models. Both deep learning and statistical learning sit under the machine learning umbrella. Statistical learning is often used when you want clear, understandable results. Deep learning, however, is better for solving more complex tasks, especially when there's lots of data to work with. And an example of statistical learning might be predicting home prices based on size and location.

Natural language processing, or NLP for short. This is when AI understands and generates human language. Examples of this might be AI finishing your sentence when you type a message in a chatbot, chatbots themselves, Siri, Google Translate.

Pattern recognition. This is where AI detects trends, relationships, and anomalies in the data. So, here common use cases might be stock market predictions and medical diagnostics.

Robotics. This is where AI controls machines that interact with the real world. So, this might be industrial robots or, as we've seen lately, the rise of robotic assistants.

Computer vision. This is where AI sees and interprets images and videos. And some examples of this might be medical imaging, self-driving cars, security cameras, and phone unlocking.

And generative AI. This is where AI creates new content in the form of text, images, and music. Now, the examples you might be familiar with or have used are ChatGPT and similar large language models, DALL-E for images, and Midjourney as well. They're a subset of AI that creates new content instead of just analyzing or recognizing it. So, LLMs are a type of AI model specifically trained to understand and generate humanlike text. These models are built on machine learning and deep learning principles. At this point, we've touched on large language models briefly, but in the next lesson, we'll dive deeper into exactly what they are, how they work, and why they're important.

[Music]

What is a large language model? A large language model is a type of AI model trained to understand and generate humanlike text. It's built using deep learning techniques, specifically neural networks, and it processes and predicts words based on patterns it has learned from vast data sets. This information from the data sets might be from books, articles, and websites. Instead of thinking like a human, it predicts the most statistically likely next word in a sentence.

So, as an analogy, think of a large language model like an ultra-powerful autocomplete system. Just as your phone suggests words while you type, large language models predict what words should come next based on context, except on a much larger scale.

How are large language models different from other AI models? Well, most traditional AI models are built for specific tasks like fraud detection or facial recognition. But large language models are general-purpose AI models trained to be flexible and adaptable.

So, if we compare some of the features between the traditional AI models and the large language models:

The purpose of a traditional model might specify that it's task-specific artificial intelligence, like fraud detection, object recognition, and medical diagnosis. Whereas the purpose of large language models might be general-purpose artificial intelligence that can answer questions, write content, and more.

The learning type of traditional AI models tends to be rule-based or trained on limited data sets, whereas the learning type from large language models uses deep learning to analyze massive text data sets. So, there's a lot more information to work with here.

In terms of flexibility, traditional AI models are designed for a single function, whereas large language models can adapt to many tasks: writing, summarization, coding, etc.

So, when we're looking at examples of traditional AI models, these tend to be things like spam filters and credit card fraud detection, which have been around quite a while. Whereas, as you know by now, large language models will be models like ChatGPT, B, Claude, Perplexity AI, etc.

Why are large language models a big deal? Well, they're versatile. Large language models can write, they can summarize, they can translate, they can code, and a lot more. So, they actually process massive data sets, making them smarter than previous AI models. And they learn from interaction, improving their outputs based on the user input.

[Music]

AI doesn't think like a human. It predicts the next most likely word based on patterns it's seen before. So, you may have seen this when you're using Google autocomplete, predicting the next word that you type when you're searching for something.

At the core of every large language model is a neural network, which is a deep learning system designed to mimic the way the human brain processes information. A human brain actually has neurons that connect and pass information. Whereas a neural network has layers of artificial neurons that process data and recognize patterns. Actually, deep learning means stacking many layers of neurons to process complex information. And this is how AI recognizes language, images, and speech.

Keep in mind, though, that all large language models are trained on a huge amount of existing text: books, websites, articles, roughly the equivalent of the words from around 35 million books. But they're not updated in real time or today unless they're connected to the internet. And even then, you have to be very cautious about what outputs it gives you. Think of the model like a digital time capsule. It learned everything it knows at a certain point in time and now just predicts what should come next based on that frozen snapshot in time. So, if you were searching for information on an event that happened in the last month, and that snapshot is only up to 3 months ago, then it's likely that when you prompt for information on that recent event 1 month ago, nothing will come up, or it will give you outdated information.

Also, note that AI doesn't know facts the way we do. It just predicts statistically probable responses. Okay? That's why it can sound confident even when it's making things up or hallucinating, as we'll see later on.

How does a large language model generate text? Well, when you type a prompt, the large language model follows a three-step process:

Step one: Tokenization. Imagine you're typing the message, "I'm going to the..." Before the AI can understand what you're saying, it breaks the sentence into smaller pieces called tokens. Tokens are like building blocks. They can be whole words, parts of words, or even punctuation. For example, "I'm" is token one, "going" is token two, "to" is token three, and "the" is token four from the phrase I just read out.

Step two: Prediction. Now, the AI looks at the tokens so far and tries to predict the most likely next token based on patterns it's seen in the past from billions of text. It might ask itself, "What word usually comes after 'I'm going to the'?" It could consider options like "store," "gym," "movies," and "beach," for instance.

Step three: Generation. So, once it picks the most likely next token, let's say "store," it adds that to the sentence. So now the sentence becomes, "I'm going to the store." Then it repeats the process again. Based on the full sentence, it keeps predicting the next token and the next until a full response is generated. Think of it like autocomplete on steroids, if you like. But instead of just guessing one word, it can write entire emails, poems, or code by stringing together one word at a time with lightning-fast predictions.

So, why does AI sometimes get things wrong or hallucinate? Well, since large language models don't have true understanding like humans, where they predict words based on probabilities, they will sometimes hallucinate, which is generating incorrect or misleading information. Why does this happen? Well, AI relies on patterns, but patterns don't always mean the right facts. So, how do we fix it? Well, with better prompting, and that's why you're here. We'll learn all about this in later modules.

So, now that you understand how AI generates responses, the next step is learning how to guide AI effectively using structured prompts.

[Music]

What is prompt engineering? Well, prompt engineering is the process of crafting clear, structured inputs that guide AI towards producing better results. A well-designed prompt guides the AI to provide useful responses. And it's designed to engage in natural language conversations, enabling you to interact with it like you're talking to a friend or a colleague. And when I say talking to a friend or colleague, I mean natural, back-and-forth conversation. So, you ask it something, it gives you output, and then you build on that by asking it something else, and then it responds, and so on and so forth.

So, ChatGPT, for example, is an advanced artificial intelligence large language model developed by OpenAI. Now, just look at the name. The "Chat" in ChatGPT refers to conversation. The "GPT" refers to generative pre-trained transformer. GPT means it's a type of AI that has been trained on vast amounts of text, for instance, the text from books, articles, and websites to understand and generate humanlike responses in a conversational manner. So, when I say it's been trained on the text from books, articles, and websites, they've obviously been written by humans. So, it over time, with more training, starts to understand the nuances of human language and expression and emotion. So, it's kind of like the way that a child would learn to speak as they're a toddler. Okay? So, it understands the phrases, it remembers the phrases, then it builds on those as the child grows, and it can refer to those phrases again and again in the right context. So, this is why we need to give the prompt more context when we're asking it a question.

So, let's look at an example of a poor prompt. You might ask the large language model, "Tell me a joke." Is this a good prompt? Well, it's fairly general, isn't it? It's not giving a lot of context about who you're targeting, what sort of tone you want. It's just lacking the context that it needs to actually draw on information to be able to provide a good output.

Now, here's an example of a better prompt: "Tell me a joke that a 10-year-old would find funny and is related to animals." So, why is this second prompt better? Well, it gives AI a clear role that is a 10-year-old. It defines the length and structure of the response, and it specifies context and constraints. For instance, the child's humor related to animals. So, not only is it targeting the humor that a 10-year-old is likely to understand, it's also targeting that into the subject matter of animals as well. Okay? So, if I was to ask it, "Tell me a joke," how is it going to know that any of that's relevant? So, what I'm getting at is that you need to add that context to make it meaningful or give a meaningful output.

[Music]

What can large language models actually do? Well, let's take a look at an overview of the main functions they can perform.

The first might be natural language understanding and generation. And this is where it processes and responds in human-like language in things like chatbots, email drafting, and summarizing articles. An example of this might be asking AI to draft an email response to refine a job application cover letter.

Another use might be text creation with contextual relevance. And this means it might generate coherent responses based on the context provided. And this could be, uh, for blog writing, storytelling, marketing copy, etc. An example of this might be using AI to generate a social media post or help brainstorm content ideas.

It might be used in multilingual translation where we translate between languages accurately, used in things like travel assistance and international communication. So, an example might be translation of a business email from English to French before sending, or traveling with your mobile app and having a conversation with someone who's trying to sell you food so that you can understand each other.

Could be question answering and general knowledge retrieval, which is where it provides factual answers based on trained data. Um, this could be, you know, research assistant, quick fact-checking, and an example might be asking AI to summarize a complex industry report before an upcoming morning meeting.

It could be used in sentiment and emotional tone analysis. So, this is where it detects emotions and tone in written text. Um, this could be used in things like customer feedback analysis and social media monitoring. And an example might be is where we use AI to analyze customer reviews and detect common sentiment trends, um, to detect whether the overall majority are favorable or negative reviews, or, you know, what the temperature in the room is at the time.

It could be used for text summarization where it condenses long content into key points. Um, might be used for things like executive briefings and meeting summaries. And an example might be to request a quick summary of a long podcast episode you missed, which potentially might save you hours.

Another might be content classification and categorization. So, this is where it groups text into meaningful categories for analysis. Might be for email sorting, topic detection in articles, and things like that. Uh, and we might use AI to automatically sort emails into priority and non-priority folders.

It can simulate humanlike dialogue where it engages in dynamic conversations, such as those seen in virtual assistants and customer support bots where we've continually chatted to them back and forth. So, we engage in this back-and-forth conversation with AI for either trying to find out answers to things or brainstorming.

It could be used in personalized content tailoring where it actually adapts responses based on the user input and their preferences. It could be used as an adaptive learning tool or make personalized recommendations in shopping or food. And an example might be to ask AI for personalized movie or book recommendations based on past preferences chosen by that user.

And it can also perform structured information extraction. So, this is where it identifies key data from unstructured text. So, when I say unstructured text, I mean it's text that's mixed up in a whole lot of subject areas, um, different sentence lengths, there's no actual structure to it. So, therefore, you can't classify it. So, what it actually does, it will extract names, dates, and financial data from reports and put them in a structured format or a tabulated format so that you can actually analyze it quickly and easily. So, an example might be extracting key data points from a contract before signing it so that you might not have to read the whole, you know, 20 pages of the contract.

So, just keep in mind that large language models can have limits, and that's where AI lacks true understanding and can make factually incorrect statements. AI relies on pattern recognition, not reasoning, and poorly crafted prompts lead to poor responses. But overall, large language models represent a huge step in the development and advancement of artificial intelligence. Large language models are much more versatile and flexible compared to traditional chatbots, which are often too rigid and fail to respond in humanlike ways. In module three, we'll look at how different AI models handle these capabilities and when to choose one over another.

[Music]

How do large language models differ? Well, not all AI models are the same. Some are better at creativity. Others specialize in accuracy, research, or coding. So, here's a quick guide:

For general help or brainstorming ideas, use ChatGPT or Claude.

For writing blogs, emails, and creative content, use ChatGPT or Claude, where they're good for blog posts, storytelling, and ad copy, and also handle complex instructions very well.

If you want to check factual accuracy with citations, use Perplexity AI. So, Perplexity provides real-time citations for factual accuracy. And another one you might want to consider here is Google Gemini, which is good at extracting key insights from data.

To do research and summarize long texts, use Claude and ChatGPT.

For business tools and productivity, use Microsoft Copilot. To automate tasks in Microsoft tools like Word and Excel, again, Microsoft Copilot.

For live up-to-date information, use Bard from Gemini, you.com, Perplexity AI, which actually retrieves live web data with source citations, or ChatGPT when web browsing is enabled.

To customize response tone or format, use you.com, where you've got the adjustable settings to be able to do this.

For AI-assisted coding and technical solutions, use DeepSeek AI, Microsoft Copilot, or ChatGPT 4 with Code Interpreter, although that's more of a paid feature, which we'll cover later on here. It's great for code debugging, generating scripts, and writing automation. Copilot also integrates directly into developer tools.

For the most advanced capabilities, ChatGPT 5, which is upcoming, will likely be the most powerful with that deep reasoning, improved multimodal capabilities, that is the ability to produce text, um, audio, and visuals, and integration of all of those GPT models I was talking about before. Free users will have access to a standard intelligence version of GPT 5, while Plus and Pro subscription members will be able to unlock higher intelligence levels, enhanced search capabilities, and research-based tools. Meanwhile, GPT 4.5 or Orian is expected to refine further response accuracy and usability over previous models ahead of the full release of GPT 5 later in 2025.

Now, one thing I mentioned in a previous lesson is that the training data that large language models are trained on is at a certain snapshot in time, which could be a couple of months ago, right? So, it doesn't always give the most updated results. Now, when performing a prompt on a large language model to get over this fact that it's a little bit out of date, is that we can do an internet search. Okay? Which will give us the latest data. Okay.

Now, if we look at the large language models we just compared, the ones that will be able to perform this internet search, if you would like the most up-to-date information included in your prompt, are ChatGPT, provided the browsing is actually enabled, and this is actually using a Bing search, not always available. Internet search is always enabled on Bard or Google Gemini. Again, always enabled on Perplexity AI, always enabled on you.com as of the date of filming. This unfortunately leaves three models where internet search is not available: and that's Claude by Anthropic AI, Microsoft Copilot, and Deepseek AI.

[Music]

Now, to start using large language models, you don't need to download anything or create an account. All you need, actually, is a web browser and internet connection. Therefore, you don't need to pay to create an account to start learning how to use it. The prompt engineering techniques covered in this course could be used across multiple large language models.

So, as a quick reference where you can access different AI models and their general features:

We've got ChatGPT, which has the free GPT 4.5 Orian and GPT 5 versions, but also the paid GPT 5 Plus and GPT 5 Pro versions for the much more advanced features.

Claude, which has both a free and paid version.

Bard, which is free.

Perplexity AI, which has both a free and paid version.

Microsoft Copilot via Edge, Windows, or Microsoft 365, which has a free and paid version, which is integrated with Microsoft 365 products.

Deepseek AI, which is free.

And you.com, which is also free.

Note: Upon the release of GPT 5, OpenAI will restructure access tiers. So, free users will have unlimited access to GPT 5 at a standard intelligence level. The Plus subscribers will have access to a smarter version with improved reasoning. And the Pro users will have advanced features, including deep research, voice commands, and AI-generated visuals.

Okay, now it's time to move on to module four, which will be crafting prompts from scratch.

[Music]

Now that we've explored what LLMs can do in module 3 and why prompts are essential from module 2, it's time to now learn how to write structured, effective prompts from scratch. Most AI users struggle with getting the right results because they're either giving vague instructions or assume AI knows exactly what they want. The GRASP+Q framework that we'll be using is designed to help you build prompts in a structured way, ensuring AI understands the context, goal, and constraints before responding. So, this module will guide you through crafting better prompts that lead to more useful AI-generated content.

Once you've logged onto your chosen large language model, a chat prompt will appear like so on the screen. You simply start typing in the prompt chat area what it is that you want. Once you've entered your command, you simply press enter, and the large language model will return its response. As you can see, I'm answering the prompt output with another question that I believe is needed to obtain a more thorough response.

Now, why is it important to define intent? Well, imagine you're about to bake a cake. You wouldn't just throw random ingredients into a bowl and hope for the best, right? You'd think about the kind of cake you want, gather the right ingredients, and follow a recipe. Crafting prompts for AI is pretty similar, where we need to provide context. So, providing context helps the AI understand your intent. If you're drafting an email, specify the tone and purpose to get the desired outcome. So, if you'd like the email tone to be more assertive and not aggressive, then tell it so, because it won't know actually whether if you're demanding something, whether you want to be aggressive or assertive.

Iterating and refining prompts over a number of back-and-forth questions between you and the large language model enhances the quality of those AI interactions, leading to better, richer results.

So, what are the pitfalls of pre-made prompts? Well, sometimes users rely on readymade prompts produced by someone else, and they just change a few words here and there. Trying to prompt solely on pre-existing prompts is like using a one-size-fits-all recipe. It might work sometimes, but often it doesn't quite hit the mark. Besides, this can be very time-consuming, as you would need to try and find a scenario that fits what you're trying to do from a library of potentially hundreds or even thousands of prompts that someone has published online. And this is why identifying your specific needs is important.

So, before you even think about a prompt, ask yourself: Firstly, what exactly am I trying to achieve? Secondly, why do I need this information or action? And thirdly, what background details are essential?

So, why this approach works: It encourages critical thinking in the user. Okay? By focusing on your specific needs, you're engaging more deeply with the task, leading to more effective and richer prompts. It promotes tailored interactions. This method ensures the AI responses are more relevant and useful to your particular situation. And it enhances adaptability or flexibility. So, instead of sifting through countless prompts, as I said, you'll have the skills to craft the perfect one for any scenario. And let me just say, employers are actually looking for people who can craft a prompt from scratch rather than trying to access a library of hundreds or thousands of prompts and just tailor-make it for their situation. So, this is why I'm teaching you this method from scratch.

So, as a quick example, let's say you inform your team about a project delay. If you're a project manager, you could search for a generic prompt email template to use in the large language model, but it might not address the specifics of your situation. Alternatively, by understanding your exact needs, you can craft that message that clearly communicates the delay, the reasons behind it, and the next steps, ensuring your team is well-informed and prepared.

[Music]

To make prompt engineering less overwhelming for beginners, I'm introducing my GRASP+Q framework, a structured, logical approach that helps you construct prompts from scratch based on your needs.

So, what is GRASP+Q and why does it work? Well, it's a six-step method that ensures AI understands your request clearly and responds accurately. It also includes an important step where you ask the large language model to ask you any clarifying questions before generating a final response.

So, let me break down the framework for you first.

So, the GRASP+Q framework starts with G for Goal. Okay? That's what do you want the AI to achieve? So, this will define the outcome of the AI's response. As an example, I might say as a goal: "Write an educational blog about the decline of English grammar standards in the higher education system." That's a clear goal.

The second letter, R, is for Role. And this is who should the AI act as. And what I mean by that is what role would you like the large language model to play to bring in, you know, certain tone, expertise, and a perspective. And an example of that might be: "Play the role of a senior lecturer with a PhD."

Following on from that, A for Additional Context. And this is what additional background details are essential for more understanding. So, an example here might be: "The educational blog should include reference to the Education Institute study in 2024 that specifically mentions how the standards have declined from the years 1900 to 2024." Okay? So, giving this extra context will ensure the accuracy and specificity of what that output gives. So, by defining this extra context, it'll ensure a more accurate and specific output.

S is for Structure. And this is basically the format that the response should follow. So, this will define the style and specific look and feel of the written piece. An example of what I might put in here is: "It should be written in a professional style targeting other education academics that would typically read academic papers. Start with a good hook that captures the reader's attention and use bullet points and bolding to highlight key points." Okay? So, you see that's all about the specific look and feel and the style. You know, they would like the language to be more academic-oriented as opposed to someone that's at high school or primary school level.

P is for Parameters. And that is the rules and limits or the refinements that AI should follow. So, this would define, say, the length of the written piece and prevent irrelevant or excessive output. So, an example here might be something like: "It's to be no more than about 500 words, and content is to be accurate and up-to-date. Please include cited sources of information."

And finally, we have the +Q, which is the clarifying question that we will always ask it to make sure that it asks us questions about what it thinks might be relevant. So, let me give you an example of what we put here. So, what I would type in here is: "Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer." So, this will ensure that the large language model will clarify any uncertainties before generating a response and will potentially, you know, save some wasted time. So, the reason we include this last field as it is, is to prevent any misinterpretations or incorrect outputs.

So, let's now apply this framework to a real situation, and I'll take you through the steps that are needed to complete it to create a prompt from scratch.

[Music]

All right. Now, let's apply the GRASP+Q framework in different scenarios. But first, I'd like to take you through the four steps that you would typically use to apply the framework effectively.

So, if we were to look at an example of writing a travel blog post. The scenario might be: Lisa wants AI to write a travel blog about Santorini in Greece for first-time visitors. So, step one, I'm going to run through how the GRASP+Q framework structures it logically. So, if you'd like a bit of practice at this stage, you'll get plenty of practice after. Um, just turn the video off and see if you can fill out the elements of the GRASP+Q framework, and then compare it to what I've come up with. Okay?

So, applying the framework to that scenario:

The Goal: Create an engaging, informative travel blog to help first-time visitors plan their trip.

Role: AI should act as the travel journalist specializing in European destinations.

Additional Context: The blog should include must-see sites, local culture, food recommendations, and ideal travel times.

The Structure: Write a casual, friendly travel blog organized into sections with headings.

Parameters: Ensure content is accurate and up-to-date, avoiding generic and outdated recommendations, and keep it under 300 words.

And of course, the clarifying question as the Q is: "Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

So, that is how I would start filling out the framework as the first step.

So, the second step, following on from that, would be to create your prompt based on the table or what you've just filled out as the elements of the GRASP+Q framework into the LLM input, or the large language model input on the screen, and that would be: "You are a travel journalist specializing in European destinations. Create a travel blog about Santorini for first-time visitors, including must-see sites, local culture, food recommendations, and ideal travel times. Organize the response into sections with headings. Keep it under 300 words. Ensure content is accurate and up-to-date. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

Okay, there you go. That's the prompt that would be entered into your large or your chosen large language model. And then you would hit enter.

And then step three, it's most likely going to ask you some questions here. So, the large language model will then ask you a series of questions, of course, because you've prompted it to do that in the Q part of the framework, or the question part of the framework, in a numbered format. So, for example, it might ask:

1. Do you require an emphasis on beautiful beaches?

2. Would you like a list of the best sunset locations?

3. Would you like me to provide luxury boat tours?

4. Do you need to cover any visa requirements?

5. Are you targeting a particular age group?

So, you then answer these questions by simply typing in the responses into the large language model to the relevant question number. And the way I would do this is that I just put the relevant number that it asked and my answer next to it. So: "1. Yes, but a maximum of three. 2. Yes. 3. No thanks. 4. No. And 5. Not really, but I'd imagine it will be for over 25s."

And then finally, step four. Once you have these answers in the large language model, it will then give you the final response of your query. So, it's as simple as that. So, when you look at the response it gives, just read it. And if you think of something else you need to add, simply type in your extra question or requirement as a follow-up and engage in a further back-and-forth conversation until you achieve the output you want. Now, of course, I add this because that's naturally what you will be doing with a large language model. It's that back-and-forth conversation. So, it's like when you're talking to a colleague or a friend or a family member, they're explaining something to you. They might give you an answer, but then you think of something else you want to ask. Um, so you ask them back, and then they'll answer it. It's exactly the same thing. Okay?

The GRASP+Q framework isn't about rigid, predefined prompts. It's a structured way to help learners think through their AI interactions rather than just memorizing or reusing prompts word for word. It also isn't a checklist; it's a conversation guide. So, GRASP+Q gives your prompt shape, but you can say things naturally as well. Think of it like sketching before painting. You still bring your own style. It's basically using a guide in order to think through a prompt logically. Over time, and with repeated use, it will become second nature. This approach strikes a fantastic balance between encouraging natural conversational prompting, which aligns with reducing the cognitive load, providing a structured foundation for those new to AI interactions so they don't feel lost or overwhelmed, and avoiding rote learning by focusing on adaptability and critical thinking in prompt crafting.

Okay, now it's your turn. So, let's practice using this framework in various scenarios. But before we dive into more examples, let's get you thinking about where you might use these prompts in your own life and work. Although the framework can be applied to any prompt, below is a list of the more likely areas in which it will be incredibly useful. And these could be:

Everyday conversations, say awkward conversations you need to have with a family member. So, it can give you the practice for doing that.

Personal development of fun, professional settings, research, creative writing, brainstorming ideas, text analysis and summarization. Say you've got a huge report, you know, a 100-page report, and you'd like a one or two-page summary because you just don't have time to do it.

Problem-solving, complex queries, and troubleshooting, where you could use a large language model to help you walk through the steps of fixing a problem that you might come up with when you're say installing software on your computer.

So, as you hear these examples, consider which of these areas are probably most relevant to you in your everyday life and work, and where do you already see opportunities to use large language models more effectively?

Okay, let's get you practicing crafting some prompts. So, what I'm going to do is read five scenarios to you. Please take note of those scenarios, and then for each scenario I give you, create the elements of the GRASP+Q framework. So, that is, fill in each of the elements as stated, and then once you've done that, take the elements of that framework and put them into a paragraph, and in effect, that will become your prompt for that scenario.

So, if I read the prompts out to you:

The first scenario is writing an email to your boss about an unrealistic deadline. So, you need to email your boss explaining that the current deadline for a major project is unreasonable while maintaining professionalism and offering potential solutions.

The second scenario is brainstorming solutions for a noisy open-plan office, where your open-plan office is way too noisy. It's making it hard to concentrate. So, you need a creative solution or solutions to address the issue.

You're hosting a dinner party on a hot summer's evening and need fresh, creative recipe ideas.

The fourth scenario is explaining blockchain in a way anyone can understand. So, you need AI to explain blockchain technology in the simplest way possible, avoiding technical jargon.

And the last scenario is summarizing a long-form article on climate change, where a researcher needs a concise summary of a lengthy article entitled "The Uninhabitable Earth" by David Wallace Wells, published in New York magazine.

Okay, so turn the video off, get cracking, and resume the video so we can go through and compare what we've got. Good luck.

Okay, welcome back. How did you go? Did you find it challenging, or was it quite intuitive? All right, let's work through each one of the scenarios, and let's compare it to what I have. You know, it won't be exactly the same, but as long as you've got a feel for each element of the framework and what it means and how you can use it to create a prompt from scratch.

Okay, so the first scenario was writing an email to your boss about an unrealistic deadline. Okay? So, as the goal, I've put: "Draft a professional email to my boss explaining why the project deadline is unrealistic." The role I've put is: "You are a project management expert with experience in corporate communication." For the additional context, I've put: "The project has tight constraints, limited resources, and unforeseen issues that make the deadline unattainable. I want to propose alternative solutions without sounding negative." For structure, I've put: "Write a clear and concise email, maintaining a professional tone." For the parameters, I've put: "Ensure the email remains respectful and solution-focused. Keep it under 250 words." And of course, for Q, the clarifying question, as always, we put: "Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

Okay. So, when I put that all together from my suggestion, the framework gives me the following prompt: "You are a project management expert with experience in corporate communication. Draft a clear and concise email to my boss explaining that our current project deadline is unrealistic due to tight constraints and unforeseen issues. Maintain a professional tone. Propose alternative solutions, timeline adjustments, or resource changes, and reinforce our commitment to the project's success. Keep it under 250 words. Ensure the email remains respectful and solution-focused. And before responding, please ask me any clarifying questions until you're confident that you can give me the correct answer."

So, you'll then take this prompt, you'll put it into the large language model of your choice, you'll hit enter, and then it will come up with the questions that you need to answer, and then you'll put in the answers to those questions, hit enter, then it will give you the response that you're hopefully looking for, and if not, obviously, you can ask it more questions.

Okay, let's look at the second scenario, which is brainstorming solutions for a noisy open-plan office. So, in my example, I've put:

The Goal: Generate a list of practical and creative solutions to reduce noise distractions in an open-plan office.

The Role: AI should act as a workplace productivity consultant specializing in office environments.

A for Additional Context: The office is noisy due to phone calls, conversations, and background distractions. Solutions should be budget-friendly and minimally disruptive to team dynamics.

So, for Structure, I put: "Provide a list of five solutions categorized into physical changes, behavioral strategies, and tech-based fixes."

For Parameters, I've put: "Ensure that solutions are realistic, feasible, and adaptable to different office sizes. Avoid overly expensive solutions unless cost-effective alternatives are also provided. Present in a way that anyone can understand."

And for Q, which is the clarifying question, as always: "Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

All right. So, then putting these elements together in my prompt, I've got: "You're a workplace productivity consultant specializing in office environments. Generate a list of at least five creative and practical solutions to reduce noise in an open-plan office. Categorize solutions into physical changes, behavioral strategies, and tech-based fixes, and present in a way that anyone can understand. Ensure that the solutions are budget-friendly, minimally disruptive, and feasible for different office sizes. Avoid overly expensive solutions unless cost-effective alternatives are also provided. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

Again, I'll put this prompt that I've just devised into the large language model and press enter. It will give me the questions I need to answer. I'll answer those questions, hit enter again, and will hopefully give me a final response. But if I read it and I'd like, I think of something extra or I don't think there's something quite right, then I will ask it to follow up with a refinement of that or ask it another question or ask it to change something.

Okay. The next scenario is generating the summer dinner party recipe ideas. So, for the Goal here, I have: "Provide a selection of summer-friendly dinner recipes that are easy to prepare and refreshing." For the Role, AI should act as a professional chef or food blogger with expertise in seasonal cooking. For Additional Context, the meal should be light, suitable for hot weather, and include options for different dietary preferences, for example, vegetarian, gluten-free, seafood, and grilled meats. For the Structure, I've said: "Provide three appetizer options, three main course ideas, and three dessert suggestions along with a brief preparation summary for each." And Parameters: I've put: "Recipes should require minimal cooking time, prioritize fresh and seasonal ingredients, and be suitable for serving to a small group of about six guests." And of course, the clarifying question at the end is always: "Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

So, putting these elements from the GRASP+Q framework together again: "You are a professional chef or food blogger with expertise in seasonal cooking. Provide three appetizer options, three main course ideas, and three dessert suggestions that are light, refreshing, and ideal for a hot summer evening. Include a brief preparation summary for each dish. The recipes should require minimal cooking time. Prioritize fresh and seasonal ingredients and be suitable for serving to a small group of six guests. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer."

For the next scenario, which is explaining blockchain in a way anyone can understand, I've put:

For Goal: Explain blockchain technology in a simple, easy-to-understand way suitable for a complete beginner.

For Role: AI should act as a tech educator or teacher who specializes in simplifying complex concepts.

For Additional Context: The explanation should avoid technical jargon, use relatable analogies, and focus on why

blockchain matters rather than the deep technical details. For structure, I've put present the explanation in three short sections. Firstly, what blockchain is, secondly, how it works, and three, why it's important. For parameters, I've put keep the explanation under 300 words, use real world analogies, and assume zero prior knowledge of blockchain or cryptocurrency. And of course for the clarifying question at the end I put before responding please ask me any clarifying questions until you're confident that you can give a correct answer. And when I put this together my prompt comes out as you are a tech educator who specializes in simplifying complex concepts. Explain blockchain technology in a simple and easy to understand way for a complete beginner. Structure the response into three short sections. Firstly, what blockchain is, secondly, how it works, and thirdly, why it's important. Keep the explanation under 300 words. Use real world analogies. And avoid technical jargon. Assume zero prior knowledge of blockchain or cryptocurrency. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer and then so on and so forth. Put that prompt into your chosen large language model. Press enter. Get the response. It will ask you questions. You answer those questions. Press enter and then take it from there. Okay. Okay.

And the last scenario summarizing a long form article on climate change. So I've put for the goal summarize the key points of the article the uninhabitable earth by David Wallace Wells. For role I've put AI acts as a research assistant proficient in climate science communication. So for additional context, I've put the article should discuss potential worst case scenarios resulting from climate change, emphasizing the urgency of the crisis. For structure, I've put provide a summary in bullet points, each highlighting a major point from the article. And for parameters, I've put limit the summary to no more than 150 words. Ensure accuracy and neutrality, avoiding personal opinions. And of course for the clarifying question at the end before responding, please ask me any clarifying questions until you're confident that you can give a correct answer. So the prompt that I would put in my large language model will be you are a research assistant proficient in climate science communication. summarize the key points of the article, The Uninhabitable Earth by David Wallace Wells. The article discusses potential worst case scenarios resulting from climate change, emphasizing the urgency of the crisis. Provide the summary in bullet points, each highlighting a major point from the article. Limit the summary to no more than 150 words. Ensure accuracy and neutrality, avoiding personal opinions. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer. All right. So, I then obviously put that into the large language model and go through the same steps that we have for all the others. So, I'm hoping at this point it's starting to become second nature to put together the framework. you're getting a better feel for each of the elements of the grasp plus Q framework itself and how easy it can be to actually apply this to any situation.

[Music] Okay, now that you've had the chance to see how the framework, when put together, can help you devise a prompt from scratch, I'd like to show you how the framework adds value overall and what happens when you don't use the framework and of course when you do as a comparison. So firstly, here's a vague prompt many people might try. So in the prompt you might type, help me write an email about a project delay. You hit enter and the large language model might give you an example output that says subject project delay. Hi, I just wanted to let you know that the project is running behind schedule. We're doing our best to get it back on track. Thanks for your understanding. Now, this isn't actually wrong, but it's very vague, weak on details, and gives no real plan. It's the kind of email that raises actually more questions than it answers. So, so I'd like to show you now what happens when you structure your prompt with just a bit more clarity using the Grasp plus Q framework. So once you go through the steps of filling in all the the elements of the grasp plus Q framework, you'll come up with a prompt that says something like, "You are a project manager with strong communication skills. Draft a professional email to a senior executive explaining that the current project deadline is unachievable due to unforeseen technical delays. Propose a revised timeline and outline next steps while maintaining a respectful tone. Keep it under 200 words. Before responding, please ask any clarifying questions until you're confident that you can give a correct answer. Now, let's have a look at what the large language model output might be. Okay. Subject: Revised timeline for project XY Z. Hi, I'm writing to inform you that due to unforeseen technical challenges, our team will be unable to meet the original project deadline of April 15. Despite best efforts, several key dependencies encountered delays that impacted our delivery schedule. We're currently rescoping the remaining work and anticipate a revised delivery date of May 10. I've attached a proposed updated timeline for you to review. We remain committed to delivering a high quality outcome and will continue to provide weekly updates. Please let me know if you'd like to meet to review the plan in more detail. Best regards. Now, having a look at the difference between the two prompts. This is why having some sort of structure matters. Having the grasp plus Q framework helps you get clearer, more thoughtful, and more useful responses from any AI model.

[Music] So in module five, I'll be walking you through a number of tips that I have that are tried and tested that will help with optimizing large language model interactions, refining their responses, and avoiding the common pitfalls that a lot of people get into with prompt. engineering. So the first few I'd like to go through are about shifting your mindset in how you should approach collaborating with a large language model. So tip one, speak to large language models like they're human. It's as simple as that. Many users treat the large language models like search engines, say Google, but engaging with them as conversational partners leads to much better responses. So engaging with the large language models like they're human improves both the response quality and the user experience. So why is this? Well, large language models are trained on vast amounts of human language data, allowing them to mimic the conversational styles. Over time, with more training, they develop a nuanced understanding of language, similar to how a child or a toddler learns humor, tone, and context as they grow older. So, when users provide open-ended, friendly prompts, the AI mirrors this with more detailed, engaging, and contextually rich responses. So this approach also reduces cognitive load for the human meaning it makes interactions easier by allowing users to speak naturally. Okay? Uh rather than overthinking how to phrase a prompt in a structured or rigid way. Just as we don't consciously think about syntax when speaking to a person, treating AI conversationally allows for a more intuitive fluid exchange. So, additionally, social cognition, our brain's ability to process social cues and engage in natural dialogue plays a key role. So, when users treat AI as a collaborator rather than a search engine, their brain activates the same neural pathways used in human interactions, which encourages more thoughtful, iterative refinement of prompts, leading to deeper engagement and better outputs. So, you might be asking at this stage, well, how are we going to talk to it like a human using a framework like grasp plus Q? Well, I'm glad you asked because when building the prompt using the grasp plus Q framework, we still use the conversational and free flowing language as if you're talking to another person giving expanded context. So as an example, as we're filling out each element of the BR of the grasp plus Q framework. So if we were to use a for additional context, say something like the team is having a really hard time working together with the toxic environment they have to work in, which is absolutely demotivating them rather than saying something like team rapport is toxic. Okay. So, you can see that the first approach I'm just brain dumping what it is I'm feeling about that additional context at the time. What I'm feeling is coming out in words naturally. I'm not having to think about it too much rather than thinking of particular keywords, which is what we tend to do in searches or Google searches.

Tip two, engage the large language model like a thought partner or colleague instead of a tool. So when defining the role are in your prompt framework, specify the AI's persona to enhance collaboration. For example, say you are an experienced business consultant. Don't just say you are a business consultant. So by saying experienced business consultant rather than just saying business consultant, we're telling the large language model that it needs to be someone who is an absolute expert. That's everything about business. And of course our framework grasp plus Q's connection here is that this is the R for role in the framework itself.

Tip three, iterate, refine and experiment. So prompt crafting is an iterative back and forth process between the user and the large language model. The large language models improve with feedback and adjusting these prompts based on responses helps refine the results. So here we rephrase and adjust the constraints to improve responses where certain context may have been missing the first time around. So, as an example, we might say, "Can you simplify this response for a 10-year-old reader?" So, we probably realized when doing this prompt as an afterthought that the text that it gave us was probably a little bit too complex for the target audience that we wanted. So the large language model because it is been trained on large amounts of text and data will understand what level of language a 10year-old will use.

[Music] The next set of tips for this lesson are structuring prompts for maximum clarity and control. Okay, so tip four, structure prompts effectively. Large language models struggle with unclear or jumbled queries. So structuring prompts clearly and providing a defined format improves the output quality. This is especially important for technical or datadriven responses. So in this case, use clear, well ststructured sentences and delimiters which mitigates ambiguous language handling. So an example of this might be list the top five trends in digital marketing separated by bullet points. And the connection to our grasp plus Q framework here is clearly specify the output formats in S for structure.

Tip five, be specific and detailed. So providing specific instructions helps prevent generic outputs. Vague prompts often lead to surface level responses while detailed prompts drive accuracy and relevance. So include purpose, audience, and key points to avoid the vague outputs. So an example of this might be explain the benefits of automation for reducing business costs rather than saying list benefits of automation. So when it talks about the benefits of automation, it will actually give more context around how automation itself is important for reducing business costs and all the associated things that need to be done to achieve that goal rather than just talking about the overall benefits of automation. And the connection to our grasp plus Q framework here will be the additional context a that ensures that precision.

Tip six, split complex tasks into simpler steps. So complex requests often confuse large language models, resulting in incomplete or disorganized outputs because in short, we're giving it too much information. So breaking down the tasks into smaller steps which is often how we explain technically difficult things to other humans ensures more accurate results from the large language model. This is also called chain of thought prompting and is especially useful here. So if we were doing financial budgeting where we're churning quite a lot of numbers with the help of the large language model we might say to it number one sum the total expenses and then number two calculate the average monthly cost. So we're not asking it to do it all at once. We're saying first concentrate on this task and then secondly concentrate on the next task. Okay to give it time to think properly and not jumbling its logic. The second example where we might be doing some problem solving is that we might say one list the possible causes of the problem and two suggest solutions ranked by likelihood. Okay, again I'm breaking it down into two separate parts rather than just blurting it all out. And the connection to our grasp plus Q framework here is that we're using P for parameters for the stepbystep outputs.

So tip seven, use clear formatting. Okay, so when we're working with a lot of data or we we're producing a lengthy report and we want to understand a number of concepts in a logical framework in a logical way. We ask it to use bullet points, lists, and tables for the output. So we'll be talking more about this in module six. But the connection to our grasp plus Q framework is that we define what format we require in S for structure.

[Music] In this lesson, I'll be walking through some tips around cross-checking large language model outputs for completeness. So the next tip, which is tip eight, is that we as the human user embrace continuous learning and crossverification. So you should stay current with large language model advancements and comparing outputs across different models to ensure accuracy and reduced chance of hallucinations. So as an example, you might want to run a particular prompt in chatbt, look at the response and then run the same prompt in Bard or Claude, okay? And see what the differences are.

Tip nine, ask the large language model to critique its own answer. So prompting the model to review its responses can help identify errors, inconsistencies, and biases. So once you've received some output, you might want to put in a prompt, can you review your previous answer and highlight potential errors or assumptions. This enhances the quality control, which encourages the AI to reanalyze its own logic and structure, improving the response accuracy. So it's actually stepping outside of itself and looking at it as a third party and saying, "Yeah, I could have done this a little bit better. This is all right but let's make some improvements on this part. So it will help detect factual inconsistencies or missing details in AI generated content. Improves clarity and readability. So that extra check will enable the AI to simplify or refine explanations for better understanding. So, it might realize that the answer it's given you is a bit long-winded and could be stated in a in a clearer way in um a shorter form. So, it ensures logical consistency which means it eliminates contradictions it might come up with and enhances the coherence of responses that you as the human reviewer might miss and it generates alternate perspectives. So this encourages the AI to present additional insights or opposing viewpoints. So when we apply this tip of critiquing its own answer in academic writing, it ensures better structure, factual accuracy and logical flow is in business reports. It will enhance the clarity and remove unnecessary complexity. In legal and research queries, it will help verify claims and identify missing citations. And in creative writing, it might improve storytelling and engagement.

Tip 10, give the model time to think. Okay, for complex reasoning or calculations, instructing the model to explain its process often results in more accurate and insightful responses. So this is where we request reasoning or multi-step breakdowns. So if we had numerical calculations that we wanted it to complete for budgeting, we might say to it, explain your reasoning before giving the final answer. And the connection to our grasp plus Q framework here might be to use G for goal to tell it to give a structured reasoning.

Tip 11, ask the large language model to search the internet if it can. So the reason this is needed is that standard large language models lack access to current events and real-time data because it has a heavy dependence on training data which is limited by the knowledge cutoff date that I talked about previously of this last data set update which might actually be a few months old. So to bridge this gap between that information cut date which might be a few months ago and today's date we use the large language models that allow web browsing for instance chat GPT. It's just to ensure that any related information between the date of the information cut off on which large language model is trained on and today's date is actually captured in the prompt output.

In this lesson, we'll be looking at tips around handling bias and ethical AI use. Tip 12, adjust the prompt to avoid biases in the data. So, biases can emerge from training data. So instructing large language models to provide diverse perspectives promotes fairness and inclusivity. Some examples of bias include historical bias. Some AI models could be trained on old hiring data favoring male candidates over female candidates for leadership roles. And of course the impact of this is that it reinforces outdated societal patterns. There could be a cultural bias. So AI translations struggle with non-western languages and the impact of this is that it limits accessibility for diverse global users. And another type of bias might be confirmation bias which is where AI provides answers that align with users beliefs instead of presenting multiple viewpoints. And of course the impact of this is that it leads to misinformation and certain echo chambers that we often see on social media. So a strategy to overcome these biases is to diversify the sources, request neutral language and cross-check outputs and consider fairness. So this could be a combination of other tips that I've talked about recently. So examples of this when prompting might be to diversify data sources. So you might tell the large language model provide perspectives on climate change from both scientific and economic viewpoints which as you can imagine might be very different. You might ask it to request natural language. So you might say rewrite this job description to be genderneutral and inclusive. You might ask it to cross-ch checkck AI outputs where you might say provide at least three different sources to support this claim. So it should give you the cited sources that you ask for. And you might explicitly ask AI to consider fairness. So you might say list potential ethical concerns in using AI for financial lending decisions. Okay. So using the grass plus Q framework, you might be including this in the parameters P or additional context A to make sure that this extra information is actually included to give that richer output.

Tip 13 and this is a very important one. Avoid using private or confidential information. So AI tools may retain or process sensitive data or confidential data. So users should avoid sharing that private information. By default, some large language models are set to actually share your information to a public audience. So if you're not aware of these settings in the large language model that you're using, uh then unfortunately it will be made public. So, so best practice to avoid this is to review the AI tool privacy policies that relate to the particular large language model that you're using. Now, in terms of avoiding confidential information getting into the wrong hands, you might want to avoid prompts like summarize my company's financial report if confidentiality is an actual concern.

Tip 14, avoid plagiarism. So in short, don't copy large language model output directly. Always alter the output that you get if you're presenting it as your own work. So the reason this is needed is that AI generated content can resemble existing materials that it's drawn from leading to large plagiarism issues. So the proper paraphrasing and citation are essential for ethical use. When I say paraphrasing, I mean actually changing the words around to make it your own. So, so if we wanted to paraphrase something, we could take something that says AI is transforming business operations to something like business operations are evolving due to advancements in AI technologies. Okay. So ensure that you are not using AI's output verbatim if you're presenting it as original work. And as an additional note here, always use citations for sourced information. And this is where large language models like Perplexity AI is particularly good in that it's known to always include cited sources for its output for authenticity.

[Music] In this module, I'm going to take you through the different AI tool capabilities, including different functions, output formats, integrations with other software, and file uploads. But first, we'll start with uploading files for artificial intelligence analysis. So here we'll talk about using AI tools to analyze uploaded documents. As an example, I may upload a PDF containing suggestions on how to become more comfortable in front of the camera for filming YouTube. Here's how to upload the PDF and ask Chat GBT to analyze that. Firstly, uploading file. Just click on the paperclip icon in the message bar to attach your file. You can either upload from your computer, Google Drive or One Drive. Secondly, analysis. After uploading the file, tell the large language model what you want it to do. You could say, "Analyze this PDF and brainstorm 15 hooks inspired by its content." Thirdly, the large language model will read the PDF and come up with new ideas related to it. In addition to PDFs, certain large language models can also analyze spreadsheets, certain comma delimited and CSV formats and images. In the next example, I'm uploading a screenshot of a system process. Then I ask chatbt to summarize how the system works without any additional context, just the picture. So uploading the image, click on the paperclip icon in the message bar to attach the image. Secondly, analyze the image and submit this prompt. Thirdly, the large language model will analyze the image and explain how the system works all without any additional context. I only fed in the image and it prompted the large language model to analyze it.

[Music] So let's talk about the different ways AI can structure its responses in the output including tables and summaries. So, so large language models can provide answers in multiple different formats including documents, table format, bulleted lists, spreadsheet format, charts, plain text, videos, code, and hyperlinks. All you have to do is ask it what output you want. So, let's look at a few of these outputs. Starting with plain text, which is the most basic form. This is what you're already using. So for example, if you ask what is artificial intelligence, the large language model will give you a straightforward response in plain easy to read text, right? This format is the default and works well most of the time. Then there's bulleted lists. So when you want information presented in an organized way, if there's an excessive amount of information that you know you're going to get back, lists are a fantastic option. So simply ask the large language model to produce numbered or bulleted lists which are helpful for step-by-step guides or collections of ideas. So here I ask what are some effective time management tips? The large language model provides a nicely formatted numerical list plus bullet points in each section. And then there's table format. So for data comparisons or displaying structured information, large language models can format outputs into tables and even spreadsheets if you like. So as an example, a business owner wants AI to generate a competitor pricing analysis for online digital marketing services presented in a clear table format. So if we put this in our framework, our grasp plus Q framework, we could specify G for goal being conduct a competitor pricing analysis for online digital marketing services. R for role. AI should act as a business consultant specializing in market pricing strategies. A for additional context. The analysis should compare pricing structures for SEO, social media marketing, and PPC services from three competitors. S for structure. Present the response in a table format with columns. Competitor, SEO pricing, social media pricing, PPC pricing, and additional services offered. P for parameters. Focus on real world industry trends, ensuring all prices are estimated based on available market data. Provide insights in a structured and easy to read format. And finally, the clarifying question cue at the end. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer. So the resulting prompt from this graphs plus Q framework we just put together will be you are a business consultant specializing in market pricing strategies. Conduct a competitor pricing analysis for online digital marketing services focusing on SEO, social media marketing and PPC. Present the output in a table format with columns, competitor SEO pricing, social media pricing, PPC pricing, additional services offered. Base the insights on industry trends and ensure readability. Before responding, please ask any clarifying questions until you're confident that you can give a correct answer. So again, we're just putting together the elements of the grasp plus Q prompt framework to give us a written prompt we can put straight in to the large language model. The next example I'll go through is creating a market research report in Excel format. So we have a scenario where a marketing professional needs AI to generate a marketing research report on consumer preferences for electric vehicles. Okay. The output should be structured in an Excel compatible format for easy analysis. So if we were to plug this into our grasp plus Q framework as normal, we can start with G for goal being generate a structured market research report on consumer preferences for electric vehicles. R for role AI should act as a market research analyst specializing in automotive trends. A for additional context, the report should include key factors influencing purchase decisions segmented by demographics. Data should be presented in a structured table format. S for structure. Provide the response as a table with columns. Demographic, key factors, preferred EV brands, price sensitivity, barriers to adoption. Output should be Excel compatible in CSV format. P for parameters. Ensure accuracy, avoid assumptions, and base insights on known industry trends. Limit the responses to five key demographics. And of course, Q for clarifying question. But before responding, please ask me any clarifying questions until you're confident that you can give a correct answer. Therefore, based on this, our resulting prompt will be, "You are a market research analyst specializing in automotive trends. Generate a market research report on consumer preferences for electric vehicles segmented by demographics. Structure the output as a table with the following columns: demographic, key factors, preferred EV brands, price sensitivity, barriers to adoption. Ensure the data is Excel compatible CSV format for easy import. Base insights on industry trends and limit responses to five key demographics. Before responding, please ask me any clarifying questions until you're confident that you can give a correct answer. And here you have the resulting output on the large language model.

Certain large language models have multimodal capabilities which means it can work with multiple modes like text and image if you like. Here I ask chat GPT to create an image of a group of kavoodles tap dancing on a theater stage in front of a large audience. Not bad. Hey So, just note that this image can actually be produced on a free plan. Another great way to use images is to load a photo of a room or space and ask it to reimagine it in a certain way. For instance, here I ask Chatch GPT to reimagine this kitchen in an industrial chic style complete with stainless steel features, wooden benches, brick feature walls, and suitable plants to make it more inviting. Great, just as I imagined it in my mind. Of course, if you needed to perfect your vision, you would ask it to change a few things to suit what you need. So let's look at another multimodal capability such as voice mode. So some large language models allow voice mode where instead of actually typing and engaging in typed text, you're able to actually prompt and converse using your voice. And the large language models own audio voice that you can generally choose. So the audio conversation will also be recorded in text as well as an output, which is pretty handy. So when speaking to a large language model using the voice command, you don't need to follow the grasp plus Q framework rigidly. Just speak naturally like you're talking to a colleague rather than formally stating all six parts such as the goal, role, etc. The user can think through them mentally as a guide or training wheels if you like and express them in a casual way just like you'd explain a request to a colleague in conversation. So having the grasp plus Q structure in mind will help you give the AI enough context to respond effectively especially in simulations or feedback based conversations. So let's give an example um in voice mode uh for say casual prompting with the grasp plus Q framework where I talk into the model and say hey I've got a tricky work situation. Can we do a role play where you're my boss and I'm trying to explain why our deadline's unrealistic. Afterward, could you give me feedback on how I handled it? On this screen recording of my chat GPT mobile app, you'll see I'm talking directly into it and I'm being answered by the voice I've selected on chat GPT. It will also keep a transcript of the recording as I'm showing you here. So, the way I said it's quite natural, but behind the scenes, it still contains G for goal. Practice a difficult conversation role. AI plays the boss. Additional context. It's about an unrealistic deadline. The structure and parameters, role play first, then feedback, and the clarifying question where AI might ask those clarifying questions naturally in the back and forth to get more context. So which large language models have voice mode? Well, chat GPT from OpenAI does um but in the paid plan only. So this will be available in the mobile app. Claude by Anthropic doesn't have it. Bard by Google does have it, but the voice input is via a mobile mic and there's no spoken replies from the model itself. Perplexity does have it, but there's a limited voice input via the mic on mobile and browser. U.com does have it, but with the mic input only and no voice output from the model, and Microsoft Copilot, which does have it and allows a natural voice interaction via the mobile and edge browser. So you'll see that most of the large language models provide the voice mode, but some will be free and limited functionality. For instance, they might not talk back to you. They'll just give a text response. And others are paid like chat GBT where you know it'll be a two-way conversation.

So besides using chat GBT on its own, this particular model can also be connected to other software using something called an API which is short for application programming interface. So this lets developers add chat GBT's language skills to different tools or applications. For example, some website builders now let you create a full website just by chatting with chat GPT in in prompt mode. So that's how I actually built my own company website by using a tool called B12 editor which was connected to chat GPT through these APIs. While this kind of setup is more advanced than what we'll cover in this course, it's actually good to know what's possible as you build your skills. If you're curious, open Chat GPT and click explore GPTs in the top left corner. You can browse readymade tools such as Canva and B12 as I said that connect chat GPT with other services. So just note exploring and using existing GPTs are free but creating your own requires a chat GPT plus subscription. Okay, that's a paid subscription.

[Music] Congratulations, you've made it through the course. You've learned how to build prompts from scratch using my Grasp plus Q framework, and you've applied those skills in real world scenarios. But before I wrap up, I want to give you a quick glimpse into what's next in the world of AI and how everything you've learned here will continue to serve you no matter where the technology goes. So today we've been working with large language models or LLMs like chat GPT, Claude, Gemini, etc. These tools are trained to understand and generate text. And they're great for communicating, brainstorming, writing, and solving language-based problems. But there's another class of models starting to emerge, LCMS, or large concept models. These are designed to do more than generate language. They aim to understand and reason through abstract ideas, structured data, diagrams, and relationships. In other words, they're not just about words, they're about meaning. So when we talk about concepts, we could be talking about general theories that are discussed by academics or are already known or we could be talking about the more abstract concepts such as what scientists might theorize about the universe and things like that. LCMS or large concept models are still new, but they are a sign of where AI is heading. So, what does this mean for you as a prompt engineer or AI user? Many people may not need to use large concept models and will more likely be used in research by academics and enterprise. But the good news is if you've learned to think clearly and structure your prompts with purpose like you've done in this course, you're already ahead of the game. So whether you're working with today's large language models or tomorrow's large concept models, the skills that matter most is structured thinking. So the ability to break down goals, give context, define rules, and ask better questions that doesn't go away. And that's why the grasp plus Q framework isn't just a framework for now. It's actually a mindset that grows with you as AI evolves. Where to next? Well, you have finished the course, but this is just the beginning. So, here are some recommended tools, communities, and next steps to keep building your skills. So, thanks for learning with me. If this course has actually sparked new ideas for you, I'd love to hear about them. You've now got the mindset, the tools, and the confidence to work with AI in a meaningful way. And this is just the beginning. So just remember, practice makes perfect. And good luck.

[Music]