Video ID: IT_GnOZTYWk
YouTube URL: https://www.youtube.com/watch?v=IT_GnOZTYWk
Added At: 13-06-25 21:16:37
Processed: No
Sentiment: Neutral
Categories: Education, Tech
Tags: AI, prompting, natural language processing, machine learning, techniques, engineering, strategies, communication
Summary
• The author disputes the effectiveness of traditional AI prompting formulas, highlighting that they might be holding users back. • Instead, they emphasize the importance of clear thinking and communication in crafting effective prompts. • The video also explores other myths about AI prompting, such as assigning roles, using longer prompts, and being polite to AI.
Transcript
There are lots of perfect prompt formulas out there, and I'm sure you must have heard of them before. But the more I prompt these advanced AI models and learn AI prompting, the more I realize those perfect prompt formulas might actually be holding you back. And the recent videos from Anthropic’s team and my own experience just confirmed my understanding. I've even talked with my friend who is an experienced machine learning engineer who trains AI models, and his insight also validates what I've been seeing. So in this video, I'll share the so called truth about AI prompting that you need to stop believing and the better way to actually approach AI prompting and to craft more effective prompts. The first one, assigning AI a role will always generate better results. Everyone will tell you that you must assign a role in the prompt. Act as an expert consultant, act as an experienced marketing professional. I'm not saying that assigning a role is not useful, but role prompting might not be always be as effective as you think. And from my experience, especially when prompting using those more advanced models like GPT 4 model 3.5 Sonnet, the difference is not quite noticeable. A study about the effectiveness of role prompting also shows that assigning roles will not consistently improve the response. Indeed, using a 2-shot chain of thought prompting is even better. And more interesting, in experiments to ask AI to act like an idiot versus a genius, the idiot prompt actually outperformed. Another research also reveals that the role prompting performance are often unpredictable and use them when there is a clear and logical alignment between the task and the role that is to say, you might not need a role in every prompt. Generally for now, I see assigning roles works better for tasks that require more creative thinking, or tasks that require high accuracy like legal document writing. So don't use roles as a blind shortcut to try and make AI sound smarter or more authoritative. Use roles when they generally reflect the context or provide a meaningful framework. And if you want to accelerate your AI prompting learning curve, I recommend this valuable resource from HubSpot, a library of 1,000+ AI prompts for marketing and productivity. I put it in the description for you to download for free. This library covers most of the common marketing and business scenarios from marketing strategies, brand strategies, to SEO, pay search, and even productivity. These are not perfect prompt formulas to follow blindly, but give you real world examples that you can learn from and customize, so you don't have to start from scratch. And this is why I like it. These prompts serve practical starting points to help you develop your own natural approach to working with AI. For example, the brand analysis section, AI really helps at analytical tasks like analyzing your brand positioning compared to competitors, identify gaps in your brand messaging. You can download this in the description below for free. And thank you HubSpot for sponsoring this video. The next one. There is a perfect prompt formula you should always follow, and this is absolutely misleading. While I agree that structuring your prompt can help because that makes your request easier for AI to follow your thought process, they shouldn't be the end point. We need to understand that why we have those frameworks at the first place. It's because we as human want AI to mimic the way we think. For example, the RTF framework, Role Task Format, RISEN framework, few-shot prompting, chain-of-thought prompting, whatever they are, these are All these are trying to mimic how a real human approach a problem in a real situation. So if you only follow these perfect formulas or technical frameworks or even just end there, you're limiting the creative side of these smart AI models. Just imagine how you solve a problem in reality without AI. If you always follow the same set of rules, you are limiting yourself to always solve the problem from rigid perspectives. And this is the problem. Even for those technical frameworks like chain of thought promptings, they might not be needed for simple tasks, but only for more complex reasoning tasks. Like this example, you can see following a rigid framework won't generate better results in the first place, and the response without using any framework is actually better. And I would say this is even more obvious when it comes to advanced AI models. So don't force every prompt into a rigid template with unnecessary sections. Do focus on clear communications and include only the important context that matters for a specific task. The next one, longer prompts will always generate better results. This is also incorrect. While for longer prompts, it may mean you can include more context, but it doesn't guarantee the results. in response quality will be better. In fact, study has shown that there is a notable decline in LLMs reasoning performance as the prompt length increase. This degradation is consistent across all tested models. That means when LLMs are loaded with too much information, they may struggle to identify and prioritize the most relevant details, consume more tokens, leading to inconsistent even low quality in response. In fact, you should aim to prompt AI in a way that you can achieve the best possible results with the least tokens, because the more tokens spent, it means higher costs. Even though adding more examples can usually create better performance, but it doesn't mean you need to write super long prompts. Also, there are still lots of different factors impacting the output and not just the number of examples alone. So don't add excessive context thinking more contexts must lead to better outputs. But to select a few high quality, relevant contexts or examples that illustrate your task needs. The next being super polite to AI will always lead to better results. Some people say that you need to be very polite to AI in order to get a better response. The reality is it depends. Although some research shows that politeness can affect LLM performance, it emphasized being overly polite does not guarantee better results. And from my own experience, just being polite won't significantly improve results. Indeed, they can sometimes lead to confusions to AI to understand the crux of the task. These LLMs are trained at using the RLHF method, reinforcement learning with human feedback, which means involving real humans in rating the response before they roll out. And that's why LLM's response is always being fine tuned to say what should be acceptable, including politeness and how it should respond to rude language. And that's why indeed I find it's less about being polite or not, it's about the emotions of the prompts. LLMs are being LLMs are being trained to understand human emotion language. So incorporate emotional context implies urgency and importance like using all capital letters. Study also shows that incorporating negative stimuli also have the same impact on LLM performance, like explicitly express disappointment. So I'm not telling you to be rude to AI, just treating AI with respect is generally a good practice. So don't add unnecessary polite or formal language, only add emotional language when it's necessary and matters to the task itself. So how to better approach AI promptings and craft more effective prompts? First, we must understand how AI actually works. AI doesn't truly understand context like human do. It just makes its best prediction by calculating probabilities based on the input, it generates response based on the pattern matching. So in general, the more specific and clear your prompt input, the better its prediction and results will be. And second, AI models will only get smarter, but it doesn't mean prompt engineering is dead. It is evolving from just focusing on techniques to be more strategic. It's not just the Claude model, it's also how OpenAI trains the O1 model to incorporate chain-of-thought techniques into the models to scale the performance. So we can expect the need for complex prompting will decrease over time and the models will trigger those technical prompting frameworks techniques without you even noticing. And therefore, to be able to craft better prompts, I realized it really boils down to two things. Clear thinking and clear communication. Basically, all those technical frameworks, techniques are all evolved from these two core components. Clear thinking, understand what you really want, it means to a thought planning process. Clear communications, including your certainty and uncertainty in the prompt using simple direct language. When you pair thoughtful planning with effective communications, you are creating a positive loop to make sure you will get a better response from LLM every time. First, begin with an end in mind, not a formula. Don't try to start with a prompting formula, but first get super clear on your end goal and the problem you're trying to solve first. Is it a proposal, an analysis, a report, or summary, or just simply getting ideas? You need to identify your current state and your desired outcome and what success looks like to you. And so AI is here to help you to fill that gap. If you even don't have any ideas, you can use the 5W1H method And I would say the what and the why is the most important. The what force you to pinpoint the real problem that you need to solve. The why reveals the motivations behind this problem to give AI more context. And so to give a more meaningful response. For example, you're analyzing your company sales data. So the current state is you have monthly sales data. The desired outcome is you want to understand the sales trends and success is about identifying the growth patterns to inform strategies. So to craft an effective prompts, analyze the monthly sales data is the What, and inform better sales strategy is the Why. Of course, you can also include other elements to improve context, but getting clear on your end goal and what you want AI to achieve will set the strong foundations of a good prompt. The next tip is to identify the type of tasks. Not all tasks are created equal. And it is so important to understand this in the first place to shape the best possible prompt. Basically, there are two types of tasks. Tasks that you do understand what to do, and you know exactly how to do it manually, even without AI. You're just seeking AI for executions. I call these tasks “Goal & Process Clear Tasks”. The second type is tasks you don't know how to do, but you're clear about the desired outcome. You just need more guidance from AI on problem solving. I call them “Goal Clear Tasks”. And each type of task also has a different level of complexity. So by knowing which category your request falls into, you can tailor your prompt more effectively. For example, doing keyword research for a mental health website and identify the best money keywords, you know exactly how to do it. And so you can just ask AI to analyze the data based on your criteria and ask AI to execute all the steps. However, for building a mobile application for a health website with specific features, you only know the outcome, but you don't have any ideas. And then instead, you need to frame the prompt to focus on exploration, brainstorming and strategy generation. And you can even express you are uncertain about the approach you're inviting the AI to add as your guide. And this will greatly change how you frame your request to align with your actual needs. Next is to communicate without assuming shared context or background knowledge. Besides the goal, whenever I start writing a prompt, I find it helpful to ask these two questions. What do I actually know and I have not yet mentioned? What background context it need to fully understand my request? For example, help me optimize my funnel's CTR is a bad prompt. It lacks all important context, like the why context we just mentioned. But this enhanced version will generate a much better response. I share all necessary context. It's a B2B software, it explains funnel structure, it shares current metrics, it includes what success looks like and makes the goal much clearer. Now you may wonder how much context is enough. There is no absolute rules, but generally three principles. First, include details that are not common sense to AI. This is usually about more specific details about your project or tasks. Just ask yourself, "Is this common knowledge?" If no, you better include some background in your prompt. And second, if any information in your prompt would confuse AI, just ask yourself if your friend who has no background information would get confused by this piece of information. If yes, clarify it or remove it. And lastly, start simple and add if needed. Keep iterating based on the AI response until you're satisfied and add some constraints, examples to shape the kind of response you need. So the key is more context is only better when they're relevant and you should avoid overloading. And that leads into the next, identify blind spot in your prompt. Sometimes you may not fully know you have given enough context to AI. So you can actually ask AI what it needs from you. And so to see it more as a thinking partner, just as if you're working on a project with a teammate. For example, explicitly ask in your prompt to encourage it to ask for additional Or even more direct, "What information it needs from you in order to solve the task?" these not only improve the response quality, but also trains you to think more about the information gap in your own reasoning. Another method you can use is to ask it to identify any contradictions or gaps. You can even set it in your custom instructions. So whenever it encounters any contradictions, it should highlight them before it generate the response to you. This way, you can try best to minimize the information gap and provide the most necessary context to AI for the best possible response. I even encourage you to direct AI to expose blind spots in your every prompt and do it for some time. And you will start understanding the thought process and naturally become more sensitive when a response might not be up to your standard. The next tip is to apply the 80 20 rules, 80 percent of your desired results can come from just 20 percent of your prompting efforts. So do not over engineer in the first place. Start simple. Give the AI a straightforward, decent prompt and see what you get. If you're not satisfied with the output, keep refining and iterate. Also, use concrete, clear natural language. I know they may sound too obvious, but this is one of the top mistakes people make when approaching AI prompting. AI models are being trained on human communication patterns. So you don't need to use the exact wordings, but more importantly is to understand the thinking process behind. Avoid vague instructions and aim for specifics. Instead of saying, "Give me some business ideas", try saying, "Propose three new distinct business concepts targeting Canadian parents." A tip I personally use is to confirm with the model its understanding about the task by asking "Do you understand?" Or ask it to recap again the task to you and find if there's any discrepancy. So you can make sure there's no misunderstanding before you move on. When you use natural, clear, direct language, you're working with the model in a way they were designed to interact. And with clear thinking, most of the time, it's already sufficient to generate a great response. When AI is widely adopted, one of the most valuable skills, and I'm talking about future proofing skills, is not following formulas. It's actually thinking critically. The book, "The 5 Elements of Effective Thinking" lays down the core five elements of improving thinking skills. And I found it applies perfectly in how we approach AI, prompting, all the tech stuff. Understand deeply, learn from mistake, keep asking question, understand how different ideas interconnected, embrace change. I've create a bonus videos about advanced techniques anyone can use to think smarter with AI. And I share practical methods to use AI to enhance your own thinking process. I have put it in a community. You can find the link in the description to join. And if you want more inspiration about prompting, also check out my other video about smart prompting specifically for AI Search Engines. I will see you next time!