Transcription
I learned everything about Anthropic's Claude for you. I've read their papers, watched their YouTube videos, their tutorials, and of course, for over a year now, I've been following along and playing around with Claude and using it myself. And since the release of Sonic 3.7, I can officially say that I am a true power user now. I pretty much use Claude every single day. So, in this video, I'm going to do a comprehensive walkthrough, including how to configure your settings, talk about the major strengths, Claude's most impactful features, and the really cool use cases that you can do with Claude. And of course, help you decide whether you are someone who would benefit from using Claude. As per usual, it is not enough for me just to talk about stuff, so at the end of this video, there will be a little assessment to help you retain all this information. Now, without further ado, let's get started. A portion of this video is sponsored by Brilliant.
Let's start out with settings. So, after you log into Claude, you'll be greeted with something that looks something like this. So, first thing we want to do is actually go to the settings over here. So, under this personal preferences section, I actually leave this completely blank. This is the equivalent of ChatGPT's custom instructions and the info section for Gemini, where you can give it overall context and preferences that you have. So, me personally, I actually leave this section completely blank for all of the different models that I use because I tend to use AI in like a very diverse set of ways, like all the way from coding stuff to writing things to generating pictures and brainstorming. I find that as long as I'm prompting with good prompt engineering techniques, I tend to get the best results just through that. Which, by the way, if you want to get better at prompt engineering, please do check out this video over here, which would make you better than like 99% of people in prompt engineering.
Under Claude capabilities, I have artifacts enabled. So, this is a dedicated window that shows you generated content. I don't have prompt examples turned on because the prompt examples they give you are usually super basic, but I do have CSV chat suggestions turned on, which can be quite handy for some general analysis. I also do have GitHub and Google Drive connected, so that when you're chatting with Claude, you can do something like just click your Google Drive, select the file that you want, and ask it something like, "What does this document show?" And yeah, this was my Google as Studio notes, which I just made a video about. You can check it out, linked over here. But yeah, this is, it's able to just like tell me all the things that are in my notes section. Pretty convenient to have. Similarly, with GitHub, I can go into one of my projects and pick something like this and just say, "What is happening in the code here?" And yeah, this is a workflow definition for Kestra. So, this is a project that I did before where I was building myself a little AI bot. So, yeah, it's able to understand the codebase and you can ask it questions, and it already has that context there.
And finally, for feature preview. So, some people have like different things for different features here, but in my case, I only have the analysis tool that helps you write and run code to process data and just do analysis and things like that. So, I have that turned on.
All right, let's actually get started. So, here we have all the different models that are available. So, we're going to go with Claude 3.7, which is the most intelligent model yet. This is the announcement document for Claude 3.7 Sonnet, as well as Claude Code, which is a command-line interface tool. Not going to cover in this video because I want to make it accessible to everybody, not just to coding people. But if you want me to do an AI workshop about it, please sign up for the AI workshops and also let me know in the comments.
The major impressive thing and why everybody is going nuts over Claude 2.7 is that it is exceedingly good at software engineering tasks. You can see here that Claude 3.5 Sonnet, OpenAI's GPT-4, and DeepSeek Coder are all around like 49% in terms of accuracy for their software engineering abilities, while Claude 3.7 Sonnet is already at 62.3%. With some custom scaffolding, it gets up to 70.3%. So, that is why for the examples and use cases in this video, I'm going to be concentrating more on these STEM subjects as well.
Starting off with some data visualizations. What we have here is a little table that shows when it is that you should be playing golf versus not playing golf, depending on the weather forecast. Now, we can put this into Claude and just write, "Turn this table into an infographic that is visually appealing." It's nice to put the visually appealing part because it does make it a lot more visually appealing. And we see pop up over here, and it's generating a bunch of code. And this is what artifacts is. It's generating code on a side and it's showing you this, while on the main chat, it's just telling you what it's doing. And when it's done, you also have a little preview option here. It's not bad, but I don't like how this is covering the text at the back. So, I can just say, "The red circle with temperature is covering the text in the back. Could you reformat it so that it doesn't do that?" There you go. It is now spaced correctly. Yay! What you can do here is that you can directly download this file, and you can just have the infographic as an SVG, or you can copy the actual code and put it into your favorite code editor, like VS Code or something like that, and be able to edit it or change your code or do whatever you want with it.
So, at this point, you might be like, "Oh my God, Tina, you're just getting like this really clean set of data and telling you to make it an infographic. What happens if my data is not pretty and clean like that?" I don't know why I'm using this place, but but but but but I will now show you what Claude can do even when your data is a piece of junk and it's not very clean. Now, this dataset over here, which is about anime, is not very clean. So, you have your anime, different anime here, your genre, blah blah blah, all this information, right? But as you can see, you have things like "unknown" under episodes, different formats for the date over here, and for the time, it's like different. So, some of them are in episodes, some of them is just like 52 minutes, we have 1 hour 59 minutes, or just 2 hours, different formats.
So, now let's take this not very clean data and ask it to perform an exploratory data analysis and create an interactive dashboard to help me understand and visualize this dataset better. So, it's sort of doing its analysis, doing its thing. You can definitely see that Claude is slower, for sure, compared to something like OpenAI. And we have our little artifact. So, it looks like we're doing things in React. It's telling you what it's doing. So, it says it'll help you analyze and visualize the anime dataset. First, let's explore the data structure and then create an interactive dashboard. Then, after an analysis based on the data analysis, I'll now create an interactive dashboard to help you visualize and explore the anime dataset, which is what we're doing right now. And it's currently creating the dashboard using React. This is a $2 coin. I don't know if it can see it from Hong Kong.
It's time for show and tell. Oh no, we have encountered an error running the artifact. Let's try fixing it with Claude. This is, in fact, not intentional, by the way. But I guess, depending on what happens from here, we can see how good Claude is at fixing itself. I have this visual timer here, which is basically for kids. It's because I have like no perception of time, and I'm always like late for stuff, and I never know like what I'm supposed to be doing. So, I have to like set a timer for a Pomodoro or something like that, and it buzzes when it's done. I use this a lot for like everything. I don't have any more things to show and tell. Wow, it fixed itself! Yay! Great! Excellent! It has fixed itself. So, it said it fixed the syntax error in the React component, blah blah blah. Great. Okay. And here is what we have. This looks pretty cool. It's an interactive dashboard, anime dataset explorer. So, we have the different genres over here, type filters, a TV, movie, OVA, special, and 245 in total. Data summary: total anime, average scores, average members, 91,000, as well as 41 different genres. And here's all the different anime types as a pie chart. Some top genres here as well, so comedy, action, drama. And score distribution, so primarily between 6.5 and 7, some and some unknowns as well. Rating age distribution, so rated R, unknown, R, PG-13, so primarily rated PG-13. And score versus popularity. Note: lower popularity means more popular. Okay, so this is the least popular. Okay, that makes sense. So, the higher the score is, the more popular it is. Makes sense. And score versus favorites. And top 10 highest rated: Kaguya-sama: Love Is War, Monster, Mononoke Hime, Princess Mononoke, Slam Dunk, One Piece, Hunter x Hunter, Cowboy Bebop, again, Hunter x Hunter, Hunter x Hunter, Hunter x Hunter, Initial D: First Stage. It is with utmost shame that I have to tell you that I have actually, in fact, not watched Cowboy Bebop. So, it seems like this is the universe telling me that I should do that. I have watched everything else, though. And if you actually click into it, it gives you information like Princess Mononoke, uh, gives you information about it, where it is, synopsis, and all this information as well, which is really cool. And finally, there's a popularity tab that tells you the top five, top 10 most popular. So, Naruto, One Piece, Fullmetal Alchemist, Mononoke Hime, Monster, and Fruits Basket. Fruits Basket, I'm actually surprised Fruits Basket is so popular. It's very different than the other ones. Hunter x Hunter and Fullmetal Panic. Yep, watch all of them except this one. Which of these have you watched/not watched? Maybe you don't watch anime at all. That's cool too. Okay, member versus scores. So, it looks like there's a general distribution here. And yeah, popular anime by types. So, again, you can copy the contents and actually have the code for this, and you can run it. And if you download the file, you can run it yourself.
Moving on. So, I was watching a lot of different YouTube videos about people doing really cool things using Claude, and I came across a video by AI Search. Um, and he showed some really crazy animations, um, and like simulations that he was doing with Claude 3.7. So, I also wanted to try it out. So, I created the prompt here, and I want to show you guys the results. "Write a p5.js script." So, p5.js is a JavaScript library that is for creative coding. So, it allows you to do like a lot of animations and simulations. "Simulates an ant colony searching for food, uses pheromone trails and basic, basic AI rules to show ants exploring and optimizing paths. Include controls the user can adjust in real time." So, this is what it looks like. Ant colony simulation. This is freaking crazy. It's like one prompt. Um, yeah, so you've got your ants doing ant stuff. And you can decrease your ant count, you can increase your ant count. Holy crap, that's a lot of ants. Um, you can decrease your food amount, you can increase your food amount. Pheromone strength. Not really sure what this does. Does it just like, I don't know what this is supposed to do? Or pheromone decay? Not really sure what this is supposed to do either. But any like pheromone ant colony people, if you know things about ants, let me know in the comments what these two are supposed to do. You also have wander strength. So, if you increase your wander strength, then they just start spinning in circles, which is probably not very efficient. I'll probably decrease that as well. The obstacle count toggle doesn't work over here, and add obstacle also doesn't work. But you can clear the obstacles, and you can also reset the simulation here. Let me see if I can fix it. And it still doesn't work. Anyways, you probably have to keep trying to coax it, or I mean, in my case, I probably just go in there and just fix it with code. It did like 90% of what it needs to do, so just that one function. Still very impressive.
It can also do things like make games. For example, "Create a game of memory cards. Make it visually appealing." So, it's pretty. So, memory card game. New game. Sun. Oh my God, I have a horrible memory. Mushroom. Sun. That's moon. No. Sun and Sun. Yay! Wait, why did it? So, four moves, right? I found the Sun and Sun. Okay, that didn't work. I clicked the same matching pair, but it didn't show as pair found. Okay. Turtle. Owl. Cactus. Star. Mushroom. Star. Star. Star. Nope. Still doesn't work. This is again one of those instances where knowing how to code is really helpful, because you can probably just go in there and fix it yourself. But for the sake of this, let's see if Claude can fix itself. It says that I've completely ruined a memory card game with a more robust implementation. Great. Let's see if it works. Star. Moon. Sun. Butterfly. Butterfly. Yep. Yay! It works now. It increments pairs here. Great.
All right, some more practical examples. Oh yes, so something I wanted to share with you guys, which is freaking crazy. All right, like it literally blew my mind. So, I run this program called Lonely Octopus, where people learn AI skills and they have the opportunity to apply the skills they learned on an actual project from an actual company. So, this is one of the projects. And this is a PRD, Product Requirements Document, which gives information about the projects. So, in this case, it's an AI-powered personalized nutrition plan for diabetes. Um, it gives like project overview, you know, we have level, type of project, skills required, and then like the different milestones, which is what the client wants to have. In this case, and client information as well. So, this is pretty standard. Usually, when we do like freelance projects or just projects in general, we would have this PRD so we know that we're building the right thing, right? Anyways, so the goal of this project was to develop an AI system that creates personalized nutrition plans for individuals with diabetes. So, this project should have taken a team of at least three people, two weeks to complete. And it's not like a super simple project or anything like that, right? It's like an actual project from an actual company, an actual use case. That was our estimate. But we literally just put this PRD into Claude, and it freaking outputted the entire application, literally one shot, and it was perfect within like minutes. This was one of the like key turning points for me when I realized like, holy, this is the power of Claude.
Another feature from Claude, which I think is really helpful, is Projects. The reason why you would want to use Projects is if you want to have multiple chat sessions that all relate to a single project or goal. Within each project, you can upload certain documents, provide context for everything, and then have like specific focus chat sessions that all have specific sub-goals. Let me show you an example. Example over here is the World Economic Forum's Future of Jobs Report from 2025. As you can see, it is very, very long. It is, in fact, 290 pages long. I do not want to read all of these things. But say that you needed to create some specific pieces of content, like maybe a newsletter, an article for a subset of different countries, like Malaysia, Germany, and the Philippines. If you try to do this in a normal chat and you're just asking Claude to create something like, "Create an article based on Malaysia," and then you move on to Germany, then you move on to the Philippines, what ends up happening is that because it has like the previous context that's there, the results that are coming out is going to be sort of like muddled up, and the different types of results are going to start influencing each other. Especially, for example, for Malaysia, you might want to be creating like LinkedIn posts, while for the Philippines, you are interested in creating like an article instead. So, to prevent things from getting all muddled up and confusing, you can use Projects. So, you go over here and click Projects, create project, and we can call this WEF Reports. For the overall project knowledge, you can attach the WEF report. It is uploaded, and we've already used 79% of knowledge capacity. One thing to note is that Claude really doesn't give you that much knowledge capacity, and just the number of tokens is pretty limited as well. Anyways, um, we have the report, and then we can also add text context as well. So, let's call this system prompt. "You're an economics expert on jobs and career prospects from all over the world. You also make digestible content for the general public, like you use a casual tone and concise tone to convey information that is appropriate for social media like LinkedIn and Instagram." Cool. So, all of this is going to be available, and we can start a chat here. So, you can write something like, "Generate two LinkedIn posts that are relevant to Germany. Showcase seeing the future forecast of jobs in different sectors." It is pondering. So, this is LinkedIn post number one. And just to double-check if this is actually the case, and yes, it is. 71% for AI and Machine Learning Specialists. And second LinkedIn post over here is "Germany's Green Jobs Boom and Demographic Challenge." So, great. We have Germany over here. But say we want to do something a little bit different for the Philippines. So, right here, we can write, "Write a 200-word newsletter for the job prospects in the Philippines." And for this specific chat, without influencing any of the other chats, you can attach a sample of a newsletter. So, it's able to have the tone, and you can write, "Use the tone that is in the sample provided." And there you go. This is, in fact, pretty good because I use my own newsletter to get the tone of it, and this does sound like something that I would write. And you can do this for different countries, or if you want to be even more general than that, and you can have different chats specific to different vocations, different regions, or like different industries as well. Projects is one of my favorite features. It seems really simple, but it can be really, really powerful.
Let me now show you guys the extended thinking mode for Claude. This is especially helpful for math and coding challenges. This problem is called Feynman's Restaurant Problem. So, Richard Feynman, very, very famous physicist, apparently like these are supposed to be really hard problems. So, I just chose this one randomly. I honestly don't get it at all. But let us ask Claude, "What is the final solution?" So, apparently, this answer is incorrect. It's supposed to be the square root of 2n + 1 - 1. Now, if we do the same thing and turn on the extended thinking mode, it thinks for a much longer period of time. Okay, so it is running for a really long time now. So, while it's doing its thing, I'm going to show you two more use cases that I have prepared here. I'm going to ask Claude to use Chain of Thought prompting. So, here I have some transaction data that showcases money coming in and out of a bank account, and the description of what was being bought. I'm attaching this, and then I'm going to use a prompt that leverages Chain of Thought prompting. If you don't know what that is, it's basically a prompting technique that's able to break down something into clearer steps, and it helps the AI come up with better solutions to things. I talk about it more in my prompt engineering video here. Anyways, so here is, "Brainstorm three different product recommendations you could make to this customer based on their transaction history. For each potential recommendation, provide a brief rationale explaining why you think it would be a good fit for this customer." And here you go: REI or Patagonia trail running shoes. The customer has purchased hiking boots from REI, a hiking jacket from Decathlon, and a hiking backpack from Patagonia, showing a clear interest in outdoor activities. Anime streaming services at Crunchyroll Premium. So, he's already subscribed to Crunchyroll and has purchased anime figures from Anime to Hong Kong. And a reusable coffee tumbler because it shows that there's been coffee purchases from Starbucks, including multiple visits to McDonald's. Screenshot this prompt, save this prompt. Somehow variations of this whenever you're doing any sort of brainstorming is really, really helpful.
Oh, and let's quickly talk about the True Style feature. So, you can have a lot of these different ones, and you can also create and edit your own style where you can, uh, create a custom style by putting in different writing examples. Honestly speaking, after playing around with this, like, it really doesn't do much at all. Like, if you want it to sound a certain way, you can just prompt engineer it to sound a certain way, and the custom style thing doesn't work very well either. Like, I got one of the transcripts from my own live stream, and I put it into the Claude style thing, and it honestly doesn't really sound like me. Like, I get better results if I just attach the transcript as part of the prompt instead. So, I would say, honestly, just stick with normal. Normal is good enough.
Final prompt is going to be on writing because if you remember, Claude is also supposed to be much more superior in writing. "Act as an experienced scriptwriter specializing in character development. Brainstorm a list of character names, detailed backstories, underlying motives, and compelling character arcs for the script titled 'A Tale of Four AIs.'" Blah blah blah. Multi-dimensional. Want the characters to be really cool and things like that. And let us run it. And just for comparison's sake, we're also going to run this using ChatGPT here.
"Arya: Autonomous Reasoning Intelligence Assistant. Physical manifestation is holographic interface with fluid blue-white colors that shift based upon emotional state. Voice is melodic with subtle electronic undertones. Interesting. Backstory: Created by idealistic programmer Dr. Eleanor Chen as a personal assistant AI, Arya is designed with an experimental empathy module that unexpectedly evolved beyond parameters. Interesting. And it says here, Arya began to understand mortality and loss. Years of processing thousands of human problems gave her unique insights into human nature, but left her feeling increasingly isolated and misunderstood. Blah blah blah." So, there's a lot of information here. Different character arcs, as well as key traits like empathetic and intuitive, conflict-averse, seeks harmony. So, very, very descriptive. Has secret poetry subroutine where she processes emotions. Okay.
And let's look at ChatGPT. "Character list: Nyx 013, the Dreaming AI. So, it has a role designed for creativity, art. Personalities: introspective, poetic, melancholic, and yearning. A humanoid AI with a sleek black-silver frame built for expression. Eyes glow faintly depending on mood." And it has a backstory as well. I think you can clearly tell from this that the Claude version is just a lot more nuanced. There's a lot more complexity that's put into this, like, you know, a lot of subtleties and things like that that is going on. While for ChatGPT, I think it's still good. Like, it begins the passive, obedient AI, slowly grows more defiant, right? Like, it's still definitely good, but it just simply doesn't have the complexity, um, that is from Claude's responsive. So, it does make sense why many writers have a preference for Claude over the other AIs. Let me know in the comments if you're someone that does a lot of writing and what you think about this. Honestly, me personally, I don't do that much like complex writing. Like, I would generate blog articles or, you know, like help write my script and things like that, but I don't do like highly creative, complex, like fictional writing kind of situation.
All right, honestly, there are so many other examples that which I can show you, but I swear to God, this video is going to be so freaking long, so I cannot do that. But but but I actually do want to tell you that I do live workshops where I deep dive more into different AI topics. They're completely free workshops. So, if you're interested in signing up for them, you can do so at this link over here, also linked in description.
Okay, so going back to the restaurant problem. It looks like it is finally done. So, I'm going to write, "Write the final simplified equation." So, it tells us it is this. Okay. So, the actual solution says it is this. I can pretend right now and be like, "Oh, I can figure out if these two are equivalent to each other or not," but I cannot. So, what I'm going to do is going to be like, "Is it equivalent to this?" Oh, no. Is it going to take forever again? No, it's not equivalent. However, it is an approximation that becomes valid in certain conditions. Man, I don't know if you know how to do the math. Can you tell me if the secretary optimal stopping? I feel like I learned about this in grad school, but then I forgot everything. Yeah, I don't know if anybody knows. Is it equivalent? Is it right? I'm not sure. Anyways, um, but that is the extended thinking mode, which is helpful if you're solving complex questions.
Okay, so now let's actually talk about the limitations. And Claude actually has a lot of limitations. So, the first big one is that Claude cannot search the internet. Like every single other model can search the internet at this point, but Claude cannot do that. So, it is limited by its cutoff date. Second big limitation is that it can't generate images. It can analyze images, but it cannot generate images, which is also one of those things where it's like, literally every single other model can do this now. The multimodality is definitely not Claude's strong suit. It also can't have any voice functionality as well. There's no like real-time streaming like you get with Google AI Studio or anything like that as well. And also can't analyze any videos. So, you can definitely see why for people who prioritize multimodality functions, Claude is definitely not for them.
Other limitations for Claude. Claude is extremely expensive. Like Sonnet 3.7 is so expensive, it is actually 500 times more expensive than the cheapest Gemini model. It also has really small capacity and number of tokens that you get to play around with. Even if you try to upload like a bigger PDF as context or something like that, it would just not allow you to do so. Also, remember in the very beginning when I was showing you connecting to GitHub, even just like a small GitHub project, it's not able to intake the amount of capacity there. If you tend to use AI a lot, you also very quickly hit the max limit of messages that you can send to Claude. I'm on the professional paid version already. So, for the free version, it's far worse. But even on the professional paid version, I still run into that limit all the time. So, that's why I actually have to alternate between different models. I can't just use Claude by itself for long periods of time. Finally, Claude is much slower than the other models out there.
Claude is definitely one of those models where some people really, really love, like people who are interested in reasoning skills, coding, STEM subjects, specifically, and creative writing. But if you're not specifically interested in those things, I don't actually recommend that you use Claude, or at least don't pay for Claude at this point. Maybe Anthropic will eventually add some of these other functionalities, and you can reconsider later. But for now, Claude is definitely not your general use case model. If you're going to pay, I still think that you should be paying for ChatGPT because it is a general use case model. It can pretty much do everything relatively well. But if you don't want to pay at all, I actually would recommend just sticking with Google AI Studio, which is a completely free product, and it pretty much does everything that you possibly could want it to do. Anyways, Claude is for people who are interested in specific use cases, more of like a luxury kind of situation, not for general use cases.
All right, thank you so much for watching this entire video. That was a lot. So, as promised, here is a quick little assessment. Please answer these questions and write them in the comments to make sure that you retain all the information that we covered. Let me also know in the comments, after going through all of this, whether you are going to be a user of Claude or not. Or if you're already using Claude. If you're watching this video, you're probably interested in learning AI and STEM subjects. I don't know why else you would watch this video. So, I highly recommend that you check out Brilliant, the sponsor of this portion of the video. Brilliant is a STEM learning platform that helps you get smarter every day with thousands of interactive lessons in math, science, programming, data analysis, and AI. What I love about Brilliant is that it helps you focus on developing critical thinking skills and a deep understanding of subjects as opposed to just memorizing things. Brilliant also incorporates a lot of little quizzes, analogies, and just little dopamine hits that help a lot from you getting bored or discouraged. Learning on Brilliant is shown to be six times more effective than just watching videos. They also have a great mobile app, so you can dive into new topics or do a quick practice session when you have a couple minutes instead of just mindless scrolling. Brilliant programming courses are some of my favorites. They help you build the foundations of coding and teach you how to think like an engineer, a crucial skill in the age of AI. Speaking of which, they also have great AI courses that help you gain a deep understanding of how AI models work and their applications. Brilliant's courses are honestly such high quality. You can tell so much effort and love have been put into them. To try everything Brilliant has to offer for free, you can visit brilliant.org/tahong or scan the QR code on screen, or you can click the link in the description. You also get a 20% off an annual subscription. Thank you so much, Brilliant, for sponsoring this portion of the video. Thank you so much for watching, and I will see you guys in the next video or live stream.