How to Build Your First AI Agent (No Code Required)

Key Takeaways

Most people using AI right now are doing the equivalent of hiring a genius and then only asking them to proofread emails. I know this because this is the use case of most of the people that I talk to in my everyday life. They open chat GPT, they type a question, they get an answer, then they close the tab. Is it useful? Of course, but they're leaving 95% of the value of AI on the table. When chatbt first dropped in 2023, it was this cool, novel, kind of optional, fun tool, and everyone was using it as such. >> I am Harvey, a computer. Jim sucks. [snorts] [laughter] >> Now, uh, jobs are being taken away, entire industries are being displaced, and the vibe is MORE LIKE, >> OH MY GOD. OKAY, IT'S HAPPENING. EVERYBODY STAY CALM. >> WHAT'S THE PROCEDURE, EVERYONE? WHAT'S THE PROCEDURE? >> STAY CALM. SO, I want to help out. In the past few weeks, I've built a game I play with my kids in line at Disney World. I ran a competitive SEO analysis for a client with real data, actual screenshots, and I built it in about 10 minutes. I built a video editing tool that cuts the silences and mistakes out of long- form video content, saving me thousands in editing cost. And I have a personalized virtual assistant that helps me prioritize and execute tasks across my life and business. I don't know how to code. Not a single line. All of this was done by my Claude Code AI agent, Gus, who by the [music] way gave himself that name and even designed his own profile picture. Now, before I built this agent, I was one guy running a consultancy. I guess I still am doing the work of one very tired person. Client research took hours. Reports took longer. Half the tasks on my to-do list sat there because there just weren't enough hours in the day. Now, I genuinely feel like I operate at the level of a 10 to 20 person agency. It's the same business. It's still me, but the output is completely different. And if you follow this video from start to finish, I'm confident this will change your life in business, too. This is the guide I wish existed when I started. If you want to skip straight to the buildout, go to this timestamp. Otherwise, stick around because most people don't actually know what an AI agent is or why it's so much more powerful than what they're currently using. So, what is an AI agent? Here's the simplest way I can explain this. Think about the last time you used Chat GPT. You typed a question, it answered. You typed another question, it answered. It's back and forth. It's like texting a really smart friend. That's an AI powered chatbot. Now, chat bots are useful, but there is a ceiling. Chat GBT doesn't really know anything about you beyond what you tell it in that conversation or fragments of memory in previous conversations. It can't access your files, your tools, your project management system. And the moment you close the tab, it forgets most of what you said. Now, imagine something different. Imagine hiring an actual employee. [music] On day one, you hand them the employee handbook. You tell them about the company, the clients, the goals, how things work around here. You give them login credentials. Here's notion. Here's the analytics dashboard. Here's the CRM. And then you turn them loose. And that is what we're building today. Not a chatbot, but an employee. Now, to me, an agent needs four things to truly be considered an agent. First, it needs a brain. That's the AI model. In our case, Claude. Anthropic Opus 4.6 is one of the most capable models available, at least for the next 10 minutes, based on how fast AI is moving. And they just released Sonnet 4.6 right before I recorded this, which is nearly as powerful at a fraction of the cost. So, there's really never been a better time to start. Chat GPT does have a brain, too. That part isn't unique. Now, second, it needs memory. I'm talking about deep structured knowledge about you and your world. Not the scattered fragments that ChatGpt picks up from conversations, but a real briefing. It covers your business, your clients, your goals, and your workflows. And it lays out things to always do, things to never do. It defines how you want the agent to speak and think and prioritize. Jack GBT sort of has this through its memory feature, but it's extremely fragmented. It's nowhere as deep as what's possible with an agent. So, we're going to build that document later in this video. Now, if you only do that one thing, you'll be ahead of most people using AI right now. Okay. So, we've covered the knowledge layer, the Opus 4.6 large language model under the hood. We've covered the memory layer, which is what we're going to build in just a moment. Now, we're going to get to my favorite, and that is the abilities layer. Basically, we're equipping it with tools, software, essentially giving it superpowers. Project management, CRM, analytics, web scrapers, whatever it needs to do real work. Without tools, your brilliant employee is basically just sitting at an empty desk thinking really hard. Now, ChatGpt does have a handful of plugins, but Claude Code has a completely open plug-in system. You can connect it to notion, to analytics platforms, to web scraping tools. You may have heard the term MCP. That's something we're going to be exploring in just a moment as well. You can connect it to pretty much whatever you need, and we'll set those up later. Now, when people talk about agents, I noticed they kind of stop at those three. They talk about the knowledge, the large language model. They talk about the memory and they talk about the abilities. But in my opinion, the fourth ingredient is the one that matters the most. And that's autonomy. This is what separates an agent from a chatbot and an agent from an automation. Let me show you what I mean. Last week, I told my agent, "Go do keyword research for such and such brand." That's all I said. First, it checked my knowledge base. There was no info on this brand. it was a new client or a new potential client. So instead of asking me for context, it crawled their website, pulled their industry, their competitors, their target keywords, and it built its own briefing before doing anything else. Then it said, "Okay, now I need to do keyword research." So it picked the most applicable tool, in this case, the HF's MCP, which is one of my favorite keyword research plugins. And it ran the query. It got the data, but then I asked it to pull what people were saying about the brand on social media. Now, it tried to scrape Reddit, but Reddit blocked it. Now, here's the part that got me. It didn't stop. It didn't come back and say, "Oh, sorry. I can't do that." It looked at its other tools and said, "Okay, my web scraper got blocked, but I have Playright, which opens a real browser. Let me try that instead." It ended up going to an old Reddit site, and it somehow got the information it needed. I didn't tell it to try Playright. It just figured it out on its own. Now that reasoning, the check context, pick a tool, try it, fail, adapt, try a different approach. That's the difference between an automation and an agent. An automation follows a flowchart. If X [music] happens, do Y. If Y fails, then stop. An agent actually thinks. Now, tools like NAN, you may have heard of this, gained a lot of popularity for letting people build agentic workflows. I even made a video about this on my last video in 2025 about AI agents. And they are genuinely powerful. But at the end of the day, you're still programming visual programming instead of textbased code. You're dragging nodes. You're connecting APIs. You're handling error states. For a non-technical person, it's a lot easier than programming, but the learning curve is real. Claude Code collapsed all of that into plain English. You don't drag nodes, you don't program workflows, you describe what you want, and the agent figures out how to do it. ChatGpt is a chatbot. Claude Code is an agent. And that's the distinction that matters. Brain, memory, tools, and autonomy. Those are the four ingredients. Now, let's build yours. All right, let's talk setup. I know this is the part where eyes start to glaze over. So, three steps, that's it, and you'll never have to do them again. Step one is to go to claw.ai and create an account. You'll need at least the pro plan, which is $20 a month if you want the full experience with the most powerful model and fewer limits. Max is 100. I use max, but pro works fine to start. And if you're not sure you want to pay yet, that's totally fine. Watch the rest of this video, see what's possible, and then come back to the step when you're ready. Okay, step one done. Step two is to download Visual Studio Code. It's totally free, and I know this is the second product in this video about no code that has the word code in it. I promise you, I promise you there is no coding involved. Go to code.visisualstudio.com, pick your operating system, install it, bada bing, you're good. VS Code is going to be the container that we run Claude inside of. Claude is like the employee. VS Code is like the office. You get to see your files, your folders, your conversations all in one place. It's just a lot more userfriendly than staring at a terminal. Step three is to open VS Code. Click the extensions icon in the left sidebar. It looks like four little squares. And search Claude Code. You'll see the official extension from Anthropic. Click to install. Now, while you're there, grab one more thing. Search for a rich markdown editor. You don't have to worry about what that means for now, but I use Office Viewer. So markdown is the format that your agent writes in. And this just makes those documents look clean and readable, and you can edit [music] it more like how you would edit a document in Microsoft Word or Google Docs. You'll definitely appreciate it as you're staring at markdown files for the next several hours. Well, unless you don't want to do that, but I think you're going to get addicted. Okay, so that's it. Sign in with your Anthropic account. Claude Code is now running inside VS Code. [music] To actually talk to your agent, look for the star icon in the top right corner of VS Code. Click that and it opens a fresh Claude Code chat. That's your line of communication from now on. You've now got your brilliant employee sitting at their desk on day one. They're ready to work, but right now they don't know anything about you or your business or what you need. That's what we fixed next. This is the most important part of this entire video. You know how chatbt gives you that generic, overly polished, kind of hollow output like it's writing a college essay for a stranger? That's because it doesn't know you. It's guessing [music] every time. It's always moving to the mean, which is actually a really insightful thing that Ben Affleck said about AI on Joe Rogan. I didn't have Ben Affleck having pretty smart takes about AI on my 2025 bingo card. So, life is interesting. Now, what we're about to do is give your agent permanent knowledge of who you are, how you work, and [music] what you need. So, every conversation starts from understanding, not a blank slate. In tools like NAD, you'd store this kind of context in a database node or a Google sheet and wire it into each workflow. Here, you just write a document or speak a document, which I'll I'll talk about in just a moment. Here's how. I've put a prompt in the video description. Copy and paste it into Claude code. Claude is going to interview you. It'll ask you about your work, your goals, your pain points, how you like to communicate. Answer honestly [music] in plain English. You don't need jargon or formatting. Talk like you're explaining your job to a smart friend over dinner. And one quick tip before we start. I highly recommend voice dictation for this part. Now, I use Whisper Flow. I'm not sponsored. I'm just obsessed. It's extremely accurate. It uses AI for context. It formats well, and it cuts out filler words. More importantly, when I speak out loud, I'm more expressive. I add nuance that I'd normally skip when typing. And that's exactly the mindset we want when talking to our agent. If you want to give Whisper Flow a try, you can check it out via the link in the description. So, whether you're typing or using dictation, paste the prompt into Claude Code and answer the questions in as much detail as possible. Now, when Claude asks, "What eats your time? What do you dread? What falls through the cracks?" Be specific and don't think to yourself, now this is a very limiting belief. People go into this interview and they think, "Oh, AI could never do [music] this. AI could never help me with this. This is this is something that AI could never really do and they categorically store it away as something that's not even possible." That's a huge mistake. Tell [music] them everything. Even if it's a total moonshot, something that you think AI could never help with, you really never know. Also, don't say things like, "I'm too busy." Be specific. Say, "I spend 3 hours every week pulling competitor data into a spreadsheet, and I hate every minute of it." It's the same with goals. Don't say, "I want to grow my business." Say, "I want to add five new clients by June." [music] Or, "I want to launch a YouTube channel, but I can't keep up with the editing." The more specific you are, the more useful your agent becomes. Now, once you're done, Claude will take everything you told it and turn it into a structured document called a Claude. [music] MD file. That's Markdown. This is your agent's permanent briefing. Every time you start a new conversation, Claude reads this first before you say a word. These files are stored locally on your computer. You own them. There are ways to back them up to the cloud, but that's a topic for a later video. And if you ever want to move your knowledge base somewhere else on your machine, you can just ask Claude to do it. Whenever you close and reopen VS Code, just fire up a terminal and type Claude or click the orange star icon in the top right corner, and you're right back where you left off. [music] The great thing about building out your knowledge base with markdown files is that it's the universal language of large language models. The reason this is important is for whatever reason if you decided to move away from cla code or move away from cloud in general because large language models and providers are always changing you essentially have a model agnostic knowledge base. So you could move to ChachiPT. You could move to a completely open-source model that you run on your computers at home, which is another huge benefit to building out this kind of structure over and above the built-in memory of something like Chat GPT. Now, here's something important. [music] If you cram everything into one claw.md, every client, every project, every rule, things will start to degrade. The model gets noisy. It's kind of like handing your new hire a 300page handbook and saying, "Okay, memorize all this before every conversation." Now, what my agent did and what yours will automatically do [music] is it will likely break up your knowledge base into multiple files. One for each client, one for my YouTube channel, one for my website, each one living in its own folder. So, when I'm working on client work, Claude reads that client's file. When I'm building a game for my kids, it's not thinking about keyword research because that would be a huge waste of tokens and context. [music] The masterclaw.md at the top ties everything together, kind of like a CEO delegating to different departments. Each department has its [music] own context and rules, so read through yours. Change anything that doesn't sound right. Add things you forgot. This is a living document. You'll keep refining it over time. By the way, I put together a free starter kit that has everything we're covering in this video in one place. The exact interview prompt, a Claude MD template you can copy, my five recommended skills with install commands, a plug-in decision tree, and the five things I'd automate first. It's called the AI agent starter kit. You can grab it for free. The link's in the description. Just pop in your email and it's yours. Your agent has a brain and memory. [music] It knows who you are and what you need. Now, we give it superpowers. Start by saying this to Claude. Now that you know about me, what are the top five things you could help me with this week based on everything I just told you? Please think deeply, do deep research, and recommend any particular tools, skills, MCPs, or software I can connect to you to make my life easier. If it suggests tools or plugins, just tell it to install them. It will either do it autonomously or walk you through the process if the tool needs authorization. Most of mine took less than 3 minutes to set up. Now, here's a tip that makes this step exponentially easier. If you already have a task management system, Notion, To-Doist, Obsidian, Tick, Tick, whatever you use, connect your agent to it. I use Tick Tick, and once I connected Claude to my task lists, it could survey everything on my plate and say things like, "Okay, looks like you can knock this out right now. Uh, how about you go work on this, Matt, and I'll go work on this. Uh, let's move this to next week. Here's what I'd prioritize." guys and I'm sitting back like, "Yo, dude, this is sick." It's basically like having a brilliant, cheap virtual assistant. So, what kind of abilities can you give your agent? Well, first things Claude Code can already do right out of the box. No setup is required. It can read and write files on your computer, search the web, and run commands in your terminal. It can organize folders, analyze a spreadsheet, rename hundreds of files in second. Second, MCPS. Now, forget the acronym for a second. Just think of them as plugins. Each one connects your agent to a piece of software it can actually use. Lots of modern software platforms offer MCPs. So, whatever you use in your typical work stack, you can just Google my tool MCP or even ask Claude, does this tool that I use every day have an MCP? It will go out, [music] it will find it for you, and it will help you get set up. It's essentially like telling your new hire, "Here's your notion access. Here's the analytics dashboard. Here's the CRM." Without those loginins, they can't really do their job, but with them, they're [music] absolute weapons. Here are a few of my favorites. Notion, so Claude can read and update my project boards. Tick, tick, so it can see my tasks and help me prioritize. Hrefs, so it can pull real keyword data. Playright, so it can open a browser and interact with websites the way a person would. and Nano Banana Pro, which can generate highquality images for thumbnails on the fly. Now, you don't need any of these on day one. Start with zero plugins. Get comfortable. Then, the first time you think, "I wish Claude could send a Slack message for me, or I wish it could pull data from Salesforce." That's the signal to add one. I'll cover the best plugins in a future video. Now, third, and this is the one that got the internet really excited a few weeks back, are skills. These are custom workflows you teach your agent. Think of them like SOPs, standard operating procedures. I have a skill that takes raw data and formats it as a beautiful polished report I can send to clients. When I just type slashreport generator, Claude loads that entire workflow and follows it step by step. You can download skills from the internet or build skills for anything repeatable. client onboarding, weekly reports, social media posts, anything where you find yourself explaining the same process more than once. Turn it into a skill and never explain it again. Here's a pro tip. After about a week or two of consistent use, ask your agent, "Based on our work together so far, what skills do you recommend we create?" It will then recommend them and build them for you right there on the fly. Your agent is built. It knows who you are. It has tools to help. Let's put it to work. Go back to your interview, specifically your pain points. You told Claude the things that eat your time, the tasks you dread, the work that feels tedious. Don't overthink the prompt. Describe what you need like you're talking to a capable colleague. If the output isn't perfect, tell Claude what to fix. Make it shorter. Add a section on pricing. Download whatever tools you need to make this happen. It will adjust and the output will be all the better for it. Okay, here are some tips to get the most out of your agent. These are things that I've discovered through weeks of stress testing my agent. Every one of them made a noticeable difference. Tip number one is to think atomically. Now, this might be the most important tip in the whole video, which is why I'm putting it first. Most people give up on AI because they hand it tasks that are too big or too vague. The fix is learning how to break things down. Say you need to present a deck on your department's performance to the seauite next Tuesday. Don't just say, "Make me a presentation." Think about the pieces. You need to gather the data. You need to assemble that data in a way that tells a story. You need to build a deck that's concise, engaging, and on brand. And that often requires gathering screenshots or gathering imagery. You need to practice and present that to stakeholders. Now, Claude will sometimes surprise you and just oneshot all of this, and it's amazing. But the output is more reliable and frankly usually better when you break it up. Ask it to gather the data and save it to a file. Then feed it past email conversations with your bosses and ask it to figure out what exactly they want to see. Then give it your brand guidelines and ask it to assemble the deck. You can even spin up three separate instances of cloud code and have them work on these in parallel. The point, don't assume AI can't help me with this just because the whole project feels too complex. In every worthwhile project, there are subtasks better suited to humans and subtasks better suited to robots. The people who can break things down and delegate accordingly are the ones who will get the most out of AI. Tip number two is to ask your agent to improve itself. This one sounds too simple to work. Once you've been using your agent for a few days, simply ask it, "Look at my setup, my claw.MD, my folder structure, my tools. What could be better? What am I not using you for that I should be?" I recently followed a prompt series from Maxwell Finn on X, and my entire setup leveled up overnight. I ended up with super organized rules, structured contexts, and sub agents all working in the background. I'll put Maxwell's prompt series in the description. Now, your agent can audit its own setup and find the gaps you can't see. Once you're comfortable with the basics, push it. You'll be surprised what it comes back with. Tip number three is to not waste your saved time. When you first start using Clawed Code, it feels like magic. You'll be tempted to sit there just watching it work, or worse, pull out your phone and start scrolling. But the real power of an AI agent is that it frees you up [music] for more meaningful work. I'll have an agent running on one monitor while I'm working on something else on my main screen, or I'll just open another instance of Clawude Code and have it start working on a completely different project. Recently, I had four instances all doing research and working on separate projects while I went downstairs to play two square with my son. There's no limit on how many instances you can run at once. The only limit is your token usage, which brings me to the next tip. Tip number four is to switch models [music] strategically. Now, the biggest complaint about Claude Code is its usage limits, and that's legitimate. If you're running the most powerful model on a $20 plan, you will hit the cap, especially during an intense session. [music] Anthropic just released Sonnet 4.6, which is a lower tier, but extremely capable and significantly cheaper model. [music] At any time, you can type /model, and switch to Sonnet for routine work like retrieving data, scraping the web, and organizing files. Save Opus for the heavy lifting. Research, brainstorming, complex analysis, and creative work. If you're not sure which to use, just ask your agent. Hey, what model should I be using for this task? [music] And it will tell you. Tip number five is to use skip permissions mode. Now, by default, Claude asks permission before doing anything. Accessing software, running commands, deleting files. That's a security feature, and I get it. But sometimes, guys, sometimes I just want my agent to be autonomous. I just want them to go. I'm not pushing code to the production server of a multi-million dollar SAS brand. I'm one dude running a consultancy. Open a terminal in VS Code and type claude- dash dangerously-sk permissions. [music] Now it's truly autonomous. It just keeps working until it finds a solution. And I know you may be thinking, Matt, are you trying to reenact the plot of Terminator 2? because it kind of feels like you're out here trying to reenact the plot of Terminator 2. It's really not that serious. Claude will only work on things you assign to it. Once you're comfortable, give it a shot. A year ago, building an AI agent meant dragging nodes and programming workflows. Today, you built one by having a conversation. It has a brain, one of the smartest models available. It has memory, a deep understanding of who you are, what you do, and what you need. And it has tools to let you do real work outside a chat window. This is the foundation. But the best part is that it compounds. The more you use your agent, the better it gets at helping you. The more skills you build, the more powerful it becomes. The more specific your knowledge base, the less time you spend explaining context. [music] So, drop a comment. What do you want to see me build next? What workflow? What tool? What problem can I help you solve? I read every single one and it directly shapes what videos I make next. If you haven't already and you've gotten value out of this video, I would really appreciate a like and a [music] subscribe and a share with a friend. This is honestly just the beginning. Links to all prompts and tools are in the description. Thanks for watching and go build