Claude Code: Build Your First AI Agent

Key Takeaways

Hey everybody, I'm Jamie. Welcome back to Teacher's Tech. So, in my last Claude Code video, we did something pretty cool. We built a full bookmark dashboard app just by typing in plain English. No coding experience needed. If you haven't seen that one, I'll link it right up here and down below in the description. Definitely go watch that one first because today's video builds on everything we covered there. But today, we're going to go way beyond building apps. Today, we're going to be building an AI agent. Let me show you what I mean. Are you seeing this? I type one sentence and it started thinking on its own. It broke the task into steps. It asked me follow-up questions. It went out to gather information, organized everything, and produced a full report, all by itself. Last time, Claude Code did what we told it to do. This time, we're going to teach it to figure out things on its own. That's called an agentic workflow, and honestly, it's the most useful thing I've learned in AI this year. Oh, and we're doing the whole thing inside VS Code this time, way more comfortable than the raw terminal like last time. Let's get into it. All right, before we start building anything, let's talk about what an agentic workflow actually is because it sounds way more complicated than it really is. Think back to the last video, we told Claude Code exactly what to build. We said, "Make me a bookmark dashboard with search, dark mode." We gave it specific instructions, and it followed them. That was powerful, but we're still doing all the thinking. We were directing every step. An agentic workflow flips that around. Let me break it down into the three levels so you can see the differences. Level one is chat. This is the most basic way people use AI. You ask a question, you get an answer, one and done. That's your standard chat GPT or Claude conversation. Most people never get past this level. Level two is building. This is what we did in the last video. You tell the AI what you want to create, you guide the process step-by-step, and it writes the code for you. You're the director, the AI is the builder. It's way more powerful than just chatting, but you're still making all the decisions. Level three is agentic, and this is where things get really interesting. Instead of telling the AI what to do step-by-step, you describe the goal. You say, "Here's what I need to accomplish," and the AI figures out the steps on its own. It makes a plan, picks its tools, adapts when things don't go as expected, and it will even ask you questions when it's not sure about something. That's what we're doing today. Here's the way I think about it. It's like the difference between micromanaging a brand new employee versus trusting an experienced one. In the last video, we were the manager saying, "Do this, then this, then this." Today, we're saying, "Here's the goal. Here's what you have access to. Figure it out." That's a completely different relationship with the AI. So, what specifically makes something agentic? Three things. First, it follows a process. It's not just answering a single question. It's working through a whole series of steps to get a result. Second, it makes decisions. It chooses what to do based on what it finds along the way. If one approach isn't working, it pivots. If it finds something unexpected, it adjust. Third, it asks before it assumes. Instead of guessing what you want, a good agentic workflow will ask you clarifying questions first. That one change alone makes the output 10 times better. And the use cases for this are endless. You could use it to research a topic across multiple sources and compile a report, analyze a folder of files and organize them by category, draft personalized emails for a whole list of contacts, compare competitors and produce a side-by-side breakdown. The pattern is always the same. You give it the goal, and it does the work. All right, let's get our tools set up now. Last time we worked entirely in the terminal, and that works, but VS Code gives us a much better experience for what we're going to do today. You can see your files, your code, and Claude Code all in the same place. When the agent is creating and editing multiple files, which it will be, having that visual file explorer makes a huge difference. So, let's get this set up. This only takes about 2 minutes. If you don't have Visual Studio Code on your computer, I'll put the link to this site down below in the description. Then you can pick what operating system you're working on. Today, I'm going to be on Windows, so I'm going to go ahead and choose this and download this for my machine. Once you have it downloaded, go ahead, open it up. Where I want you to go is right here to extensions, and you're going to type in just Claude Code up here, and you're going to see that we can install Claude Code here. So, I'm just going to go ahead and click on this. I'm going to trust the publisher and install. All right, it's all ready to go. We have it installed. I can see the Claude Code. I'm going to click on it to open. I'm going to go and mark this as done. I'm just going to go and close these extensions over here. I'll just click on them again to give myself more room. Notice you can uh stretch these uh to whatever you want that works best for your preferences. If you already installed Claude Code from the last video, you're all set. The extension uses the same installation. If you haven't installed it yet, go watch that first video because we covered the whole setup there. Let me give you a quick orientation of VS Code. Over here on the left, we've got our file explorer. This is where you're going to see every file the agent creates in real time. Folders appearing, documents being written, it's really satisfying to watch. Over here, we have our Claude Code panel. This is where you talk to the agent. You go down here, type your prompts, you see what it's thinking, and you watch it work. Now, let's create our project folder. I'm going to go ahead and create a new folder where we're going to be working out of. I'm just going to click open folder, and I haven't created one yet. If you know what folder you want, you can go ahead and choose that one. I'm going to go in my documents here, and I'm just going to create a brand new folder, and I'm going to go call it my first agent. Select folder. I'm just going to go back and click on Claude Code here and new session just to open that back up. And if I click on the explorer, you can see here's the folder that I just created. Now, there's a couple of features in VS Code that are going to be really important for you today. First, you can see the files updating in real time as the agent works. Watch the file explorer. When the agent creates something, it just appears. Second, when the agent plans a multi-step task, you'll see a to-do list pop up showing you exactly what it's going to do and where in the process. And here's the big one, planning mode. Press shift plus tab, and you'll toggle into planning mode. In this mode, Claude Code thinks and plans, but doesn't actually change any files. It's like a dry run. We're going to use this a lot today because the key to agentic workflows is planning before building. All right, our environment is ready. Now, let's set up one more thing that's really going to make all the difference. Okay, this is something I didn't cover in the first video, and it completely changes how Claude Code works for you. It's a file called claude.md, and I'm not exaggerating when I say this is the single most important thing you can set up. So, what is claude.md? It's a simple markdown file that you put in your project folder. Claude Code reads it automatically every time it starts up. Think of it like an onboarding document for a new employee. It tells the agent who you are, what the project's about, and how you want things done. Let's build one together right now. In VS Code, I'm creating a new file in the root of our project. I'll name it claude.md, all caps. And now, let's fill it in section by section. Okay, the first section is going to be project context, and this is going to tell the agent what the workspace is for. So, I am going to give it a heading right here at the beginning, and to do this, use the pound key and then a space, and then you can just type your heading. You'll notice how it turns blue. I'm just going to give it a couple spaces just so everything's easier to read and so I can take a quick glance at and see what's happening. I'm going to go and put this. So, for project context, this is my AI agent workspace. I use it for research, content creation, and productivity workflows. I'm going to give it a couple spaces. Next is the about me. This is where you tell the agent about who you are, what what you care about. So, I'll give it again another header and give it a couple spaces. And this is what I'm going to put. I create content about technology and productivity. My audience is people who want practical, no-nonsense tutorials. I prefer clear, jargon-free output. Now, the most important part, rules. This is where you set the guardrails. These are the instructions the agent will follow every single time. Let's give it a heading and a couple spaces, and my rules are going to be listed with dashes in front of them. So, you can have a clear list, just like this. So, always ask clarifying questions before starting a complex task, show your plans and steps before executing, keep reports and summaries concise, bullet points over paragraphs, save all output files to the output folder, and cite sources when doing research. And finally, project structure. This tells the agent where things go. And this is where I want things. I have my workflows, workflow instruction files, my output are the finished deliverables, and resources are the reference docs and templates. I'm just going to save this now by hitting control S together. That took us about 2 minutes, but now, every single time you open Claude Code in this folder, it already knows your preferences, your rules, and exactly how you want things organized. No repeating yourself. Let me show you why this matters so much. Now, I'm going to use the same prompt, research remote work trends. On the left, without claude.md, I made a separate folder with no claude.md file in it. So, I get this generic wall of text. You can see as it goes through, there's not really much structure. I get one question just about allowing something. Uh just kind of a data dump. But on the right, with the claude.md file, the agent's asking me many different questions as it goes through it. Then it gives me organized bullet list, clearer sections, source cited, and saved the file exactly where I told it to. Let's take a look at this end file here. Here's the link to the report and it's in the output. Remember I told it to go ahead and put it the finished output in here. So, if I click on it, it brings me to this. Now, you might be thinking, well, that doesn't look very good, but there's a shortcut to see what it looks like. So, if you go to control shift and V together when you click in here, if you press all those keys together, you get the report what it created for you. So, as I go through it, you can see how much more detailed this was here. We have our key takeaways. We have all of this information versus that one little short link that we had with the without the claw.md file. And this is why the claw.md is non-negotiable. Two minutes of setup saves you hours of frustration down the road. And here's something that was just added to Claude Code that makes this even more powerful, automatic memory. As the agent works through your project, it now automatically records things it learns and recalls them in the future sessions. So, it's just not reading your claw.md at startup anymore. It's also building its own understanding of your project over time. The more you use it in this folder, the smarter it gets about your preferences and how things are organized. Before we start building our agent, I want to give you the mental model. There are three pieces that make agentic workflows work and once you understand them, everything else clicks into place. Layer one, workflows. These are just plain English files that describe a process step-by-step. Think of them like a recipe. The agent reads the recipe and follows it, but it's smart enough to adapt if anything's off. Workflows live as a simple markdown file in your project. Nothing fancy. Layer two, the agent. This is just Claude Code itself. It reads your workflows, thinks through the steps and makes decisions. You don't program the agent. You don't write the code for it. You just give it clear instructions and let it reason. Layer three, tools. These are what the agent uses to actually get things done. Out of the box, Claude Code can read and write files, run terminal commands, and search the web. That's right, web search is built right in. No extra setup needed. The agent can go out, find current information, and bring it back into your project. Now, you can also connect external services like Gmail, Notion, and databases using something called MCP. That's model context protocol. That opens up a whole world of possibilities and we'll cover that in depth in the next video. But for today, the built-in tools are more than enough for everything we're doing. And here's the key insight that most people miss. Most people skip straight to the fancy tools and plugins, but the real power is writing good workflows. A well-written workflow with basic tools will outperform a sloppy prompt with every plugin in the world. The workflow is where the magic lives. Okay, enough theory. Let's build our first agent. All right, I'm in our empty project. You can see I just have the claw.md file that we created. I deleted those other test folders when I ran through that quick test. The first thing we're going to do is go and change it to plan mode. So, when I'm inside the chat down here, I'm just going to press shift and tab, but notice if I press it once, it goes to edit automatically. I need to press it one more time. It brings me to plan mode. If I press it again, it brings me back to where I was. So, two times pressing shift and tab brings me over to plan mode. Now, plan mode means Claude Code will think and plan, but won't change any files. We're going to design a workflow before building it. Always plan first. Now, I'm going to describe what I want and watch carefully. I'm not giving it step-by-step instructions. I'm describing my goal and this is what I'm going to say. I want to build a workflow where I can give you a topic, you'll research it thoroughly, organize the findings, and produce a clean structured report. Before you start researching, you should ask me clarifying questions about the scope, the audience, and what I specifically want to know. I'm going to submit. Look at this. I gave it a goal and it's already breaking it down into steps. It's thinking about what the workflow needs. A clarifying question phase, a research phase, a synthesis phase, an output phase. It's even suggesting quality checks. That's the agentic thinking in action. I didn't tell it these steps. It figured them out. But I want to add one more thing before we build. Before I add one more thing, let's go ahead and take a look what it has already here. We have a research plan workflow and if I scroll through, I can see there's some files that are going to be created. And remember, once we accept it to keep an eye over on explorer here cuz you're going to see it get created. It's going to go ahead and create this research report.md. As I go down, step two, confirm the plan, three, research, four, write the report, five, save the output. It gave me some formatting. So, if I like everything here, I could go ahead, yes and auto accept or yes manually approve or keep planning. But what I'm going to say is this. I also want the report to include a section at the end with key takeaways and recommendations for the next step. So, I'm just going to put that in and hit enter. So, it's going to update the plan. It doesn't just start over. It just folds the new requirement into what it already had. So, that's why plan mode is so valuable. You can shape the approach before anything gets built. So, I can see now it says right here, plan updated. The report template will now include two additional sections. Here I have recommended next steps, key takeaways. So, I'm going to go ahead and you know what? I'm going to I'm happy with it. Let's build it and I'm going to go to auto accept. So, it's just going to go through and accept everything. So, you're going to see the mode change from plan and as I click on this, just keep an eye on explorer. So, notice we have our workflows. If I go ahead and open it, it's still working, but we have the research report.md that it said it was going to create. Now, let's look at what it actually wrote. This is the important part. So, I'm going to go to the research report.md file and just double click on it and you'll see it opens up. We have the chat over here with it, but here we have the file. Now, at the top, I've got its objective or purpose here. A clear statement of what the workflow does. Given a topic from the user, ask clarifying questions, conduct structured research. Okay, so let's move down to the next. We have the qualifying question steps. Ask the user these questions and you get an idea what are specific angles of the questions you want covered, who will read the report, do you want a quick overview or deep dive? And you can see the other ones that include format preferences. Now, this is what makes it agentic rather than just a dumb text generator. It's going to check with you before it runs off and does the work. So, no more garbage output because it guessed wrong about what you wanted. So, then it moves down. We can see it confirmed the plan after receiving the answers. Confirmed the research plan with users before starting. We have step three, the research topic as we go through. And look at the details for each of these. You know, like avoid opinion only posts. For recent developments, include sources from the last 12 months where possible. What to look for while researching. Now, we're onto writing the report with the writing rules, the tone. And then saving the output. And we do have some error handling. If the topic is too broad to research meaningfully, say so and ask the user to narrow it. Now, here's the beautiful part about all this. Every single word in this file is just plain English. There's no code, no special syntax. If you wanted to change something like adding a step where it creates a summary table or removing the clarifying questions for a quick task, you just edit the text. The agent adapts to whatever you write. So, that's the power of this approach. I want to go back over to Claude Code and ask this question right here. What workflows do you have available? And look, it lists our research report workflow and describes exactly what it does. It knows what it's capable of now. So, let's put this to work. Now, for the fun part. I'm going to give it a real research task and we're going to watch the whole agentic workflow run from start to finish. I'm going to go and put this in. I want to research the current state of AI agents in 2026. What are people actually using them for? What's working? What's over hyped? And where things are heading. Now, watch. The first thing it should do is ask me questions because that's what the workflow says to do. So, there it is. Look, it's asking, you know, the audience, the scope, the depth, all these questions. Now, I need to answer it, but I can answer it naturally like I'm talking to a colleague and this is how I'm going to do it. Keep it broad. I want a hype versus reality breakdown. The audience is tech-savvy people who are curious about AI, but not necessarily developers. Medium depth. Include specific tools, platforms. Stick to the last 6 to 12 months and formatted structured posts with clear sections and bulleted points. So, let's submit. Now, it's going to take those answers and start working. Watch the to-do list. Now, here's the plan before it starts. You can see AI agents 2026, what people are actually doing, all the things I just answered. What questions I'll research are these ones and I'm just going to say proceed, yes. And notice now even on the side, we have resources, a folder that's there. I'm going to say allow, yes. So, it's searching the web, reading through results, pulling out relevant information. You can see it thinking through what it finds. And remember, this is web search built right into Claude Code. No plugins, no MCP setup, no BAI API keys. It just works out of the box. Now, I have to allow it to write. All right, it's all done. The link to the finished report is here. It's also in my output. It lives right here. It's the same thing. Now, the one thing I just wanted to point out here, this right here, the most honest number, even the best models only succeed 45.7% of the time on real-world tasks. Human oversight is an optional yet. I wonder how fast this number will change. I must say it's kind of comforting to know this, but let's go ahead and open up the report. So, remember when it opens up the MD file, this is in markdown uh language here, so it doesn't look quite as nice as the formatted report, but if we hit control, shift, and V together, it's going to give the preview of it. So, we have the topic and everything that we said said wanted in it that were from the depth of 3 to 5 pages, the date range, the audience. We have our executive summary up top here, background, uh what are they what's actually working, key findings. So, we have this detailed report now going all the way through uh really narrowing down specifically what I told it to search. So, each time I do this, it's going to be very specific. It gives me all the sources to this. So, the agent did all the thinking, organizing, and all the writing. But, we're not done yet. Let me show you one more thing that makes the agentic workflow so powerful, iteration. I'm going to say this, the executive summary is a bit long. Trim it to three bullet points. Now, watch. It doesn't start over again. It goes into the report, finds the executive summary, and trims it down. Clean, targeted, edited. And one more thing I wanted to do, add a comparison table of the top five AI agent tools mentioned in the report. So, look at that. It scans through what it already wrote, identifies the tools it mentioned, and created a formatted comparison table. It didn't need me to list the tools. It pulled that from its own research. Let's go check out those changes. So, I'm just going to open back up this. I'll go back to preview. We should see a shorter summary. There's the three, 1 2 3. And now, uh before this was six, it didn't have quite the comparison on it, but now 1 2 3 4 5 on it. So, it made those quick updates to the files. That's what I mean by agentic. It remembers the context, and it knows the project, and it builds on what's already there instead of starting from scratch each time. Once you start working this way, it's really hard to go back. All right, before we wrap up, let me save you some time with the five biggest mistakes I see people make and some pro tips that go with them. Mistake number one, skipping claw.md. I know I keep saying it, but this one file changes everything. Two minutes of setup saves you hours of frustration. Just do it. Mistake number two, being too vague. Do some research gives you generic garbage, but research remote working trends for general audience, focus on productivity data from the last 6 months, and format it as a structured report with bullet points, that gives you gold. Being specific is everything. Mistake number three, not using plan mode. If you skip planning and jump right to execution, the agent might go down the wrong path, and you've wasted your time. Plan first, build second. It takes an extra 30 seconds, but it's well worth it. Mistake number four, not telling it to ask questions. If your workflow doesn't say ask clarifying questions before starting, the agent will assume, and assumptions lead to wasted time and mediocre output. Always build that check-in step. Mistake number five, trying to build everything at once. Start with one workflow, get it working well, refine it, then build the next one. Don't try to automate your entire life in a single weekend. Trust me on this one. Now, here are some quick pro tips. Keep your workflow files in a workflow folder. It keeps things organized as you build more of them. Read the agent's to-do list and reasoning as it works. It's the best tool you have for understanding what's happening and catching mistakes early before they snowball. When the output isn't right, tell the agent specifically what to fix. Don't start over. Targeted feedback is way faster than redoing the whole thing. And save your best claw.md and workflow files. You can copy them into new projects as starting points. Over time, you'll build a library of workflows that make it incredibly efficient. One more, and this is a really useful one. If the agent goes off in the wrong direction, hover over a previous point in the thread, and you have the option to roll back the conversation. And this undoes file changes that the agent made, so you can rephrase your prompt and try again. No need to manually undo edits or start over again. It's a great safety net, especially when you're still learning how to write good prompts. And finally, effort levels. Type effort followed by low, medium, or high to control how deeply the agent thinks before responding. Low is fast and lightweight, great for simple edits. High means it really slows down, reasons carefully, and produces its best work. Use that for more complex research or multi-step task. If you're not sure, just leave it on the default, and Claude will figure it out. But, knowing this dial exists is useful once you start running longer workflows. Let's bring it all together. In the first video, we used Claude code to build an app. Today, we used it to build an agent. The first one does what you tell it, the second one figures things out on its own. That's a huge jump, and that's what you did today. So, here's my challenge for you. Build one workflow this week, just one. Pick something you do regularly, research, writing, organizing, analysis, whatever it is, and turn it into a workflow file. It doesn't have to be perfect. Just get it running and see what happens. Let me know down below in the comments what you come up with. I'm always curious to see what people are thinking about what they could automate through one of these workflows. Thanks for watching this time on Teacher's Tech. I'll see you next week with more tech tips and tutorials.