Transcription
I use Claude more than any other AI, but even I dismissed Claude as just for developers, and I was wrong after using it myself. I am completely sold, and this is too powerful to ignore if you want to use AI to build systems and do work for you end to end, not just chatting and copy-pasting.
So in this video, I'll share how to get started with Claude Code, even if you don't use it for coding, and most importantly, how to use it to build an AI work team and system that you can actually use. Let's go.
So why is Claude Code even needed if Claude is already so powerful? So most people use Claude projects with custom instructions to turn Claude into AI agents, but they are siloed. You end up copy-pasting between them. You are the coordinator. But with Claude Code, it's different. So these Claude agents share the same workspace, access to the same context file. They can talk to each other and hand off work automatically, each with their own token context window. And Claude now becomes the team lead. You give the direction, they execute together, and that is what we are building today.
So each of these Claude Code agents you create needs at least three core components, all defined in one single markdown file: their role and responsibilities, their knowledge like workflow details, reusable Claude agent skills, and the MCP tools that they have access to. For this demo, we'll be building a five-agent work team, and Claude, our main agent, will act as our team lead to coordinate between these agents. A bonus tip is when designing your AI teammate, give them specific and non-overlapping roles to minimize potential conflicts.
Now, to get started with Claude Code, there are four main ways to access it. First, the local terminal. This is the raw method and will give you the best experience. Luckily, Anthropic just launched a native installer, so which makes it really easy to install. I'll put the official link below. And second, IDE extension. So you can add Claude Code to IDEs like VS Code or Cursor and let you run the Claude Code and see the project file structure side by side. Third, the Claude web app, where you can click the code tab and run the Claude Code in the cloud, but it requires you linking to a GitHub repository. And fourth, the Claude desktop app, which has Claude Code built in. And this is what we'll be using today.
Now, whatever way you choose, I still highly recommend installing Claude Code through the terminal first, because the terminal gives you the Because the terminal gives you the most complete experience. All the commands and features work fully. Since we are running Claude locally, so the first thing is to set up a local project folder. So think of it like a dedicated workspace for Claude Code. So let's say I run a marketing agency and I'm setting up the team system. So let's create a folder called "marketing team," and then I highly recommend to include all useful business context about this agency brand. Then access Claude Code using whatever method that you choose. So for our case, we will use the Claude desktop application. So navigate to Claude Code, pick "local" as the environment, and then pick the folder we just created.
Now, before we even create any agents, the very first step is to initiate this project by creating a Claude.md file. So this is just like a project system prompt. It gives all the important context to Claude about what it should know. So ask Claude to read the business context file and create a Claude.md file, and to build the necessary folder structure. So immediately, Claude has created all the folders for this agency brand, and we can even open up an IDE to double-check all the folders and the Claude.md file.
All right, now let's build our first teammate, the content strategist. So this is the simplest version, no Claude skills, no MCP tools, just the basic context file. A key feature of Claude Code is the slash command, so you can type slash to trigger these built-in shortcuts. And there are lots of them, and most of them are really for coding purposes, so I won't go through them one by one today. But for agent creation, we need the CLI terminal because the agent commands are not fully supported in the chat interface yet. So first, type Claude just to initiate a session. So this way you can trigger the terminal, and don't be scared. Basically, this is just a raw session for Claude Code, and a chat interface is just like a wrapper. Now, type slash agents to trigger the built-in Claude agent configuration. Then select "create new agent." Pick "project-level agent," as this agent is for this project only. Then pick "generate with Claude." So here you can describe what this agent will do. So in our case, this is a content strategist agent that will do web search and prepare content research documents in MD files. Then select the tools, default Claude model, color for this agent to use. And now we're done. We have created our first Claude agent for this work team. And whenever you want to call this agent, just tag this agent name in any Claude Code session. So we can double-check using the VS Code IDE, and you will see from the folder it has created the agent under the /agents folder. So basically, it's just an MD file containing the system prompt defining the role and the capabilities for this agent. And you can always change it.
Since this agent will also use different strategy templates, we can also ask it to create all the necessary templates that this agent will use. And this is the power of Claude Code because it not only creates the files for you, but also puts it in the right folder according to this team system structure. So no more copy-pasting. Immediately, these templates have been created under the template folder, like the campaign plan, content brief, content strategy template. So they are really detailed, and this is fantastic.
Now, to use this agent in a chat interface, just tag this agent or just use plain language to specify the agent name and ask it to research the latest SEO challenges and prepare a content brief. So you will see it will trigger different web searches and also read the content brief template file that it just created. And then after a few minutes, the content brief is ready. And this is super comprehensive, and the agent has followed everything stated in the template. So do not just create an agent, always give a context, templates, resources, standard rules to follow. So it will always produce quality output like this.
Now that you have seen how easy it is to build content agents with AI, here is the catch. Consumers have strong opinions about AI-generated content. If you want to understand what they actually think and how to implement AI without losing trust, this free report from HubSpot and SurveyMonkey, "How Brands Are Earning Loyalty in an Automated World," covers exactly that in detail. You can find the link below. Based on the data from over 15,000 consumers globally, only 24% of consumers actually like seeing AI in their emails, ads, or customer support, and 84% want to know when brands are using AI. That is a trust gap you need to close. And what I also found valuable is a section about operational best practices for staying authentic with AI. It gives you five actionable strategies to implement AI without losing the human connection. Download this for free in the description, and thank you HubSpot and SurveyMonkey for sponsoring this video.
All right, let's create our second agent, a presentation specialist. And this time, let's also add Claude agent skills. So skills are just like a reusable instruction manual that tells Claude how to execute your tasks. And particularly for this agent, we'll use the official document creation skills. So Anthropic provides a list of official skills, but they are not pre-built inside Claude Code. So you can go to this official skills GitHub repository, and the easiest and reliable way is to use the built-in plugin command inside Claude Code. So type slash plugin and register the official skill repository and then install the document skills. And also, I recommend installing the example skills to enable all other official skills as well. Then we can use this prompt to double-confirm these skills are successfully made inside Claude Code.
Now, again, use the embedded terminal inside Claude Desktop and use the agent configuration mode to create this presentation specialist agent. This agent will transform all the data into polished slide decks. Note, I also specified the brand color palettes here. So this agent will always follow these color tones when generating the slides. Then quickly, we have our second agent built. So let's say we have this performance dataset inside this data folder. Then we can just tag this agent name and then ask it to create a presentation deck for this file under our folder. And if you are using the terminal, you can just simply tag the file path, which will be much easier. Now it triggers the document creation skills that we have installed, and after a few minutes, it creates the performance report presentation for us with 10 slides. So it is quite decent and is using the color scheme that we have defined with the call-out boxes, charts, graphs for storytelling. Although it is not using very sophisticated graphics, it's definitely ready to present with minimal adjustment. And so perhaps maybe you have different datasets, documents, reports. You don't even need to upload them. You can just put them inside the folder and then let Claude to scan and load them and call this agent to build the deck for you.
All right, our third agent, which is a data analyst. So this time we'll be adding MCP tools. So MCP lets Claude connect to external systems and retrieve live data or context. And this is where your AI team becomes truly powerful. So first, we need to install all the necessary MCP servers and make them available on Claude Code. So if you have been using Claude Desktop, you probably already have a list of MCP servers connected. So the easiest way is to import them through the local terminal. So we need to use the computer's local terminal because the embedded one in the Claude Desktop does not support these import commands. So open the terminal application. So navigate to our project folder path using this CD command. Then use this command MCP add-from-claude-desktop to import the MCP servers that you want. So in this case, I have imported three MCP servers: Notion, Ahrefs, GA4. Note, depending on the MCP server that you install, some of them might require further authentication before you can use them on Claude Code. And to confirm it is already installed, go back to Claude Desktop, open the terminal, type slash MCP, and we can see all the three MCP servers are available.
Now, repeat the same process to build this data analyst agent with the agent command. So this agent will connect to GA4 through the MCP tool and create polished, interactive HTML dashboards. And again, we will use the same brand color palette to keep everything on brand. And then now we have three agents in the team. So let's test this agent to analyze our website performance and build a dashboard. So it's now calling the GA4 MCP server to retrieve the data, and quickly it has built us a dashboard with the folder path. So it is in the same brand color we have specified with the executive summary, key metrics boxes, interactive charts, trend lines, and this is what Claude really excels at: about strategic storytelling and insights. Then we can even export this directly as a PDF file, which is ready to send out to the team for quick sharing.
Now, let's level up further. So we'll create custom skills that make our agents even more brand-specific, and then we'll build our fourth agent, a social media specialist, who will use this custom skill. So let's create a custom skill called "branded social visuals," which will be used for creating branded social graphics. So upload some visual examples for Claude, and then ask it to read the official canvas design skills and extend it for visual creations with the specified brand color layout style for this agency brand. So now it triggers the official skills creator we have enabled and studies the canvas design skills and then extends it and creates the new custom skills following the official structure. And from the IDE explorer, we can also see a new Skill.md file is generated. And so using this same method, we can create a few more custom skills, perhaps like a keyword research skills that will call the Ahrefs MCP, an LLM-optimized blog writer skills that generate AI search-friendly blogs, and a custom brand voice skills. So if you want to learn how to build these custom skills, go check out this video.
So let's see this in action with the social media specialist agent, which will use this custom skill and a Notion MCP. So we can give it some topic ideas and ask it to draft the post copy with visuals and schedule them on our Notion calendar. So it is now calling the branded social visual skills we just created for visual creation and then calls the Notion MCP to upload all the post details to our calendar. And now all the visuals have been created under this folder. Although these visuals are not that fancy, it's also not Claude's strength, but definitely you can see everything is on brand using the brand color palettes and with the brand text logo. So this workflow automation is 100% working, and you can always let Claude use other image generation models, perhaps Gemini Nano, Banana, for more stunning visual images. And Claude also successfully scheduled all these posts on the Notion content calendar following our posting schedule with the platform details, posting status, post copy, hashtags. So this is what makes Claude Code so powerful. Everything can be done like in 30 minutes.
Now we've built our AI team. Four agents you have seen, plus an SEO content specialist I created the same way. So here comes the most exciting part: watching them working together. But before we do that, Claude needs to know when and who to delegate to. So remember the Claude.md file we created in the very first step. So the key here is to ask Claude to review all the agents in this workspace and then update that Claude.md file with the agent routing rules. So these are the instructions that tell Claude which agent to call for different tasks and what are the trigger phrases. And this is an important step to make the whole system work. So now it has added these routing rules that define when it should delegate to this agent, and it even added a section for the multi-agent workflow, which is amazing.
And let's test it with a multi-agent workflow. So we will ask it to do content research about a content topic and then create a blog post and also a presentation deck based on this content research findings, and that means both the blog writing and presentation tasks will depend on the first task of content research. So you can see it starts with the first task and calls the content strategist agent to do the research through the web search. So once the task has been completed, it now moves on to the next task for the blog post creation and delegates to our SEO Content Specialist. And the agent now triggers the keyword research skills that calls the Ahrefs MCP, and then it also triggers the blog writing and document skills for the actual blog content creation. Now it moves on to the client presentation task and calls the presentation agent, which will use the PowerPoint creation skills. So all the tasks have been done. And Claude is just like your team lead, giving you the summary for all the deliverables and the agents that it used. So first, the content brief about the topic, backed by research, that is really detailed and follows the predefined format, which is a great starting point. And then the blog post is pretty impressive as well. So following all the required formats to make it AI search-friendly, use of bullet points, FAQ sections. And I particularly like this section about the 10 common pricing mistakes, which is great for AI search engines. And what is amazing is I can see this blog is following the content brief done by the other agent, the content structure, talking points, and has almost 5,000 words. And you can further enhance it, perhaps also let Claude to upload it for you to your final CMS platform. And if you want any of your blog documents to match your branding and formatting, you can always create custom branded skills that use your branded templates like how we did just now. And for the deck, it is also using our brand theme. So everything is on brand, and the content is also using the same content research findings. So basically, all the deliverables are coherent and aligned to deliver the same story. And this is the true AI orchestration in action. And now we have different agents to achieve a big task and execute agentic workflows like this.
All right. There are still so many possibilities and use cases with Claude Code. If there are any other topics I should cover more, let me know in the comments. And if you enjoy this video, please also give me a thumbs up. I also invite you to join our community. Every month, we have deep-dive workshops, and you'll unlock all the prompt resources I share on this channel, including this one. I hope to see you there. And before you go, also watch this video about Claude skills and learn how to apply them in designing your Claude agents and systems. I will see you next time.