Most of us have hit that wall where you’re jumping between five AI tabs, copy-pasting context everywhere, and still doing everything one prompt at a time. It’s exhausting and honestly, it doesn’t scale. That’s exactly the problem this video tackles head-on.
A clever build just shipped, and it’s worth your attention. The creator, Grace Leung, a digital growth consultant and educator, walked through how to turn Claude Code into a fully functioning AI marketing team. We’re talking five specialized agents, 12 skills, all working together inside a single project. The twist? It’s not just about prompting better. It’s about designing an actual system where AI agents delegate, collaborate, and pick up tasks from a shared Notion board. You can even control the whole thing from your phone.
What’s new here
The core idea is treating Claude Code like a company org chart. Instead of dumping every marketing task into one conversation (which makes Claude lose focus fast), the expert breaks work into dedicated agents and shared skills. Agents are your specialized team members. Skills are the playbooks they follow. Each agent has its own role, tools, and responsibilities defined in a markdown file.
The setup covers a fictional travel brand called Go Travel, but the framework applies to any business. Five agents handle distinct functions: data analyst, content creator, market researcher, creative designer, and campaign strategist. Each one is equipped with specific skills from a library of 12.
The twist
Here’s what makes this different from typical “use AI for marketing” advice. The author doesn’t just build tools in isolation. She connects them through a CLAUDE.md file that acts as the brain, telling Claude when to use an agent versus when a simple skill is enough. So when you throw a complex campaign brief at the system, Claude automatically routes research tasks to the researcher agent, content tasks to the content creator, and design tasks to the creative designer. No manual delegation needed.
Even better, the Notion integration means your human teammates can drop tasks on a Kanban board, and AI agents pick them up based on priority. Tasks move from “To-Do” to “Complete” with output file paths attached.
Step-by-step mini-workflow
Here’s how the expert builds the whole thing from scratch:
- Set up your project structure. Create folders for context (brand voice, style guides, product info) and working output (ads, pages, presentations). Load your brand context first. This is what makes your AI team actually sound like your brand.
- Create your CLAUDE.md file. This is the custom instructions for your entire project. Let Claude scan your folder and draft it, then keep updating it over time. It’s not a one-off thing.
- Install official skills. Use the `/plugin` slash command in Claude Code, add the official skills GitHub repository, then install the document skills pack. These are prebuilt capabilities you can extend.
- Build custom skills using the reference method. Drop branded templates (like a slide deck template) into your templates folder. Have Claude analyze the template first, generating a detailed report. Then ask Claude to create a new skill based on that analysis. The expert demonstrates this with a branded deck skill and a social creative designer skill that connects to an external image generation model via MCP.
- Create specialized agents. Use the `/agents` slash command in Claude Code’s terminal. Define each agent’s role, skills, tools, and model. Claude generates an Agent markdown file that lives in your project’s Agents folder.
- Add routing rules to CLAUDE.md. This is the step most people skip, but the creator emphasizes it makes the system way more reliable. You explicitly tell Claude when to delegate to agents versus using skills directly.
- Connect a Notion task board. Set up a simple Kanban board where anyone on your team can drop tasks. Prompt Claude to scan pending tasks and assign the right agent based on priority.
- Enable remote control. Type `/remote-control` in your Claude Code chat to get a link you can open on your phone. Now you can send tasks to your AI team from anywhere.
🔧 Pro tips
- Context is everything. Loading brand voice guides, style guides, and product info into your context folder before building any skills makes a massive difference in output quality. The expert stresses this repeatedly.
- Don’t overload one conversation. The more skills you pile into a single chat, the less focused Claude becomes. That’s the whole reason for splitting work across dedicated agents.
- Use the reference method for design skills. Instead of describing what you want, show Claude examples and let it study the patterns. The branded deck skill produced slides that were 90% done, following the exact template style.
- Skills for execution, agents for synthesis. Research and campaign planning benefit from agents because they require synthesizing multiple inputs. Straightforward content creation works fine with skills alone. Defining this split in your CLAUDE.md makes routing smarter.
- Keep your remote control link private. It gives full control of your Claude Code session. When you’re done, archive the connection, and all work stays saved locally.
The full campaign demo is honestly impressive. The author throws a Japan Cherry Blossom campaign at the system, and within about 10 minutes it delivers market research, a campaign brief with the tagline “Sakura like a Local,” social posts tied to trending conversations, creative visuals, and a fully branded landing page. All connected to the same campaign theme.
If you want to see the actual outputs and watch the whole build process in real time, check out the full video. It’s one of the most practical walkthroughs on building an AI marketing system I’ve come across.