NemoClaw Knowledge Wiki

Claude Code and agent team - Grace Leung

5 min read

https://www.youtube.com/watch?v=0J2_YGuNrDo Here is a comprehensive summary of the video tutorial on building an AI Work Team using Claude Code, formatted in Markdown.

Building an Autonomous AI Work Team with Claude Code

Video Summary: This tutorial demonstrates how to move beyond basic chatbots and use Claude Code to build an orchestrated system of AI agents. Instead of copying and pasting between chats, these agents share a workspace, context, and files, allowing them to collaborate on complex tasks.

🧠 Core Concept: The AI Team Structure

Unlike standard Claude chats, Claude Code allows you to create a local workspace where multiple agents operate under a “Team Lead.”

  • The Workspace: A local folder on your computer containing all project files.
  • The Agents: Specialized sub-personalities defined by Markdown files.
  • The Team Lead: The main Claude interface that routes tasks to the correct sub-agent.
  • Shared Context: All agents have access to the same files, templates, and data within the folder.

🛠️ Prerequisites & Setup

1. Accessing Claude Code

While available via Web, IDE extensions, and Desktop App, the Local Terminal is recommended for the most complete feature set.

  • Recommended Setup: Install Claude Code via terminal, then use the Claude Desktop App to interface with the local folder.

2. Initializing the Project

  1. Create a local folder (e.g., marketing-team).
  2. Add a Business_Context.docx (context about your brand/company).
  3. The System Prompt (**CLAUDE.md**): Ask Claude to read your business context and generate a CLAUDE.md file. This file acts as the “brain” of the project, defining the folder structure and rules.

🏗️ Step-by-Step: Building the Agents

The tutorial builds four distinct agents, each adding a layer of complexity.

Agent 1: The Content Strategist (Basic Agent)

  • Function: Research topics and create strategy documents.
  • How it works:
    • Use the command /agents in the terminal to create a new agent.
    • Define the role via natural language prompts.
    • Claude creates a definition file (e.g., content-strategist.md).
    • Key Feature: Ask the agent to generate its own reusable templates (briefs, campaign plans) and save them to the templates/ folder.

Agent 2: The Presentation Specialist (Adding Official Skills)

  • Function: Turn text/data into PowerPoint slides (.pptx).
  • New Concept: Skills. Reusable instruction manuals that give Claude new capabilities.
  • Setup:
    • Use /plugin command to install official Anthropic skills (specifically document-skills or pptx creation).
    • Workflow: Tag the agent (@agent-presentation-specialist) and ask it to convert a specific data file into a slide deck using specific brand colors.

Agent 3: The Data Analyst (Adding MCP Tools)

  • Function: Connect to external live data (Google Analytics 4, Notion, etc.).
  • New Concept: MCP (Model Context Protocol). Allows Claude to connect to external systems.
  • Setup:
    • Use terminal command claude mcp add-from-claude-desktop to import configured tools.
    • Workflow: The agent queries live GA4 data and generates an HTML dashboard with interactive charts, stored locally.

Agent 4: The Social Media Specialist (Adding Custom Skills)

  • Function: Create social posts with on-brand visuals.
  • New Concept: Custom Skills. Creating your own specific “tools” for Claude.
  • Setup:
    • Feed Claude examples of your brand style (hex codes, fonts, layout).
    • Ask Claude to create a branded-social-visual skill based on the official Canvas design skill.
    • Workflow: The agent generates post copy and uses the custom skill to generate images that strictly follow brand guidelines, then uploads them to a Notion calendar via MCP.

🤖 The Final Step: Orchestration

To make the agents work together without manual intervention for every step, you must update the routing rules.

  1. Update **CLAUDE.md**: Ask Claude to review all created sub-agents (.md files in the agent folder).
  2. Define Routing: Have Claude update CLAUDE.md with specific rules on when to call which agent (e.g., “If the user asks for slides, call the Presentation Specialist”).
  3. Multi-Agent Workflow: Define sequences (e.g., Research → Blog Writing → Presentation Creation).

The Result: “Mega-Prompting”

You can now give a single complex command:

“Help do content research for ‘Small Business Pricing’, then create a blog post, and finally create a presentation deck based on the findings.”

The Chain Reaction:

  1. Main Claude analyzes the request.
  2. Delegates to Content Strategist (Web search + Research doc).
  3. Delegates to SEO Writer (Writes blog based on research).
  4. Delegates to Presentation Specialist (Creates slides based on the blog/research).
  5. Main Claude presents the final summary.

💡 Key Takeaways & Tips

  • Non-Overlapping Roles: When designing agents, give them specific, distinct responsibilities to avoid conflict.
  • Don’t Copy/Paste: Let the agents read and write directly to the local file system.
  • Custom Skills are Powerful: You can code specific brand guidelines into a “skill” so the AI never “hallucinates” the wrong colors or fonts.
  • Data Privacy: Consumers are wary of AI. Using local LLMs and transparent workflows (as highlighted in the HubSpot report mentioned) helps build trust.

Graph View