Directory: launch-agent-skills/README.md
January 29, 2026 · View on GitHub
Launch Agent Skills
Learn Claude Skills: Teach AI to Follow Your Playbooks
📹 YouTube Tutorial: Agent Skills Explained with Real Examples
🚀 X Post: Link
💻 Related Repos: launch-rag | launch-agentic-rag | launch-mcp-demo
☕️ Buy me a coffee: Cafe Latte
🤖 Discord: Join our community
🎬 Video Story: What You'll Learn
This tutorial teaches you Claude Skills - a way to create reusable AI playbooks so you never have to re-explain the same instructions again.
The Journey
Part 1: The Problem (WhatsApp Example) You export a WhatsApp chat and want to:
- Parse it into structured data
- Get a summary of key decisions
- Extract all action items
Without skills, you'd need to explain the format, edge cases, and expected output every single time.
Part 2: The Solution (Skills)
Define the knowledge once in skill.md files, and Claude references them whenever needed.
We'll build:
- WhatsApp Parser - Extract messages into structured JSON
- Chat Summarizer - Generate executive summaries with key topics
- Action Extractor - Find all TODOs and action items
Part 3: Real-World Application (RAG Backend) Then we'll use skills to set up a production-ready FastAPI backend:
- RAG Setup Skill - Clone and configure launch-rag or launch-agentic-rag
- Database Skill - Set up Supabase with pgvector for vector search
- Tools Skill - (Optional) Add Google Calendar/Gmail integration
From simple chat parsing to full AI backend - all using the same skill pattern.
🧠 Core Concepts
What Are Skills?
Skills are markdown files that teach Claude domain-specific knowledge. They follow a "progressive disclosure" pattern:
| Component | When Loaded | Purpose |
|---|---|---|
| Metadata | Always | Name, description, trigger keywords |
| Body | On demand | Detailed instructions, step-by-step guide |
| Resources | Referenced | Scripts, templates, examples |
Skills vs MCP vs Subagents
| Concept | What It Does | Analogy |
|---|---|---|
| Skills | Teaches Claude HOW to do something | A playbook or recipe |
| MCP | Connects Claude to external data/tools | A key to the library |
| Subagents | Spawns specialized workers for subtasks | A team of specialists |
Key insight:
- MCP = Access (can I read this file?)
- Skills = Knowledge (how do I parse this format?)
- Subagents = Orchestration (who handles what part?)
💡 Note: This tutorial focuses on Skills. Subagents are typically implemented at the framework level (LangGraph, CrewAI) or through Claude's natural task decomposition. Skills provide the "playbooks" that any agent (main or sub) can use.
📁 Repository Structure
launch-agent-skills/
├── CLAUDE.md # Project instructions for Claude Code
├── README.md # This file
│
├── skills/ # 🎯 The Skills (main content)
│ │
│ ├── whatsapp-parser/ # Part 1: WhatsApp Skills
│ │ └── skill.md # Parse WhatsApp chat exports
│ ├── whatsapp-summarizer/
│ │ └── skill.md # Summarize conversations
│ ├── whatsapp-action-extractor/
│ │ └── skill.md # Extract TODOs and action items
│ │
│ ├── rag-setup/ # Part 2: RAG Backend Skills
│ │ └── skill.md # FastAPI project setup (launch-rag/agentic-rag)
│ ├── rag-database/
│ │ └── skill.md # Supabase + pgvector setup
│ └── rag-tools/
│ └── skill.md # Google Calendar/Gmail (optional)
│
├── examples/ # Test data
│ ├── sample-whatsapp-chat.txt # Synthetic WhatsApp conversation
│ └── expected-output.json # Reference output
│
├── scripts/
│ ├── export_to_google_docs.py # Export analysis to Google Docs
│ ├── requirements.txt # Script dependencies
│ └── README.md # Script setup and usage
│
└── docs/
├── skill-anatomy.md # Deep dive on skill structure
└── mcp-vs-skills.md # MCP vs Skills comparison
🧪 Testing the Skills
📋 See TESTING.md for detailed step-by-step testing guide with exact prompts and expected outcomes.
Quick Start
# Clone the repo
git clone https://github.com/ShenSeanChen/launch-agent-skills.git
cd launch-agent-skills
# Open in Claude Code
claude .
Sample Test Prompts
Try these prompts in Claude Code:
Part 1: WhatsApp Skills
1. Parse WhatsApp Export
Parse the WhatsApp chat in examples/sample-whatsapp-chat.txt
Expected: Structured JSON with participants, message count, and messages array.
2. Summarize Conversation
Summarize the WhatsApp chat in examples/sample-whatsapp-chat.txt
Expected: Executive summary, key topics discussed, decisions made, participant stats.
3. Extract Action Items
Extract all action items and TODOs from examples/sample-whatsapp-chat.txt
Expected: List of action items with assignees, deadlines, and priority levels.
4. Chain All WhatsApp Skills
Parse the sample WhatsApp chat, summarize it, and extract all action items
Expected: Claude uses all 3 skills in sequence for comprehensive analysis.
Part 2: RAG Backend Skills
5. Set Up FastAPI RAG Backend
Set up a FastAPI backend for my new AI agent project using launch-rag
Expected: Claude clones repo, sets up environment, installs dependencies, configures .env.
6. Configure Supabase Database
Set up Supabase with pgvector for my RAG application
Expected: Automated CLI setup with project creation, API keys, SQL migrations.
7. Add Agent Tools (Optional)
Set up Google Calendar and Gmail tools for my agentic RAG agent
Expected: Service account setup, API enablement, credentials configuration.
Sample Data Explanation
The examples/sample-whatsapp-chat.txt contains a synthetic B2B sales conversation (no real data) between:
- Marcus Chen - Sales rep at DataFlow Analytics (SaaS analytics platform)
- Jennifer Wu - Buyer at TechRetail (VP of Sales, potential customer)
The conversation includes:
- Prospecting - LinkedIn connection, discovery of pain points
- Needs analysis - Scaling from 12→40 reps, forecast accuracy problems
- Solution pitch - Real-time analytics, AI forecasting ($3,500/month)
- Multi-threading - Looping in Raj (Head of Sales Ops) for technical buy-in
- Demo scheduling - Thursday 2pm PST with specific requirements
- Objection handling - Security concerns (SOC 2, GDPR compliance)
- Trial close - 30-day trial, board meeting alignment (Feb 20th)
- Expansion discovery - Additional needs uncovered (SDR tracking)
- Multiple action items: prepare security docs, custom demo, trial setup
Perfect for automanus.io use cases:
- B2B sales teams managing deals over WhatsApp
- Sales conversation tracking and action item extraction
- Multi-stakeholder coordination and next steps
- Discovery question analysis and pain point identification
- Demo/meeting scheduling and follow-up tracking
- Deal progression and timeline management
This gives you realistic B2B sales patterns to test all skills without any privacy concerns!
🤖 AI + Automation: How It Works Together
Claude (AI Skills) handles the smart stuff:
- 🧠 Parsing WhatsApp chats with context understanding
- 💡 Extracting action items intelligently (not just regex)
- 📊 Summarizing with business insights
- 🎯 Identifying decision makers and buying signals
- 🔍 Detecting sentiment and deal risk
Scripts (Automation) handle the integration:
- 📄 Exporting to Google Docs, Notion, PDFs
- 🔗 Syncing to CRMs (Salesforce, HubSpot)
- 📅 Creating calendar reminders
- 💬 Posting to Slack/Teams
- 🔄 Batch processing and workflows
Example workflow:
# 1. Ask Claude to analyze (AI does the thinking)
"Analyze examples/sample-whatsapp-chat.txt and save as JSON"
# 2. Export to Google Docs (automation distributes results)
python scripts/export_to_google_docs.py examples/chat_analysis.json
# 3. Share with team
# Google Doc link → ready to collaborate!
See scripts/README.md for setup instructions.
📝 Skill File Anatomy
Every skill follows this structure:
---
name: Skill Name
description: One-line description
triggers:
- keyword1
- keyword2
---
# Skill Title
## Purpose
What this skill accomplishes.
## Instructions
Step-by-step guide for Claude to follow.
## Example
Input/output examples.
## Edge Cases
How to handle unusual situations.
## Related Skills
Links to complementary skills.
How Claude Discovers Skills
- Trigger keywords in metadata match user query
- CLAUDE.md project file tells Claude which skills exist
- Claude loads the relevant skill body on demand
- If multiple skills apply, Claude chains them
🎨 Visual Guide (Diagram Ideas)
Use these ASCII diagrams as references for creating Excalidraw visuals.
Diagram 1: Skills Progressive Disclosure
Shows how Claude loads skill components on-demand:
┌─────────────────────────────────────────────────────────────┐
│ Progressive Disclosure │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────────┐ │
│ │ METADATA │ ◄── Always loaded (lightweight) │
│ │ - name │ Claude scans these to find │
│ │ - description │ relevant skills │
│ │ - triggers │ │
│ └────────┬────────┘ │
│ │ │
│ ▼ (user query matches trigger keyword) │
│ ┌─────────────────┐ │
│ │ BODY │ ◄── Loaded on demand (detailed) │
│ │ - instructions │ Full step-by-step guide │
│ │ - examples │ loaded only when needed │
│ │ - edge cases │ │
│ └────────┬────────┘ │
│ │ │
│ ▼ (if implementation needed) │
│ ┌─────────────────┐ │
│ │ RESOURCES │ ◄── Referenced (external files) │
│ │ - scripts │ Helper code, templates │
│ │ - templates │ loaded as needed │
│ └─────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
Diagram 2: MCP vs Skills vs Subagents
Visual comparison showing what each concept provides:
┌─────────────────────────────────────────────────────────────┐
│ MCP vs Skills vs Subagents │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌───────────────┐ ┌───────────────┐ ┌───────────────┐ │
│ │ MCP │ │ SKILLS │ │ SUBAGENTS │ │
│ │ (ACCESS) │ │ (KNOWLEDGE) │ │ (WORKERS) │ │
│ ├───────────────┤ ├───────────────┤ ├───────────────┤ │
│ │ │ │ │ │ │ │
│ │ "Can I read │ │ "How do I │ │ "Who does │ │
│ │ this file?" │ │ parse it?" │ │ what task?" │ │
│ │ │ │ │ │ │ │
│ ├───────────────┤ ├───────────────┤ ├───────────────┤ │
│ │ • File system │ │ • skill.md │ │ • LangGraph │ │
│ │ • Supabase │ │ files │ │ • CrewAI │ │
│ │ • Stripe API │ │ • Domain │ │ • Claude task │ │
│ │ • GitHub │ │ knowledge │ │ breakdown │ │
│ │ • Notion │ │ • Playbooks │ │ │ │
│ └───────┬───────┘ └───────┬───────┘ └───────┬───────┘ │
│ │ │ │ │
│ └──────────────────┼──────────────────┘ │
│ │ │
│ ┌──────▼──────┐ │
│ │ POWERFUL │ │
│ │ AGENT │ │
│ │ │ │
│ │ Access + │ │
│ │ Knowledge + │ │
│ │ Workers │ │
│ └─────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
Diagram 3: WhatsApp Analyzer Skill Chaining
Shows how multiple skills work together:
┌─────────────────────────────────────────────────────────────┐
│ WhatsApp Analyzer - Skill Chaining │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌───────────┐ │
│ │ User │ "Parse, summarize, and extract actions │
│ │ Query │ from my WhatsApp chat" │
│ └─────┬─────┘ │
│ │ │
│ ▼ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │
│ │ WhatsApp │──▶│ Chat │──▶│ Action │ │
│ │ Parser │ │ Summarizer │ │ Extractor │ │
│ │ Skill │ │ Skill │ │ Skill │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │
│ │ │ │ │
│ ▼ ▼ ▼ │
│ ┌───────────┐ ┌───────────┐ ┌───────────┐ │
│ │ Parsed │ │ Summary │ │ Action │ │
│ │ JSON │ │ Report │ │ Items │ │
│ │ │ │ │ │ │ │
│ │ • 42 msgs │ │ • Topics │ │ • 6 TODOs │ │
│ │ • 3 users │ │ • Decisions│ │ • 2 high │ │
│ │ • 1 media │ │ • Sentiment│ │ • owners │ │
│ └───────────┘ └───────────┘ └───────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
Diagram 4: Skill File Structure
Simple breakdown of a skill.md file:
┌─────────────────────────────────────────────────────────────┐
│ skill.md Structure │
├─────────────────────────────────────────────────────────────┤
│ │
│ --- (YAML frontmatter) ─────────────────────┐ │
│ │ name: WhatsApp Parser │ METADATA │
│ │ description: Parse chat exports │ (always │
│ │ triggers: │ loaded) │
│ │ - whatsapp │ │
│ │ - chat export │ │
│ --- ───────────────────────────────────────┘ │
│ │
│ # WhatsApp Parser ─────────────────────────┐ │
│ │ │
│ ## Purpose │ │
│ Parse WhatsApp exports into JSON. │ BODY │
│ │ (loaded │
│ ## Instructions │ on demand) │
│ 1. Read the file │ │
│ 2. Parse with regex │ │
│ 3. Extract metadata │ │
│ │ │
│ ## Example │ │
│ Input: [28/01/2026, 09:15] Alice: Hi │ │
│ Output: { "sender": "Alice", ... } │ │
│ │ │
│ ## Edge Cases │ │
│ - Multi-line messages │ │
│ - Media placeholders │ │
│ ────────────────────────────────────────────┘ │
│ │
└─────────────────────────────────────────────────────────────┘
Excalidraw Tips
When creating these in Excalidraw:
- Use boxes with rounded corners for components
- Use arrows to show flow/relationships
- Use color coding:
- 🟢 Green for Skills
- 🔵 Blue for MCP
- 🟡 Yellow for User input
- 🟣 Purple for Output
- Keep text minimal - diagrams should be glanceable
- Add icons where possible (📁 files, 🔧 tools, 👤 user)
🛠 Creating Your Own Skills
Step 1: Create the folder
mkdir -p skills/my-skill
touch skills/my-skill/skill.md
Step 2: Write the skill
---
name: My Custom Skill
description: Does something specific
triggers:
- my trigger
- related keyword
---
# My Custom Skill
## Purpose
Explain what this skill does.
## Instructions
1. First, do this
2. Then, do that
3. Finally, return this format
## Example
Input: ...
Output: ...
Step 3: Update CLAUDE.md
Add your skill to the project instructions so Claude knows it exists.
Best Practices
| Do | Don't |
|---|---|
| Be specific about formats | Leave instructions vague |
| Include input/output examples | Assume Claude will figure it out |
| Handle edge cases explicitly | Ignore error scenarios |
| One skill = one capability | Cram everything into one skill |
🔗 Related Resources
- MCP Demo: launch-mcp-demo - Learn Model Context Protocol
- RAG Tutorial: launch-rag - Build a RAG system with Supabase
- Agentic RAG: launch-agentic-rag - Add tool calling to RAG
👤 Connect
- YouTube: @SeanAIStories
- Twitter/X: @ShenSeanChen
- LinkedIn: in/shen-sean-chen
- Discord: Join our community
- GitHub: @ShenSeanChen
📄 License
MIT License - see LICENSE for details.
Built with ❤️ for the developer community
Define once, use forever. That's the power of Agent Skills.