Disclaimer: This is an independent educational project by shareAI Lab. It is not affiliated with, endorsed by, or sponsored by Anthropic. "Claude Code" is a trademark of Anthropic.
Learn how modern AI agents work by building one from scratch.
We created this repository out of admiration for Claude Code - what we believe to be the most capable AI coding agent in the world. Initially, we attempted to reverse-engineer its design through behavioral observation and speculation. The analysis we published was riddled with inaccuracies, unfounded guesses, and technical errors. We deeply apologize to the Claude Code team and anyone who was misled by that content.
Over the past six months, through building and iterating on real agent systems, our understanding of "what makes a true AI agent" has been fundamentally reshaped. We'd like to share these insights with you. All previous speculative content has been removed and replaced with original educational material.
Works with Kode CLI, Claude Code, Cursor, and any agent supporting the Agent Skills Spec.
After completing this tutorial, you will understand:
- The Agent Loop - The surprisingly simple pattern behind all AI coding agents
- Tool Design - How to give AI models the ability to interact with the real world
- Explicit Planning - Using constraints to make AI behavior predictable
- Context Management - Keeping agent memory clean through subagent isolation
- Knowledge Injection - Loading domain expertise on-demand without retraining
Start Here
|
v
[v0: Bash Agent] -----> "One tool is enough"
| 16-50 lines
v
[v1: Basic Agent] ----> "The complete agent pattern"
| 4 tools, ~200 lines
v
[v2: Todo Agent] -----> "Make plans explicit"
| +TodoManager, ~300 lines
v
[v3: Subagent] -------> "Divide and conquer"
| +Task tool, ~450 lines
v
[v4: Skills Agent] ---> "Domain expertise on-demand"
+Skill tool, ~550 lines
Recommended approach:
- Read and run v0 first - understand the core loop
- Compare v0 and v1 - see how tools evolve
- Study v2 for planning patterns
- Explore v3 for complex task decomposition
- Master v4 for building extensible agents
# Clone the repository
git clone https://github.com/shareAI-lab/learn-claude-code
cd learn-claude-code
# Install dependencies
pip install -r requirements.txt
# Configure API key
cp .env.example .env
# Edit .env with your ANTHROPIC_API_KEY
# Run any version
python v0_bash_agent.py # Minimal (start here!)
python v1_basic_agent.py # Core agent loop
python v2_todo_agent.py # + Todo planning
python v3_subagent.py # + Subagents
python v4_skills_agent.py # + SkillsEvery coding agent is just this loop:
while True:
response = model(messages, tools)
if response.stop_reason != "tool_use":
return response.text
results = execute(response.tool_calls)
messages.append(results)That's it. The model calls tools until done. Everything else is refinement.
| Version | Lines | Tools | Core Addition | Key Insight |
|---|---|---|---|---|
| v0 | ~50 | bash | Recursive subagents | One tool is enough |
| v1 | ~200 | bash, read, write, edit | Core loop | Model as Agent |
| v2 | ~300 | +TodoWrite | Explicit planning | Constraints enable complexity |
| v3 | ~450 | +Task | Context isolation | Clean context = better results |
| v4 | ~550 | +Skill | Knowledge loading | Expertise without retraining |
learn-claude-code/
├── v0_bash_agent.py # ~50 lines: 1 tool, recursive subagents
├── v0_bash_agent_mini.py # ~16 lines: extreme compression
├── v1_basic_agent.py # ~200 lines: 4 tools, core loop
├── v2_todo_agent.py # ~300 lines: + TodoManager
├── v3_subagent.py # ~450 lines: + Task tool, agent registry
├── v4_skills_agent.py # ~550 lines: + Skill tool, SkillLoader
├── skills/ # Example skills (pdf, code-review, mcp-builder, agent-builder)
├── docs/ # Technical documentation (EN + ZH + JA)
├── articles/ # Blog-style articles (ZH)
└── tests/ # Unit and integration tests
- v0: Bash is All You Need
- v1: Model as Agent
- v2: Structured Planning
- v3: Subagent Mechanism
- v4: Skills Mechanism
See articles/ for blog-style explanations.
| Skill | Purpose |
|---|---|
| agent-builder | Meta-skill: how to build agents |
| code-review | Systematic code review methodology |
| PDF manipulation patterns | |
| mcp-builder | MCP server development |
# Use the agent-builder skill to create a new project
python skills/agent-builder/scripts/init_agent.py my-agent
# Specify complexity level
python skills/agent-builder/scripts/init_agent.py my-agent --level 0 # Minimal
python skills/agent-builder/scripts/init_agent.py my-agent --level 1 # 4 tools# Kode CLI (recommended)
kode plugins install https://github.com/shareAI-lab/shareAI-skills
# Claude Code
claude plugins install https://github.com/shareAI-lab/shareAI-skills# .env file options
ANTHROPIC_API_KEY=sk-ant-xxx # Required: Your API key
ANTHROPIC_BASE_URL=https://... # Optional: For API proxies
MODEL_ID=claude-sonnet-4-5-20250929 # Optional: Model selection| Repository | Description |
|---|---|
| Kode | Production-ready open source agent CLI |
| shareAI-skills | Production skills collection |
| Agent Skills Spec | Official specification |
The model is 80%. Code is 20%.
Modern agents like Kode and Claude Code work not because of clever engineering, but because the model is trained to be an agent. Our job is to give it tools and stay out of the way.
Contributions are welcome! Please feel free to submit issues and pull requests.
- Add new example skills in
skills/ - Improve documentation in
docs/ - Report bugs or suggest features via Issues
MIT
Model as Agent. That's the whole secret.