shareai-lab/learn-claude-code

Learn Claude Code is an educational repository that teaches you how to build the infrastructure layer that lets an AI language model operate as a practical coding agent. The project makes a foundational distinction: the intelligence in an AI agent comes from the model itself, which was shaped during training, while the surrounding code is the harness, meaning the environment, tools, and interfaces that allow the model to take real-world actions. This repository focuses on teaching harness engineering rather than model training. The central concept is that a harness consists of five components: tools (things like file reading, shell commands, and network calls), knowledge (documentation and context the agent can reference), observations (information the agent can perceive about its environment, such as error logs or git diffs), action interfaces (the ways the agent can affect the world), and permissions (safety boundaries and approval mechanisms). The repository builds a minimal clone of Claude Code, Anthropic's AI coding assistant, from the ground up to illustrate how these components fit together. It uses TypeScript as its primary language, and the README contains extensive philosophical framing about the history of AI agents, explaining why genuine agency is a product of trained neural networks rather than complex scripting logic. A recurring argument is that simple prompt-chaining systems dressed up as agents do not represent true agentic behavior. The project is described as going from zero to one, meaning it walks through building the harness incrementally. You would use this repository when you want to understand the architecture behind tools like Claude Code from first principles, or when you are building your own agent system and want a solid conceptual foundation before writing code.