The learn-claude-code repository is an educational project designed to teach developers how to build agent harnesses for large language models, specifically using Claude as the foundation. Hosted at learn.shareai.run, the repository contains a comprehensive tutorial series that emphasizes a critical distinction: agency comes from model training, not from orchestration code. The project argues that building an agent product requires both a trained model and a harness, where the model provides intelligence and the harness provides the operational environment, tools, and context management that allow the model to function effectively.
The repository's core philosophy challenges the prevalent misconception that stringing together LLM API calls with conditional logic and prompt chains constitutes building an agent. Instead, it positions agent development as a two-part endeavor: either training a model through reinforcement learning and fine-tuning, or building the harness infrastructure that enables a pre-trained model to operate in a specific domain. The README extensively documents this distinction through historical examples, from DeepMind's DQN playing Atari in 2013 to modern LLM-based coding agents in 2024-2025, demonstrating that agency has consistently emerged from trained models rather than from procedural logic.
The repository is written primarily in Python and covers practical harness engineering tasks including implementing tools that give agents capabilities like file operations and shell execution, curating domain-specific knowledge, managing context and memory, controlling permissions and boundaries, and collecting trajectory data from agent interactions. The educational material uses Claude Code as a reference implementation, not because it contains secret techniques, but because it exemplifies clean harness design that avoids imposing rigid workflows and instead provides tools and context while allowing the model's own judgment to drive behavior.
GitGenius activity data reveals the repository maintains active community engagement with a median issue and pull request response latency of 22 hours and a mean of 287.1 hours across 158 tracked items. The most frequently used issue labels are discussion with 8 instances, feedback with 7, and enhancement with 5, indicating a community focused on collaborative learning and iterative improvement. The primary contributors tracked by GitGenius are Gui-Yue with 148 events, CrazyBoyM with 74 events, and Bill-Billion with 35 events, showing concentrated but distributed leadership. The repository connects to related projects including hesreallyhim/awesome-claude-code, alirezarezvani/claude-skills, and anthropics/claude-code through overlapping contributor networks.
The repository currently offers two tutorial tracks. The canonical version consists of root-level folders labeled s01 through s20, each containing full narrative READMEs, translations, runnable code.py files, and diagrams. A legacy track preserved in the docs, agents, and web directories maintains an older 12-lesson version for existing readers and the web platform. The repository is classified across multiple domains including Claude AI, Code Generation, LLM, AI Programming, Prompt Engineering, Tutorials, Code Examples, Software Development, AI Assistant, and Generative AI, reflecting its broad educational scope and relevance to the AI development ecosystem.
