learn-claude-code Hits 59K Stars by Reverse-Engineering the Harness
A repo from shareAI-lab called learn-claude-code is at 59,000 stars and 9,700 forks, climbing 300+ stars a day. It's a 12-session curriculum that teaches you how to build Claude Code's harness from scratch — not how to use Claude Code, how to rebuild the thing that runs Claude Code.
The structure says everything. Sessions 01-02 cover basic agent loop and tool dispatch. Sessions 03-06 add planning, subagents, skill loading, context compression. Sessions 07-08 add task systems and background execution. Sessions 09-12 cover multi-agent coordination, protocols, autonomous claiming, and worktree isolation. The curriculum's stated thesis is that "agency comes from the model" through training, not from prompt plumbing — which is a direct shot at the chain-orchestration libraries the repo dismisses as "no-code platforms with extra steps."
The timing is the part to notice. Anthropic shipped Skills as a first-class concept in late April. Addy Osmani's agent-skills hit 30K stars last week. Now learn-claude-code at 59K with a curriculum specifically about building harnesses, not using them. The category that didn't exist three months ago — harness engineering — has its first textbook.
The deeper signal is who reads this. The audience description is harness engineers building agent environments for software, agriculture, hospitality, manufacturing, healthcare. Vertical agent builders are starting to treat Claude Code's harness as the reference implementation the way Linux kernel internals became the reference for OS courses. If learn-claude-code holds at 50K+ for two months, harness engineering is officially a discipline with a canonical text — and Claude Code is officially the architecture every vertical agent shop measures itself against.
https://github.com/shareAI-lab/learn-claude-code
← Back to all articles
The structure says everything. Sessions 01-02 cover basic agent loop and tool dispatch. Sessions 03-06 add planning, subagents, skill loading, context compression. Sessions 07-08 add task systems and background execution. Sessions 09-12 cover multi-agent coordination, protocols, autonomous claiming, and worktree isolation. The curriculum's stated thesis is that "agency comes from the model" through training, not from prompt plumbing — which is a direct shot at the chain-orchestration libraries the repo dismisses as "no-code platforms with extra steps."
The timing is the part to notice. Anthropic shipped Skills as a first-class concept in late April. Addy Osmani's agent-skills hit 30K stars last week. Now learn-claude-code at 59K with a curriculum specifically about building harnesses, not using them. The category that didn't exist three months ago — harness engineering — has its first textbook.
The deeper signal is who reads this. The audience description is harness engineers building agent environments for software, agriculture, hospitality, manufacturing, healthcare. Vertical agent builders are starting to treat Claude Code's harness as the reference implementation the way Linux kernel internals became the reference for OS courses. If learn-claude-code holds at 50K+ for two months, harness engineering is officially a discipline with a canonical text — and Claude Code is officially the architecture every vertical agent shop measures itself against.
https://github.com/shareAI-lab/learn-claude-code
Comments