// the find
shareAI-lab/learn-claude-code
Bash is all you need - A nano claude code–like 「agent harness」, built from 0 to 1
A 20-lesson Python tutorial that teaches you how to build the harness around an LLM agent — tools, context management, subagents, task systems, multi-agent teams. The core loop is 15 lines; every lesson adds one mechanism on top of it. Aimed at developers who want to understand how Claude Code-style systems actually work rather than use a drag-and-drop framework.
The progressive structure is genuinely good pedagogy — same loop throughout, one concept per chapter, nothing hidden behind a library. The README's philosophical framing (agency comes from training, you're building the vehicle) is correct and cuts through a lot of industry noise. The coverage is wide: permission systems, context compaction strategies, JSONL mailbox protocols, worktree isolation — things most tutorials skip entirely. Nearly 70k stars suggests the content resonates with a large audience actively looking for this kind of ground-up explanation.
The repo is mid-migration: two parallel lesson tracks (legacy 12-lesson and new 20-lesson) with mismatched numbering and a web app that still renders the old one, which will confuse anyone who arrives from a Google result or old link. The JSONL mailbox protocol is explicitly described as a teaching approximation, not a real implementation — that caveat is easy to miss and matters if you're trying to reverse-engineer how production multi-agent coordination actually works. The Python code is illustrative, not production-ready; error handling, retries, and token budget management are mostly left as exercises. There's also a promotional tail at the end pushing Kode CLI and Kode SDK from the same org, which muddies whether this is a learning resource or a funnel.