Claude Code 最佳实践

★ 64kAgentic Engineering

#claude-code#best-practice#agentic

Overview

One of the most authoritative AI coding repos (64k+ stars), built around 'from vibe coding to agentic engineering'. It takes Claude Code from casually generating code to engineering-grade agent orchestration — with runnable implementation examples you can copy into your own projects.

Who it's for

Developers already using Claude Code who want AI to reliably participate in production code. Beginners can follow along, but the value compounds when you return to compare after real usage.

What you'll learn

01Core idea: from Vibe Coding to Agentic Engineering
The repo splits AI coding into two levels: Vibe Coding is 'throw requirements at the AI and accept its output' — fast but fragile; Agentic Engineering is 'you act as architect, decompose tasks, manage context, review every diff' — reliable and shippable. Claude Code is the best vehicle because it can actually read/write files, run commands and self-test in the terminal.
02The trio: Subagents / Commands / Skills
The repo systematically covers Claude Code's three extension points: Subagents (.claude/agents/<name>.md — dedicated 'expert personas' per task), Commands (.claude/commands/<name>.md — solidify recurring flows into one-shot prompts) and Skills (reuse Claude Code skill packs). Each comes with best practices and ready-to-run implementations.
03Best practice + implementation: practice beats theory
Every concept has two files: best-practice explains 'how and why', implementation provides 'drop-in working code'. This pairing of theory with runnable examples is what sets the repo apart from ordinary tutorials — copy the implementation, then revisit the theory for maximum learning.
04Orchestration workflows: multi-agent collaboration
The orchestration-workflow docs show how to coordinate multiple agents — planning, coding, reviewing and testing each take a role while you own decisions and sign-off. This is where Agentic Engineering lands: not letting AI write everything at once, but orchestrating it to work in steps, self-check and deliver.
05Tips & tricks (Boris Cherny and community practice)
The repo gathers hard-won community experience: how to write CLAUDE.md so every session starts from the same understanding, how to give agents explicit acceptance criteria, and how small commits spread risk. These field notes often move the needle more than official docs.

Practice tips

Start small: add a CLAUDE.md to one project documenting architecture conventions and common commands, then compare one session's difference.
Copy one subagent example from the implementation folder into .claude/agents/ and observe its effect on task quality.
Run one multi-agent orchestration flow (plan → code → review → test) and compare it against your usual single-agent loop.
Re-read the best-practice folder weekly — as your Claude Code skill grows, the same docs will hit differently.