Vibe Coding 中文指南

★ 22.7kGetting Started

#tutorial#chinese#beginner

Overview

A 22.6k+ star Chinese Vibe Coding playbook, positioned as an 'ultimate workstation' for turning ideas into reality through AI pair programming. Instead of vague concepts, it lays out an end-to-end flow — ideation, tech stack selection, implementation planning, step-by-step development, debugging and extension — built on 'planning-driven, fixed context, AI pair execution'. The repo ships a full prompt library, Skills packs and multilingual translations you can use as a daily workstation.

Who it's for

Developers who want to systematically master Vibe Coding — especially those burned by AI code that runs off track and leaves projects a tangled mess.

What you'll learn

01Core belief: planning-driven, don't let AI run wild
The repo insists 'planning is everything': be careful about letting AI plan on its own, or your codebase becomes an unmanageable mess. The whole workflow is document-driven — first write a Game Design Document (GDD) / PRD, tech stack and implementation plan as Markdown, then have the AI execute step by step, instead of letting it write everything in one prompt.
02Four layers of method: philosophy, method, technique, tools
Content is organized around a classical Chinese thinking framework: principles (e.g. 'whatever AI can do, don't do by hand', 'context is the first principle of Vibe Coding'), methods (e.g. 'one-sentence goal + non-goals', 'interfaces first, implementation later', 'change one module at a time'), concrete techniques (e.g. debug by giving only 'expected vs actual + minimal reproduction', switch sessions when code grows large), and tools (Claude Code / Codex CLI, Cursor, Warp, Neovim, etc.).
03End-to-end workflow: from documents to a memory bank
The onboarding flow is very concrete: have the AI draft a game design document and recommend a tech stack, run /init to generate CLAUDE.md/Agents.md rules that force the AI to 'read memory-bank before writing any code'; then produce a step-by-step implementation plan, paired with progress.md and architecture.md as a memory bank so every new session can pick up where the last left off.
04Prompt and skill assets: ready to use
The repo's biggest asset is a reusable prompt system: coding_prompts (the full chain from requirement clarification to planning to execution), system_prompts (constraining AI behavior), meta-prompts (prompts that generate prompts), plus an integrable Skills library — even a small tool that converts prompts between Excel and Markdown.
05Pitfalls and tips
It packs hard-won experience: tell the AI clearly what it can and cannot change; use /rewind or git reset to roll back a derailed iteration; when truly stuck, bundle the whole codebase into a single file and hand it to the model; it also provides a model-tier reference suggesting complex tasks should only use the top-tier models.

Practice tips

Pick one real small idea (swap GDD for a PRD if it's an app, not a game) and walk the full flow: design doc → tech stack → implementation plan → memory-bank.
Set up CLAUDE.md rules before writing code — especially an 'Always' rule forcing the AI to read memory-bank first, so it won't dump one giant monolith file.
When debugging, give the AI only 'expected vs actual + minimal reproduction', and roll back in small steps with /rewind or git reset instead of letting it keep digging.
Pick a few prompts and skills from the repo's prompts/skills folders that fit your work, integrate them into daily sessions, then gradually build your own versions with the meta-prompts.