Chinese AI Coding Ecosystem¶

The Chinese AI coding ecosystem has converged on several patterns that differ from Western approaches: specification-first development, role-based multi-agent systems, cost-sensitive model selection, and layered CLAUDE.md modularization. Tools like Trae (ByteDance) and frameworks like MetaGPT originated here before gaining global adoption.

Tools Landscape¶

Trae (ByteDance AI-native IDE)¶

Three modes in a single IDE:

SOLO mode is the key differentiator: fully autonomous agent that decomposes tasks, executes, and monitors results without per-step user confirmation. Uses MCP protocol. Multi-model: Doubao-1.5-pro, DeepSeek R1/V3, Claude 3.5 Sonnet, GPT-4o. v2.0 released March 2026, pay-per-token pricing.

MetaGPT / MGX¶

Multi-agent framework simulating a software company:

Product Manager → Architect → Project Manager → Engineer → QA Engineer

Each agent has a defined role, receives context from predecessors, and passes structured output forward. Shared memory + SOP (Standard Operating Procedures) across all agents. Planning an AppStore-style plugin marketplace for 2026.

Unique aspect: human-AI hybrid teams where the human takes one of the roles (e.g., Product Manager) and AI agents fill the rest.

Hermes Agent¶

GitHub: NousResearch/hermes-agent (17K+ stars, MIT, Python, v0.8.0)

Built-in learning loop: - Auto-skill creation from successful patterns - Self-improving skills (agent rewrites underperforming skills) - Cross-session persistent memory - Multi-platform: Terminal, Telegram, Discord, Slack, WhatsApp, Signal, Home Assistant

Widely adopted in Chinese community as a reference implementation of "agent that learns."

Dify¶

All-in-one toolchain for RAG applications. Local deployment, multi-model integration. Primary use: enterprise Q&A bots, domain-specific assistants. Simpler alternative to LangChain for teams that want visual workflow building with local deployment option.

OpenSpec - Specification-Driven Development¶

Chinese-origin contribution to AI coding methodology:

GitHub: ForceInjection/OpenSpec-practise
Docs:   radebit.github.io/OpenSpec-Docs-zh/

Workflow:

Create Change (Proposal) → Implement (Apply) → Archive

A lightweight spec layer that forces AI-human consensus on architecture BEFORE code generation. Each change request is a structured document:

# Change: Add user authentication

## Context
Current state: public API, no auth

## Proposal
JWT-based auth with refresh tokens

## Acceptance criteria
- POST /auth/login returns JWT + refresh token
- Protected endpoints return 401 without valid JWT
- Tokens expire in 1h, refresh tokens in 30d

Only after this document is approved does the AI start generating code. The agent cannot modify the spec after approval (spec freeze before build - mirrors agent design patterns pattern).

AGENTS.md integration: ships with pre-configured agent workflows that reference spec documents.

Chinese CLAUDE.md Patterns¶

@import Modularization¶

Chinese developers use @import for CLAUDE.md more aggressively than Western counterparts. The main CLAUDE.md acts as an index; detailed docs live in separate files:

# CLAUDE.md

@import ./rules/git-workflow.md
@import ./rules/code-style.md
@import ./rules/api-conventions.md
@import ./rules/chinese-specific.md

Rationale: bilingual teams often need separate Chinese and English contexts; language-switching works better in modular files than inline.

Up to 5 levels of recursive import supported. @import is disabled inside code blocks.

Conciseness Guidelines (community consensus)¶

• Keep CLAUDE.md under 300 lines
• Every line must answer: "Would removing this cause mistakes?" - if not, cut it
• Frontier CoT models follow ~150-200 explicit instructions with reasonable consistency
• Don't duplicate rules already enforced by tooling (eslint, prettier → don't write lint rules in CLAUDE.md)
• IMPORTANT/MUST/NEVER/ALWAYS increase compliance but use sparingly - overuse degrades effect

Layered Configuration Structure¶

~/.claude/CLAUDE.md       - global (all sessions)
./CLAUDE.md               - project root (team-shared)
./CLAUDE.local.md         - personal project overrides
.claude/rules/*.md        - conditional context injection

Vibe Coding vs Context Coding vs Spec Coding¶

Chinese discourse developed a taxonomy more explicitly than English-language content:

Key insight from guangzhengli: "Claude Code chose Unix tools (grep, find, git, cat) for context retrieval instead of RAG. By 2026 the real skill is Spec Coding - humans and AI reach consensus on architecture before coding."

Chinese Tech Company Practices¶

DeepSeek¶

• Internally uses Claude Code, Cursor, Copilot for their own development
• DeepSeek-V3.2 training includes agent synthetic data with reasoning capabilities in tool-calling scenarios
• Actively hiring for: Tool Use, Planning, long-term memory, Multi-Agent collaboration positions

Zhipu AI (GLM-5)¶

• GLM-5 (754B params): SWE-bench-Verified 77.8, Terminal Bench 2.0 56.2
• First Chinese model company to offer a Coding Plan tier
• GLM-OS concept: AutoGLM (50+ step autonomous operations), GLM-PC (desktop agent)
• AutoGLM claimed to approach Claude Opus 4.5 performance on agent benchmarks

Meituan¶

Published research on AI Coding and unit testing co-evolution: shifting from verification-driven to test-driven AI coding patterns.

Key Repos¶

claude-code-chinese/claude-code-guide  - community guide, API proxy patterns for China
cfrs2005/claude-init                   - 9 sub-agents, 10 slash commands, 8 rules, hooks
xianyu110/awesome-claudcode-tutorial   - 212+ articles, 25 chapters
KimYx0207/Claude-Code-x-OpenClaw-Guide-Zh - combined Claude Code + OpenClaw, 21 lessons
phodal/build-coding-agent-context-engineering - context engineering for agents
datawhalechina/vibe-vibe               - systematic Vibe Coding tutorial
ForceInjection/OpenSpec-practise       - spec-driven dev implementation

Network/Access Infrastructure¶

Chinese developers treat API access as a first-class infrastructure concern: - claude-init ships with built-in proxy support ("no-VPN" access) - API relay/routing configuration is part of default templates - Cost comparisons weight token-per-yuan heavily; domestic models (GLM-5, DeepSeek) positioned as cost-effective fallbacks

Gotchas¶

• OpenSpec "spec freeze" requires enforcement, not trust. Agents will bypass spec documents if not explicitly blocked from editing them. Implement a hook or file permission to make spec documents read-only after approval
• @import chains can create invisible CLAUDE.md bloat. Each imported file counts toward the effective prompt prefix. If 5 imported files total 2000 lines, that's 2000 tokens loaded every session even when irrelevant. Use conditional imports (.claude/rules/*.md) for context that's only sometimes needed
• Role-based multi-agent systems (MetaGPT pattern) require strict output contracts. When Product Manager agent hands off to Architect, the output format must be precisely defined. Ambiguous handoffs cause downstream agents to hallucinate upstream context

See Also¶

• ai coding assistants
• context engineering
• multi agent systems
• agent design patterns
• claude code ecosystem