vibe-coding-cn: a Chinese-language curriculum for AI pair programming, not a tool

What problem vibe-coding-cn actually addresses

Most AI coding material is a list of prompt tricks. This repository takes a different position: it argues that the hard part is not getting a model to emit code, but keeping generated code trustworthy across a project's lifetime. The README frames the problem as a workflow standard, described as "Prompt + Skill + Context + Quality Gate + 工程闭环", and the docs are organised around that claim rather than around a single tool. The audience is explicit. The README's quick-start path targets newcomers who need network setup, CLI configuration, a development environment and a Git loop before they can use any AI assistant productively. A second audience is teams already using Cursor, Claude Code, Codex or Gemini CLI who want reusable prompt and skill assets instead of retyping instructions. The repository is written in Chinese, and that is a deliberate constraint, not an accident: the topics list includes "chinese", and the entry points, badges and navigation are all in Chinese. If you cannot read Chinese, the repository's value drops sharply, because the prose is the product.

The five propositions and the isolation-review rule

The README opens with five propositions that function as the repository's design rationale. The first quotes Demis Hassabis on solving AI first. The second defines a "generation domain" (生成域): anything a model's output can directly or indirectly implement, drive, constrain, modify, verify or influence, with the slogan that reachable output means reachable capability. The third, "model devouring" (模型吞噬), predicts that intermediate layers built to compensate for weak models, including prompt techniques, workflow scaffolding, indexing and external memory, will be absorbed as models improve. That is an unusually self-undermining claim for a repository whose contents are prompt techniques and workflow scaffolding, and the README does not resolve the tension. The fourth proposition is the one with practical consequences. Called isolated review (隔离审查), it states that AI output is a candidate solution, not verified fact, and that generation, review and verification should be split across separate contexts. The README cites NIST on independent review, OpenAI and Microsoft on external testing and red teaming, and LLM-as-a-judge research on self-preference bias. The concrete instruction is to open a fresh session, tell the reviewing model that the previous result is untrustworthy, and require it to re-read the original materials, business code, goals, constraints and verification results. Whether that instruction survives contact with a large codebase is not addressed.

Glue coding: search for mature parts before generating new ones

The fifth proposition, capability orchestration (能力编排), is the repository's engineering thesis and the source of its "拼好码" (glue coding) concept. The argument is that AI-assisted programming should shift from an implementer mindset to an integrator mindset: identify existing mature capabilities, assess their maturity, define adaptation boundaries, and write the minimum amount of original code needed to close a business loop. The README gives a seven-step practice flow: write the requirement with goals, inputs, outputs, constraints and acceptance criteria; have the AI decompose the requirement into capability domains and search for official capabilities, de facto standards, toolchains, mature repositories, mainstream SDKs and platform services; evaluate candidates on maintenance status, licence, documentation quality, production usage, ecosystem compatibility, replacement risk and integration cost; choose a combination and justify why other options and self-development were rejected; fix input and output contracts, data models, interface contracts, error handling, dependency isolation and rollback paths; generate only glue code for connection, adaptation, orchestration, configuration, business rules and tests; and verify with tests, types, schemas, CI, scripts and checklists while keeping evidence, alternatives and rollback paths. The closing line is that you should not rebuild what can be reused and should not invent what can be orchestrated. The steps are sensible, but step three asks for production usage evidence and step five asks for rollback paths without naming a mechanism for either.

Repository layout: knowledge base, assets and Codex configuration

The README's entry-point table is the clearest description of what the repository contains. docs is the knowledge-base root and splits into getting-started (network, Codex CLI, development environment, Git loop), workflow (turning a requirement into a plan, changes, gates, commits and a retrospective), concepts (problem solving, glue coding, system construction, keyword system), references (technology stacks, quality gates, templates, common pitfalls), philosophy (thinking models, programming craft, software engineering fundamentals) and research (new technologies, notable repositories, engineering trend judgements). Separate top-level directories hold prompts, skills, tools and assets. The prompts directory points to an online prompt table, and the skills directory is described as reusable executable AI capability modules. tools contains Codex configuration with a one-click install path under tools/config/.codex, plus conversion tools and external tools. assets holds external resources, AI citation corpora and static files. A Wiki is also linked. The Python primary language is worth noting: nothing in the README describes a Python package, module or CLI entry point, so the language label most likely reflects helper scripts and conversion tooling rather than the repository's main deliverable, which is Markdown.

Getting it running: Codex configuration and the Git loop

There is no install command for the tutorial content, because it is read rather than executed. The one setup path the README advertises is Codex configuration: the badge links to tools/config/.codex/README.md with the label "Codex 配置一键安装" (one-click install), and the setup section of the README is referenced as the full setup flow. The learning map at docs/getting-started/learning-map.md is the recommended first stop and covers network environment, CLI configuration, development environment and the Git loop in that order. For reusable assets, the README points to prompts/README.md under the anchor 在线提示词库 (online prompt library), skills/README.md under 当前保留 (currently retained), and assets/README.md under 外部资源本地注册表 (external resource local registry). The wording of those anchors is itself informative: prompts live in an external table rather than in the repository, and skills are marked as "currently retained", which implies earlier skills were removed. The README does not document what happens to a skill that is dropped, nor does it give a version pin for the Codex configuration, so a one-click install today may not match the CLI you are running next month.

Where the repository is weak

The first limitation is that the README promises a standard and delivers a curriculum. A standard implies conformance criteria, versioning and a change process. The README lists no releases, so there is no changelog, no semantic version for the prompt or skill assets, and no way to tell whether a prompt you copied last quarter still reflects the current guidance. The second limitation is the self-undermining proposition three. If prompt techniques and workflow scaffolding are destined to be absorbed by stronger models, then a repository whose core value is prompt techniques and workflow scaffolding has a stated expiry it never quantifies. The README does not say which parts it expects to survive. The third is the isolation-review instruction. Telling a reviewing model that the previous result is untrustworthy is a prompt-level control, and the README offers no mechanism for enforcing it, no session-management tooling, and no evidence that a fresh session actually avoids anchoring on prior context. The fourth is the glue-coding evaluation criteria. Maintenance status, licence and production usage are named as things to check, but the repository provides no registry, no licence scanner and no template that records the outcome. Finally, the README does not document rollback for the Codex configuration or for skill removal, and it does not state a support commitment. The repository is not archived and its last push was on 2026-09-13, so it is current, but currency is not the same as a maintenance guarantee.

Alternatives and the difference in approach

The closest alternative is the official documentation for each CLI the repository covers: Codex, Claude Code, Gemini CLI and Cursor all ship their own setup guides, and those guides are the authoritative source for flags and configuration keys. The difference is scope and language. Official docs explain one product and go stale against that product's release cadence; vibe-coding-cn tries to sit above the products and describe a vendor-neutral workflow in Chinese, at the cost of being one step removed from any specific CLI's behaviour. A second alternative is a general prompt library, which offers copy-paste prompts without a workflow, quality gates or a review discipline. vibe-coding-cn's distinguishing claim is exactly that prompts alone are insufficient, so a prompt library is a strict subset of what this repository attempts. A third alternative is a framework such as an agent orchestration library, which enforces workflow in code rather than in prose. That is the real trade-off: an orchestration library constrains what the model can do and can be tested, while this repository constrains the human and cannot be tested at all. If you want enforcement, this is the wrong shape of project. If you want a reading path and reusable Chinese-language assets, it fits.

Licence, upgrade cost and who should adopt it

The repository is MIT-licensed, which permits commercial use, modification and redistribution provided the copyright notice and permission notice are retained. MIT does not grant trademark rights or patent rights, and it carries no warranty, so if you copy prompts or skills into a product, you inherit the obligation to keep the notice and you accept the software is provided as is. The README does not discuss third-party licences for the external resources listed in assets, and the glue-coding flow explicitly tells you to check licences of the repositories you integrate, so that check is on you rather than on this project. Upgrade cost is low in the mechanical sense and unclear in the semantic sense. Pulling the latest Markdown costs nothing; knowing what changed costs everything, because there is no release history. The practical mitigation is to pin the commit you read from and to re-read the specific files you depend on, rather than to assume the repository's guidance is stable. Teams that want a shared vocabulary for AI pair programming in Chinese, and individuals working through the learning map from network setup to the Git loop, are the natural adopters. Anyone who needs a supported library, an English resource, or a guarantee that a prompt will keep working should look elsewhere. The first thing to verify is whether the Codex configuration under tools/config/.codex matches the CLI version you run, followed by whether the skill you intend to reuse is still listed under the currently retained section of skills/README.md.

Editorial conclusion

Adopt it if you are a Chinese-speaking developer or team that wants a structured reading path through AI pair programming and is willing to treat the prompts, skills and Codex configuration as starting points rather than finished assets. Do not adopt it if you need a library with a stable API, a benchmark, or an English-language resource; the repository is a tutorial and reference collection, and the README does not document versioning, deprecation or a support commitment for the prompts and skills it ships. Before relying on anything, verify that the specific document or skill you intend to use still matches the CLI versions you run, because the last push was on 2026-09-13 and the material lists no releases, so there is no changelog to tell you what changed between revisions.

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. tradecatlabs/vibe-coding-cn on GitHub