Vibe Coding Chinese Guide
A Chinese end-to-end Vibe Coding curriculum that turns ideas into runnable products via prompts, skills, context discipline and quality gates.
Per-dimension scores and reasoning
Evidence shows a tutorial/knowledge-base repository. The README explicitly requires human confirmation for sensitive credentials, production databases, dangerous commands and irreversible operations, and the Codex config install backs up before writing and supports restore, so least_privilege/user_confirmation/rollback each score 1 as thin but real. However, no dependency manifest, lockfile or vulnerability-scanning evidence exists, so dependency_security is 0. Data-flow and sensitive-data handling remain principle-level statements with no concrete mechanism, scoring 1 each. source_attribution has a LICENSE and a third-party notices entry point, but the many external links and citations in the README are not individually attributed, so 1.
The README is internally structured and its entry table broadly matches the directory layout, but there is a clear inconsistency: badges and links point to tukuaiai/vibe-coding-cn while the reviewed repository is tradecatlabs/vibe-coding-cn, and referenced model versions (e.g. Claude Opus 4.7, gpt-5.5) cannot be corroborated from in-repo evidence, so self_consistency is 1. No runtime dependencies are declared; only external tools and an online spreadsheet are referenced, so dependency_availability is 1. Failure messaging is limited to scattered hints (e.g. explain pasted errors line by line) with no systematic error-handling documentation, so 1.
The README explicitly separates beginner, developer, team, Prompt, Skill, quality-gate and GEO/SEO learning routes, so audience_and_scenarios scores 2. capability_boundaries has 'experimental methods' and 'may be rewritten at any time' disclaimers but no systematic definition of Agent capability scope, so 1. trigger_precision has prompt and skill entry points but vague trigger conditions, so 1. environment_fit covers WSL2/Ubuntu/macOS/Linux Server and multiple CLIs, so 2.
Information architecture is clear with docs/prompts/skills/tools/assets layering and an entry table, so 2. Install notes only point to a Codex one-command install and a CLI setup doc; full steps are not in this repository, so 1. Naming stability is weakened by tukuaiai vs tradecatlabs mixing, so 1. Examples and FAQ exist as AI-citation material and a recommended summary, but runnable examples are absent, so 1. Known limitations appear as 'experimental' and 'adopt dialectically' phrasing without a systematic list, so 1. LICENSE is a complete MIT text, so 2. No version number or changelog, so 0. Maintenance responsibility has a Telegram group and channel but no explicit maintainer commitment, so 1.
Output usability is real: copy-paste prompts, entry tables and a layered method are directly usable by the target audience, so 2. Marginal value is limited because the content is largely methodology and link aggregation with high overlap with generic AI-coding tutorials, so 1. Cost-benefit: the repository is large with many links and high maintenance cost, but no quantified benefit or validation data is provided, so 1.
claim_traceability: the README cites NIST, OpenAI and LLM-as-a-judge claims without concrete source links, so 1. cross_source_corroboration: in-repo test files (e.g. test_review_case.py, test_analyzer.py) do not directly correspond to the README's claims, so corroboration is weak, scoring 1. fact_inference_separation: the README blends opinion, experience and factual statements without clear separation, so 1.
- Repository naming and badge links point to tukuaiai/vibe-coding-cn while the reviewed repository is tradecatlabs/vibe-coding-cn, an identity/attribution mismatch that should be resolved before relying on it.
- Referenced model versions (e.g. Claude Opus 4.7, gpt-5.5) and external claims (NIST, OpenAI, LLM-as-a-judge) lack verifiable citations and should not be treated as established fact.
- No dependency manifest, lockfile or vulnerability-scanning evidence is provided; installing the Codex config or running scripts requires independent dependency-risk assessment.
- Many external links and an online spreadsheet (Google Sheets, Telegram, third-party repos) carry availability and content-change risk, and some links appear promotional.
- In-repo test files come from third-party reference projects under a research subdirectory and cannot be counted as test coverage for this repository's own Agent product.
What does this agent do, and when should you use it?
vibe-coding-cn is a documentation-and-resource repository that teaches a complete AI pair-programming workflow in Chinese, from machine setup to delivery review. It is not a runnable service: the codebase is a knowledge system made of docs (getting-started, concepts, philosophy, references, workflow), a research domain, prompt and skill entries, a tools directory, automation scripts, and an external-resource registry. The method is organized into five layers — Prompt (one-shot instructions), Skill (reusable capabilities), Context (durable project context), Quality Gate (tests, CI, types, schemas, checklists) and an engineering loop. Codex CLI is the documented default local execution entry, with Claude Code, Gemini CLI, Cursor and others listed as alternatives. Outputs are structured prompt assets, skill modules, engineering templates, research notes and auditable Git history. Repository quality is enforced through a Makefile plus GitHub Actions that run Markdown lint, local link checks, docs structure, metadata and AI-citation consistency checks, rather than through deployed services. Licensed under MIT.
The repository's execution path is documentation plus scripts. The README routes readers through docs/getting-started/learning-map.md into beginner, developer, team, prompt, skill, quality-gate or GEO tracks. docs/getting-started/cli-setup.md covers Codex CLI installation and login, while tools/config/.codex/ ships an installable config baseline that writes to ~/.codex/ with automatic backup and restore. skills/ contains executable skill modules such as skills/auto-skill (validated with validate-skill.sh --strict) and skills/auto-tmux, which wraps tmux capture-pane and send-keys into capture/scan, send, rescue, record and hub subcommands for supervising multiple AI terminals. prompts/ points to a hosted spreadsheet of meta, system, coding and user prompts. tools/prompts-library/main.py converts between Excel and Markdown and splits a JSONL Excel workbook into per-sheet JSONL directories. Verification is handled by scripts/check-local-links.py, check-markdown-details.py, check-directory-docs.py, check-metadata.py, check-ai-citation.py, check-wiki.py, sync-doc-toc.py and Makefile targets such as lint, check-links, check-details, check-doc-structure and test; CI runs markdown-lint, local link checks, docs structure checks and a link checker on develop pushes and pull requests.
- A developer on a brand-new machine or new to AI coding who wants a step-by-step path through network setup, Codex CLI, Node.js/Python, editor choice and Git workflow from learning-map.md.
- An independent developer already using Cursor who keeps fighting context drift and AI-induced breakage, and wants memory-bank documents, implementation plans and test-based gates instead.
- A tech lead who needs one shared AI-coding standard for the team and can reuse the prompt assets, AGENTS.md operating rules and the quality-gates-and-pitfalls checklist.
- An engineer who wants to convert existing research into an executable plan, using the research domain contract, evidence and maturity rules, plus research-transfer-synthesis.md.
- A developer running several AI terminals in parallel who needs capture/send/rescue/record/hub workflows from skills/auto-tmux to build a commander-plus-worker bench.
- A technical writer maintaining bilingual documentation who wants Markdown lint, local link checks, directory-structure and AI-citation consistency checks wired into CI.
What are this agent's strengths and limitations?
- Genuinely systematic: it covers five layers (prompt, skill, context, quality gate, engineering loop) instead of dumping a prompt list, with explicit directory responsibilities for docs, research, skills, tools, scripts and assets.
- Verifiability is built into the repository itself — Makefile plus scripts/check-*.py plus .github/workflows/ci.yml cover Markdown lint, local links, collapsible-block structure, docs TOC structure, directory README/AGENTS coverage, metadata and AI-citation consistency.
- The default execution entry is stated plainly (Codex CLI) while Claude Code, Gemini CLI, Qwen and Ollama are offered as alternatives, so readers are not locked into one model vendor.
- Addresses concrete Chinese-language developer pain: new-machine setup on WSL2/Ubuntu, context management, AI losing control, and unmaintainable code each get their own document rather than slogans.
- Released under MIT with CONTRIBUTING.md, CODE_OF_CONDUCT.md, CODEOWNERS and issue/PR templates, making internal forking and contribution straightforward.
- Experimental areas are explicitly fenced off: the tmux swarm workflow is declared not to be an unattended production system, and the included research projects document that they run their own separate verification.
- It is a documentation and knowledge-base project with no runnable application, deployable service or public API; the README confirms there is no dev server and no fixed port, so anyone expecting a turnkey tool will be disappointed.
- Many core prompt assets live in external Google Sheets links, so offline usability and long-term stability depend on a third party; the repository only keeps an entry point.
- The local quality gates need Node.js 22+, Python 3.8+, pip dependencies and Git submodule initialization, which is extra environment upkeep compared with just reading docs.
- The author describes it as a project that keeps growing and self-negating, and the experimental methods (meta-methodology, glue coding, tmux swarm, philosophy toolbox) are flagged as potentially being rewritten, so readers must judge maturity themselves.
- Recommendations for specific model versions (e.g. Claude Opus 4.7, gpt-5.5 xhigh) age quickly as models iterate, something the repository itself acknowledges.
- Review enforcement rests on the path-level .github/CODEOWNERS baseline; there is no release versioning, CHANGELOG or compatibility matrix to support adoption decisions.
How do you install or deploy this agent?
This is a documentation repository: there is no Dockerfile, docker-compose.yml or Kubernetes/Helm entry point, and no fixed service port. "Installing" means cloning the repo and setting up its verification toolchain.
Prerequisites: Git, Node.js 22+ (used for the pinned [email protected]), Python 3.8+.
git clone https://github.com/tradecatlabs/vibe-coding-cn
cd vibe-coding-cn
git submodule update --init --recursive
pip install -r tools/prompts-library/requirements.txt
# only if you need the prompts-library Google API / JSONL helper scripts
pip install -r tools/prompts-library/scripts/requirements.txtIf you only want an AI CLI environment, the README offers a zero-dependency starter: paste the prompt from the "1 minute quick start" section into ChatGPT / Claude / Gemini web chat, and it will generate OS-specific Codex CLI and local-agent setup steps. The Codex baseline in tools/config/.codex/README.md installs to ~/.codex/ with automatic backup and restore.
How do you use this agent?
- Run the local checks first to confirm the toolchain works:
make help # list available targets
make lint # repo-wide markdown lint
make check-links # local relative link check
make test # all local quality gates- Pick a track in docs/getting-started/learning-map.md. The beginner path is learning-map → vibe-coding-experience → problem-solving → glue-coding → quality-gates-and-pitfalls.
- With an AI CLI configured, let the local agent read the repository docs and drive requirements → plan → edits → gates → commit → review per docs/workflow/development-process.md.
- Pull the ready-made assets: prompts/README.md links the hosted prompt library, skills/README.md lists currently retained skills, and assets/README.md plus assets/external-resources/ are the local registry of external resources.
- Convert prompt files:
cd tools/prompts-library && python3 main.py- Validate a skill in strict mode:
skills/auto-skill/scripts/validate-skill.sh skills/auto-skill --strict- Experimental tmux swarm mode: use the capture / scan / send / rescue / record / hub subcommands in skills/auto-tmux to supervise multiple AI terminals. The README explicitly warns that credentials, production databases and irreversible operations require human confirmation.
How does this agent compare with similar options?
The README names and contrasts several alternatives. First, the upstream EnzeD/vibe-coding repository: vibe-coding-cn includes a dedicated "original repository translation" section crediting Nicolas Zullo, and the contributors block renders both repositories' contributor images. Second, Cursor: the FAQ argues that Claude Code and Codex CLI currently outperform it, while pointing readers who prefer Cursor to version 1.1 of the original guide. Third, Claude Code versus Codex CLI, where the README says each better exploits its native model (Claude Opus 4.7 and gpt-5.5 respectively). Separately, tools/chat-vault states it saves chat history for Codex / Kiro / Gemini / Claude CLIs.