ai-coding-guide: a 92-part Chinese tutorial for Claude Code and Codex

What the repository actually is, and what it is not

This is a documentation project, not software. There is no package to install, no binary, no library to import. The README describes 92 articles and roughly 520,000 Chinese characters, split into a 53-article Claude Code track and a 39-article Codex track, with Codex labelled as the primary recommendation. The repository's own framing is that it is written for people starting from zero, moving from installation through to engineering practice. The topics list confirms the scope: agent, ai-coding, anthropic, chinese-tutorial, claude-code, cli, codex, llm, openai, tutorial. The primary language field comes back as unknown, which fits a repository whose substance is prose and images rather than source files. The homepage, coding.stormzhang.ai, is where the reading experience lives; the repository is the source of that site. Anyone arriving expecting a CLI wrapper around Claude Code or Codex has the wrong artefact. The problem it addresses is narrower and more real: the official Claude Code and Codex documentation is in English, is organised around product surfaces rather than a learning sequence, and assumes comfort with a terminal. A Chinese reader who has never used a coding agent has no obvious starting point. This project supplies one, and says so explicitly.

The two-track structure and why Codex leads

The split is not cosmetic. The Claude Code track runs 53 articles and the Codex track runs 39, yet the README names Codex as the main recommendation and points new readers there first. That ordering is a judgement about where a beginner should spend attention, and it is worth reading as one. The Claude Code track is the more feature-dense of the two, with dedicated articles on MCP, subagents, plugins, Agent Skills, hooks, checkpoints, settings.json, output styles, the Agent SDK, GitHub Actions integration, Chrome control, parallel tasks, voice mode and environment variables. There is a closing sequence on best practices, anti-patterns, troubleshooting and a glossary. The Codex track is described as covering four entry points, AGENTS.md, sandbox approval, config.toml, a memory feature called Chronicle, worktrees, and migration from Claude Code. That last item is the interesting one: a migration article implies the author expects readers to arrive already using the other tool, which makes the two tracks less parallel than the numbering suggests. The Claude Code track also carries a Remotion video article marked as optional reading, a sign that the author let scope drift past the core product.

How a reader moves through it: entry points and article shape

Each article has a stable URL under the homepage domain, following a pattern like coding.stormzhang.ai/claude-code/02-install or coding.stormzhang.ai/codex/03-install. The README states that every hands-on section gives a complete command plus the expected output, so a reader can run the step and see feedback immediately. It also states that each article contains at least three first-person accounts of mistakes or decisions, with concrete details and real numbers. That is a claim about editorial method rather than a feature, and it is the kind of claim a reader can check in one article. The stated sourcing rule matters more: the README says all features, commands and default behaviours are verified against the Codex official site at developers.openai.com/codex and the Claude Code official docs at code.claude.com/docs/zh-CN, and that the project does not copy third-party guesses. If that holds, the Chinese text is a translation and sequencing layer over primary sources rather than an independent account. The documentation also claims 81 SVG and PNG illustrations in a consistent dark style, capped at ten nodes each. No release has been published, so the repository is distributed as a branch rather than versioned artefacts.

The beginner-facing writing method, and where it can mislead

The README describes a three-part pattern for each new concept: a scenario to introduce it, a life analogy, and a real usage scenario. For an audience that does not know the command line, that structure lowers the entry cost considerably. The risk is specific to this domain. Analogies for concepts like context windows, permission scopes or sandbox approval tend to compress the parts that cause damage when misunderstood. An analogy for a permission prompt is fine until the reader has to decide whether to grant an agent write access to a repository. The article list shows the author is aware of this: there is a dedicated Claude Code article on permissions, one on security and risk boundaries, and one on context management framed around losing memory and burning tokens. Those three carry more weight than the rest of the track for anyone who intends to let an agent touch a real codebase. A reader who skims the install and first-run articles and skips the permissions and security pair has taken the analogy without the constraint. The troubleshooting and anti-pattern articles exist for the same reason, and they are placed at the end, which is the wrong position for a reader who wants to avoid the mistakes rather than diagnose them.

Licence, distribution and what the MIT grant covers

The repository carries an MIT licence, and the README links a LICENSE file. MIT covers the text and the repository contents under the usual terms: reuse, modification and redistribution with the copyright notice and licence text retained. It does not extend to the third-party product documentation the articles are verified against, and it does not cover the Claude Code or Codex products themselves, which are governed by their own vendor terms. The 81 illustrations are part of the repository, so the same MIT grant appears to apply to them, though the README describes them as original work without stating a separate media licence. The homepage is a separate surface from the repository and its terms are not stated in the material. For a team that wants to fork the content, translate it, or embed it in internal onboarding, MIT is permissive enough to do that, with the caveat that the underlying facts about two commercial products will age independently of the licence. Nothing here is legal advice; read the LICENSE file and the vendor terms before redistributing.

Maintenance cost and the staleness problem built into the format

The repository's last push is recorded as 2026-09-02, and no releases have been published. That means there is no versioned snapshot to pin against, and a reader who clones gets whatever the branch holds at that moment. The maintenance burden here is unusual. A tutorial about a stable library degrades slowly. A tutorial about two fast-moving commercial agent products degrades with every vendor release, and this one covers config keys such as config.toml and settings.json, approval and sandbox behaviour, pricing and subscription articles, and model integration pages for DeepSeek and other Chinese providers. Pricing and third-party model support are the most volatile sections in the list. The README's stated policy of verifying against official docs is the right mitigation, but it is a manual process, and the article count works against it: keeping 92 articles aligned with two vendor doc sets is ongoing editorial labour, not a one-time write. There is no changelog in the supplied material, so a reader cannot tell which articles were rechecked after a given vendor release. That is the single most useful thing the project could add.

Alternatives and how their approach differs

The obvious alternative is the official documentation itself. Claude Code publishes Chinese-language docs at code.claude.com/docs/zh-CN, which the README names as its own verification source. If a Chinese translation of the reference material is what you need, that page already provides it, maintained by the vendor and updated with the product. The difference is shape, not language. Official docs are organised by feature and assume you know what you are looking for. This project is organised as a course, with an install article, a first-run article, a how-it-works article and a capstone project, which is a sequence the official docs do not attempt. The second alternative is English-language community guides and video courses for the same two tools. Those trade the Chinese-language barrier for currency and often for depth on a single feature. A reader comfortable in English gets faster updates from vendor docs and community write-ups; a reader who is not gets the sequenced path here at the cost of lagging behind vendor changes. Neither alternative gives you both, and the README's own sourcing policy concedes that this project sits downstream of the official pages.

Who should read it, and the first thing to check

The fit is a Chinese-reading developer or technically curious non-developer who has not used a coding agent before and wants a path rather than a reference. The Codex track is where the README sends that reader, and the four-entry-point article plus the install article are the natural start. The wrong fit is an experienced user of either tool looking for advanced or unusual material, an English-only reader, and anyone who needs guaranteed-current API or pricing facts. A team evaluating whether to standardise on Claude Code or Codex will find orientation here but not a decision framework, since the two tracks are written as parallel curricula rather than a comparison. The first thing to check before committing time is whether the specific article you need has been rechecked against the vendor page it cites. Open the linked official page for the command or config key in question and compare. If the article and the vendor page disagree, the vendor page is the source of truth, and that gap tells you how much of the track is currently in sync.

Editorial conclusion

Adopt it if you are a Chinese-reading beginner who needs a sequenced path through Claude Code or Codex, and start with the Codex 39-article track the README names as the main recommendation. Do not adopt it as an English-language reference, as a substitute for the official docs it cites, or as a source for API and pricing details that change between releases. Before relying on any single article, open the linked official page for the command or config key you intend to use, and check the repository's last push date against the vendor doc it points at.

Official sources

1. Issues
2. License: MIT
3. Project website
4. README
5. stormzhang/ai-coding-guide on GitHub