The AI coding guide is organised by who you are, not by which tool you use

A handbook with a reading order for six kinds of reader

The repository is described as a practical guide to AI programming tools, with 66 Claude Code tips, best practices for ten tools and copyable configuration templates. It ships a Simplified Chinese front page with an English version alongside it, and the licence is Apache-2.0.

The organising idea is unusual and it is what makes the repository worth reading rather than skimming. The front page opens with a comparison sheet that puts the ten tools side by side across type, context, configuration file and common commands, and then it gives a recommended path per reader profile rather than a table of contents.

Those profiles are concrete. A complete beginner who has never used an AI coding tool starts with the sheet, then Copilot in its free form inside VS Code, then prompt engineering. A frontend or everyday coding developer goes to Cursor first. A backend or refactoring or large project developer starts with Claude Code, then task decomposition. Someone migrating from Copilot reads Copilot, then Claude Code with a comparison of the differences. Someone on a ChatGPT Plus or Pro subscription is pointed at Codex CLI and then at a comparison with Claude Code. Someone controlling costs or working alone is pointed at Aider with a local model, or Codex CLI with its --oss flag. A team aiming at quality delivery is pointed at Kiro, then code review, then testing strategy.

That last set is the part most documentation does not attempt. A guide that tells a complete beginner where to start, a migrating user what changed, and a team lead which review process to wire in is answering the question people actually have, which is where do I begin, rather than the question a tool vendor would like asked.

Ten tools, and the one axis that separates them

The tool table is the most information-dense thing in the repository, and the distinguishing column is type, not feature list.

There are three CLI agents, one of which is the largest part of the repository. Claude Code is the CLI Agent with 66 tips and a complete Agent, Skill and Hook workflow. Codex CLI is the CLI Agent from OpenAI, open source in Rust, with a sandbox and an AGENTS.md file, and usable with a ChatGPT subscription.

There are IDE plugins and IDEs. Cursor is an IDE with .cursor/rules/*.mdc rule files, Agent and Plan modes and a parallel Agents Window. GitHub Copilot is an IDE plugin with inline completion, an Agent mode and custom Agents and instructions. Devin Desktop, formerly Windsurf, is an IDE with a local and cloud Agent board. Gemini CLI is a CLI from Google with a one million token context, and the note that personal users have moved to Antigravity CLI. Kiro is an IDE and CLI from AWS built around spec driven development. Aider is a CLI that is git native and supports almost every large language model. Trae is an IDE from ByteDance with a free tier and a domestic edition that connects directly.

The axes that decide between them are therefore not features but working style. A CLI agent is driven by prompts in a terminal and works over your whole repository. An IDE plugin completes lines and dispatches an agent inside the editor. A spec driven IDE asks for the requirement first and derives the work from it.

That distinction is the one a selection guide should draw, and the guide draws it in a single table column.

The common layer, and the seven topics that are not about any tool

Underneath the ten tool directories sits a common directory of methods that apply whatever you use, and the guide is careful that these are not generic advice.

Prompt engineering is described as prompt techniques for AI coding scenarios rather than general prompt engineering, which is a distinction that matters because the failure modes are different. Task decomposition exists to break a large task into pieces an AI can do in one go. Code review covers how to get an AI to review code. Debugging is presented as a systematic method rather than a list of tricks. Context management is about controlling the context window so the model does not get worse as the conversation grows. Testing strategy covers how to use AI for tests and where it goes wrong. And security covers the risks of AI-assisted coding and how to guard against them.

The per-tool pages all follow the same five part structure, which the front page states explicitly: core concepts, quick start, prompt techniques, advanced usage, configuration templates. That consistency is what makes the ten pages comparable, and it is why the repository has both an English and a Simplified Chinese front page rather than only one.

The debugging, context management and testing pages are where the guide is most likely to earn its place, because those are the topics where a team's own experience matters more than a vendor's feature list, and where a written method can be applied without switching tools.

The pitfalls directory, written as four parts per failure

The most distinctive part of the repository is the pitfalls directory, and it describes itself precisely: 31 real stumbles across Claude Code, Cursor, Copilot and Aider, each written in a four part structure of symptom, root cause, how to get out, and how to prevent it.

That format is the point. A pitfall written as a symptom alone is a support ticket. A pitfall written as symptom, root cause, exit and prevention is an entry in a runbook, and the fourth part is what makes it worth keeping.

The count is specific and the coverage is specific, which tells you these are observed failures rather than hypothetical ones. Thirty-one across four tools is roughly eight each, and the tool spread matches the guide's claim that these problems are not unique to one product. A context window that overflows, an agent that edits the wrong file, a config file that is picked up from the wrong directory: these are behaviours of the interaction between a model and a repository, and they show up in every tool that offers one.

The two workflow collaboration pages follow the same logic. One covers Claude Code with Cursor, where Claude Code handles architecture and complex refactoring and Cursor handles everyday coding. Another covers Claude Code with Copilot, where Claude Code takes the agent tasks and Copilot does inline completion. Splitting the work by what each tool is good at is a more defensible strategy than picking one and using it for everything, and it is a claim the pitfalls directory supports, because the same failure modes show up in both halves.

The tool selection guide is described as one table that says which tool for which scenario.

The wider ecosystem the guide sits inside

The front page closes with a related projects section, and it is worth reading because it shows the guide is one node in a set of repositories rather than a standalone document.

There is a trilogy project that reorganises the same content by reader profile into three volumes, as PDF and online editions. There is a work methodology project offering twenty skills that teach an AI how to work, covering test driven development, debugging and code review, and an expert role library with 277 ready-to-use AI expert roles, including 64 original ones. There is an orchestration engine that takes one sentence and runs 276 experts together across fifteen models and eleven keyless providers. And there is a security middleware with eight layers of defence, data loss prevention on data flows and injection detection, with zero dependencies, including an MCP server.

Two things are worth noting about that list. First, the branch of AI video creation sits alongside it, with a prompt methodology and a short video production line, so the same author is building a parallel stack for video. Second, the repository says the main body, the tutorials, templates and scripts, is permanently free and open source under Apache-2.0 for individuals, but that corporate training for groups of ten or more, and paid courses or bootcamps that integrate the repository, are cases the author asks to discuss first.

That distinction tells you what the project is. It is an open documentation repository with a commercial training and content business attached to it, and the commercial interest is in selling courses built on the material rather than in gating the material itself.

Where this guide is the wrong thing to read

Three cases, and they are about audience and currency rather than quality.

The first is if you do not read Chinese. The front page is Simplified Chinese with an English version linked, but the repository's own structure suggests the English version covers the front page, the comparison sheet, the resources list and the changelog and contribution files, given that the top level carries README.en.md, cheatsheet.en.md, ecosystem.en.md, resources.en.md, CHANGELOG.en.md and CONTRIBUTING.en.md, while the per-tool directories are single-language. That is an inference from the file listing rather than something the front page states, and it is the first thing to check.

The second is if you want a spec. The three end-to-end conversation scripts under the scenario workflows page are described as copy and modify, and the configuration templates are described as ready to use. Neither is an executable artifact you install. For a team that wants a repeatable process rather than a set of ideas, this gives you the material to write that process and not the process itself.

The third is currency. There are no GitHub releases for the repository, the last push was on 2026-10-19, and the content documents a set of tools that changes quickly: one entry already notes that Gemini CLI personal users have moved to Antigravity CLI, and another names a tool that has been renamed from Windsurf to Devin Desktop. The CHANGELOG.md and CHANGELOG.en.md are where you would look for what changed, but a guide that documents ten tools at once ages unevenly, and some of its directories will be correct while others are not.

The last limitation is that this is a curated handbook with a point of view, and the front page says so. It recommends Claude Code first for backend and refactoring work, pairs Claude Code with Cursor for collaboration, and steers cost-conscious readers to Aider with local models. Those are defensible choices, but they are choices, and a reader who weighs the trade-offs differently should use the tool selection table rather than the recommended paths.

Editorial conclusion

Adopt this guide if you are a Chinese-speaking developer choosing between AI coding tools or structuring how you use one, since it gives a per-profile reading order and a side by side comparison rather than a single opinion. Do not adopt it as a reference you can follow mechanically, because it is written for an audience with existing tool experience and its scripts are described rather than executable. Verify first that the tool directories cover what you actually use, and check the CHANGELOG because there are no tagged releases and the tool landscape it documents moves faster than the repository does.

Frequently asked questions

Official sources

1. Issues
2. jnMetaCode/ai-coding-guide on GitHub
3. License: Apache-2.0
4. README