A guide to Claude Code whose main instruction is to route your prompts through a third party

The reason this guide exists is a commercial relay for a blocked region

The page opens with a section on using Claude Code in China, and its argument is short. Anthropic does not permit access from that region, so the vendor's site and documentation may be unreachable, downloading the tool or using the API requires a workaround, and the sentence reassures the reader that there is a simple solution. What follows is a named third-party API site, marked as recommended for beginners, and the site returns later in the configuration section as the value to put in the base URL setting, alongside a placeholder for the key and a model name. Between those two points the guide walks through registration, a token management page, and creating a token with an unlimited quota box ticked. Nothing on the page discusses the vendor's terms, what the intermediary does with request content, or how a reader would tell the difference between the two.

The migration instructions end in a colon with nothing after it

Three install routes are offered. The global npm install of the official package comes first, with a warning not to use sudo because of permission and security problems. Then an official standalone client that does not depend on Node, given as a piped shell script, a Homebrew cask and a PowerShell variant. Then migration for readers who already installed the npm version, and this is where the text breaks. It says to close every open session first, then to type a command in the terminal, and the sentence ends in a colon with no command following it. The next paragraph mentions that in practice it is better to run the install first and then type something, and the block that finally appears is the uninstall of the npm package. So the one command the migration section promises to show you is missing.

The repository is one file, and the setup script is pasted rather than shipped

The repository contains a single entry, the readme. There is no licence file, no code, no script directory and no release, and the primary language is recorded as unknown. That matters because the guide offers a convenience script for setting things up, and it is not a file you download: the readme tells you to edit a line in the script, fill in your key, save the script under a name of your choosing, make it executable and run it. The script as pasted is also cut off mid-word, ending on a truncated constant name, and its configuration block holds the key, the base address and a note that the rest should not be modified. So the one artefact this guide asks you to run is text inside a documentation page, with a placeholder in it and a line missing.

The two reference tables did not survive as tables

The command reference and the slash command reference are the most useful parts of the page, and both are broken in the same way. Each was authored as a three-column table with headers for the command, its description and an example, and in the text every cell has become its own paragraph separated by blank lines. So the reader gets a bare header row, then a command name on one line, its meaning on the next, and an example on the third, with no column alignment and no visual grouping. The content itself is fine, including the flag list with its allowed-tools example, a model flag, a plan permission mode and the flag that skips permission prompts entirely. What is missing is the structure that made those a hundred lines scannable into two tables. The launch modes survive as blocks, since they are short enough to have been fenced: an interactive session started with an opening prompt, and the same query run non-interactively through the SDK path and exiting.

claude "explain this project"

claude -p "explain this function"

A bare language label sits above every code block

The formatting carries artefacts from whatever produced this file. Every code block is preceded by its language as a word on its own line, so the reader sees a line reading sh, then a blank line, then the block; the same happens with bash, json, powershell and markdown. The section headings carry escaped punctuation, so the first one renders as a numbered title with a backslash before the dot, and most headings end with invisible spacing characters that survive into the text. The headings also mix two levels of numbering in a way that makes the outline hard to follow, with a 3.3 subsection for working directories sitting between a settings section and a script section with no parent above it. None of this changes the instructions, and all of it makes them harder to trust.

The performance claim has no study attached to it

The introduction contrasts Claude Code with a competing coding assistant and attributes a specific number to the difference: research shows it can raise development speed by more than five times. No study, author, sample or measurement is named, and no link is given. The rest of the introduction is more careful, describing the tool as agentic rather than a completion engine, listing that it reads a whole codebase, runs shell commands, looks at git history and can run tests, and then naming its own limits plainly, that it leans toward developers, does not suit non-technical users, and assumes command line basics. The limits paragraph is the most credible text in the document. The multiplier next to it is the least.

Permission escape hatches next to a template that ends in a push

The permissions material is accurate and includes the sharp edges. The tool is described as conservative by default, asking before operations that might modify the system, with three ways to widen it: a slash command for tools, editing the settings file in bulk, and a session flag for allowed tools. Additional directories can be granted in the settings file, and the flag list documents a permission mode for planning and the flag that skips permission prompts, annotated as use with care. The custom slash command section then supplies a worked template, a markdown file in a commands directory with a placeholder argument, whose eight steps end with writing and running tests, checking lint, creating a commit message, pushing and opening a pull request. So the document hands you a template for an unattended commit and push in the same section where it lists the flag that removes the prompts.

Editorial conclusion

Read the relay section first and decide for yourself whether to follow it. The guide's premise is that the vendor will not serve readers in their region, and its answer is a commercial intermediary that will, which means your prompts, your code context and your key pass through a company the vendor does not know about, on terms the guide never states. It also walks through creating a token there with an unlimited quota selected, which is a billing arrangement you are making with that company rather than with the vendor. The rest of the guide is ordinary, useful documentation that has nothing to do with the relay: three install paths, two settings files, the permission model, the flags, the slash commands and how to write your own. If you already have a working account, that material stands on its own. Before you route anything through a relay, settle three things the page skips: what happens to your requests if it stores them, what the vendor's terms say about intermediaries, and what you would do if the base URL in your settings changed under you.

Frequently asked questions

Official sources

1. claude-code-chinese/claude-code-guide on GitHub
2. Issues
3. Project website
4. README