Text tutorial: Save Kanban from installation to reading by giving Claude Code and Codex save tokens with RTK图文教程：用 RTK 给 Claude Code 与 Codex 省 token，从安装到读懂节省看板

The most convenient way for Windows users is to install with winget. After a command is installed, the path is also easily configured without manually changing the environment variables.

macOS users install with Homebrew; Linux and macOS can also use the official install script, which will load the executable file into the bin in the user directory.

Users who have installed the Rust toolchain can choose to compile and install it from the source code, and the source code package accompanying this tutorial can be used directly.

After installing the executable version command and the save Kanban command once: you can see the version number, you can see the Kanban, and the instructions are correct.

If the Kanban reports an error, it is most likely to be loaded into another project of the same name on crates.io, and you can reload it by installing it from the source code.

Execute the global initialization command. By default, it is oriented to Claude Code and Copilot.

Use the Gemini CLI to add the corresponding parameters, and use Codex to add codex parameters. Cursor and Windsurf also have their own agent parameters.

Initialization writes the hook configuration and automatically overwrites commands such as git status with the RTK prefix.

Once completed, restart the AI tool once for the new hook configuration to take effect.

Have it execute git status once in the AI assistant to see if the returned output has changed to a compact statistical format.

Let it run the test command again to see if only the failed use cases are retained and the passed folds are counted.

If the output is still in the original long format, check that the hook configuration file was written successfully and that the AI tool has finished restarting.

Note that the built-in file reading class tool does not pass through the shell hook, and this part of the command will not be automatically overwritten.

The Kanban board displays the cumulative number of processed commands and the compression ratio, which can be used to quickly determine which commands are most profitable.

The official compression ratio is "reduced bash output", which is not equal to the same percentage of bill drop. The input token also contains system prompts and conversation history, and the output token is completely unaffected.

The number reported by the tool is estimated by the number of bytes divided by four, and there is no built-in word breaker; the proportion is credible, and the absolute number of tokens is only approximate.

If you want to squeeze further, you can fix the commands commonly used by the model into rtk prefix writing, and go directly to the compression channel without relying on hooks to rewrite.

The four categories that are most worth compressing on a daily basis are testing, building, Docker logs, and git diff, which have the largest output of a single command.

If multiple people in the team share the same repository, you can submit the configuration of RTK together to the repository, so that everyone's AI assistant uses the same set of compression rules.

Compression will discard some details. If you suspect that the information is truncated when troubleshooting difficult bugs, you can temporarily run a comparison with the original command.