Graphic tutorial: five minutes to install the Caveman province token plugin, so that Claude Code can do more work with less nonsense图文教程：五分钟装好 Caveman 省 token 插件，让 Claude Code 少说废话多干活

Download this accompanying Caveman v4.0.1 full plugin package (zip, 8.79 MB) and extract it to a fixed directory containing the skills file, CLI, install.ps1/install.sh installation script and browser extension source code.

Users with a smooth network can also skip the decompression and run npx @ caveman-ai/cli @ latest directly, with the same effect. Requires Node.js 22.13 or higher, first use node -v to confirm the version.

The project is Apache-2.0 open source, 110,000 + stars, feel free to review the code and install it.

Windows users open PowerShell in the unzip directory to run.\ install.ps1; macOS/Linux to run./install.sh. The script automatically scans the local Claude Code, Codex, Gemini CLI, OpenCode, etc. configuration and lists the entries to be modified.

Press Enter to confirm and restart your agent session after installation is complete. After that, the normal startup of claude is already in Caveman mode, no additional settings are required.

For classmates who only want to use ChatGPT/Claude on the web, you can use the browser extension in the unzip package: on the Chrome extension page, select "Load the unzipped extension" and point to the extension directory.

The new session asks a lot of usual questions, such as "Explain why the React component is being re-rendered". The answer should be significantly shorter, but the code and commands remain intact.

In the dialogue, enter/caveman as the standard file;/ultracave is more minimal;/megacave extreme saving token, Chinese scenarios recommend trying the standard file first and then descending.

Use/caveman-stats to see how many tokens have been saved in this session, and/caveman-learn to learn your preferences.

The Read less module compresses the logs and test output by default; if your workflow depends on the full raw output, you can turn this block off separately in the configuration.

The Auto file will be used according to the task selection model and inference strength, and the subscription limit will be used to save the most money, and the official measurement will save 54% of the cost.

The team uses the suggestion to lock the version number and see the release instructions before upgrading to avoid changes in skill behaviors affecting the existing process.

This article is compiled from public sources and ships with the matching resource. Product features and pricing are subject to the official page. Resources are for learning and exchange only — please respect the original license.