Caveman plugin: 110,000 star explosion "token saving artifact", AI answer cut off two-thirds, code registrationCaveman 插件：11 万星爆火「省 token 神器」，AI 回答砍掉三分之二，代码照准

Too much AI programming assistant is just a pain point: explain a large section, the token burns fast, the context window is squeezed by nonsense, and the code that really needs to be changed has no space to put it.

Caveman's idea was simple and brutal: add an output discipline to the agent: "If you can say three words clearly, don't use thirty words." In the same React re-rendering question, the normal answer is 63 tokens, Caveman mode 20 tokens, and the fix is exactly the same.

It is not just a prompt word: the built-in agent will compress the log, JSON, and test output before entering the model, and then cooperate with the output to slim down, and save both ends together.

The official ran the Claude Code full session 54 times: the input token decreased by an average of 33.2%, 18 questions were all correct, and the answer quality did not decline.

Auto gear will automatically pick the model and inference strength according to the task difficulty, 85 coding tasks compared to Opus 5.5, the highest grade, cost saving 54%, the score is the same.

The project was cited by Adobe Research papers, JetBrains did A/B testing, and the official Elasticsearch blog was specially reengineered and replicated, topping the Hacker News and GitHub Trending double list.

Say less: The answer is extremely simple and the code remains accurate; the Ultracave and Megacave gears can continue to press to 14 and 13 tokens.

Read less: logs, CSV, test output advanced compression pipeline feed to the model; Blocks: scripts written by the agent will be retained and reused, and the test output will be reduced from 397 token to 76 at a time.

Waste fixes and memory modules are responsible for finding out where your tokens are spent and patching them automatically. Each stone has a separate switch, so you can turn it off if you don't want to use it.

The download button provides the Caveman v4.0.1 full plug-in package (zip, 8.79 MB), containing all the skill files, the Windows installation script install.ps1, the browser extension source code, and the CLI.

Minimal installation: After unzipping, run install.ps1, or directly execute npx @ caveman-ai/cli @ latest, it will automatically find the agent configuration of the machine, and it will take effect after confirmation.

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.