Developers tired of Claude Code’s verbose pleasantries now have a radical solution: Caveman, a plugin and skill that forces the LLM to strip away greetings, hedging, and recaps in favor of raw facts. Version 3.1.0 of the plugin, available via the JuliusBrussee/caveman repository, introduces three distinct modes—caveman, ultracave, and megacave—designed to minimize output token consumption without sacrificing the integrity of code blocks or technical terms.

The Mechanics of Brevity

The core of Caveman is a skill file that dictates a strict writing style: answer first, reason second, next step third. Articles like "a" and "the" are optional, but critical semantic words like "not," "never," and "only" must remain. Crucially, the plugin enforces "payload verbatim" for code blocks, commands, and file paths, ensuring that while the prose shrinks, the executable logic remains character-for-character accurate. The README illustrates this with a React re-rendering example, where a 60-word explanation is compressed into a 12-word directive: "New object ref each render, so React re-renders. Wrap the prop in useMemo."

Benchmarking the Token Savings

According to the repository’s own eval on claude-opus-5-5, the impact is measurable but nuanced. Across ten developer questions, a standard reply totaled 6,983 output tokens. A simple "Answer concisely" prompt reduced this to 4,334 tokens. The /caveman mode achieved 4,119 tokens, offering a marginal 3% improvement over the concise baseline. However, the /ultracave mode, which fragments grammar further, dropped output to 2,693 tokens—a 35% reduction from the concise baseline. The /megacave mode, which answers in Classical Chinese, showed only a 9% median reduction, with the project noting that token savings in this mode are "small and noisy."

Installation and Configuration

Users can install the plugin directly into Claude Code using claude plugin marketplace add JuliusBrussee/caveman && claude plugin install caveman@caveman. The plugin automatically activates caveman mode in every session unless configured otherwise via the CAVEMAN_DEFAULT_MODE environment variable or a user config file. For those who only want the skill without hooks, npx skills add JuliusBrussee/caveman -g provides a lighter installation. The plugin also includes utilities like /caveman-commit for one-line Conventional Commit messages and /caveman-compress for shrinking memory files like CLAUDE.md, which the repository claims reduces file size by about 46% across test fixtures.

Key Takeaways

  • /ultracave mode offers a 35% reduction in output tokens compared to standard concise prompting, making it viable for cost-sensitive or speed-critical agent workflows.
  • The plugin does not reduce input tokens; it only shortens the model's response, meaning the bulk of context costs (system prompts, file reads) remain unchanged.
  • Built-in exceptions automatically revert to plain prose for security warnings, irreversible actions, and step sequences to prevent ambiguity.
  • The optional @caveman-ai/cli proxy can compress input data (logs, JSON) but may disable Claude Code Remote Control and sends telemetry by default.

The Bottom Line

Caveman is a necessary corrective for terminal users who value speed over ceremony, but its benefits are strictly limited to output tokens. For most users, the default caveman mode offers the best balance of clarity and brevity, while ultracave should be reserved for rapid, low-context checks where ambiguity is low.