AI BriefingCaveman - Saving Claude/Codex Tokens with Caveman-Speak
Key point
It's a plugin that uses 'caveman-speak' in Claude Code and Codex to reduce output tokens by up to 75%.
Details
Caveman is a plugin that converts LLM responses into concise 'caveman-speak,' reducing output token usage by about 65-75%.
It focuses on removing filler unrelated to the technical core, such as greetings, preambles, and unnecessary recommendation phrases, while preserving key information like code blocks, technical terms, and error messages to maintain technical accuracy. However, this approach only targets output tokens and does not affect the model's thinking/reasoning tokens.
Compression intensity can be adjusted across three levels:
- Lite: Maintains grammar while removing only unnecessary words
- Full: Omits articles and is composed mainly of short sentences
- Ultra: A maximally compressed telegram-style approach
Actual benchmark results showed high efficiency in technical explanations such as React re-rendering explanations (87% reduction) and PostgreSQL configuration (84% reduction), providing benefits of faster response speed and cost savings.
Installation is done via the npx skills add JuliusBrussee/caveman command, and it can be applied throughout sessions in both Claude Code and Codex environments.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Coding Agent Tool Selection Analysis: Claude Code Relies on Prior Knowledge, Codex Leverages Web SearchDiagram Design Skill for AI Agents Supports 39 Diagram Types