What problem does it solve? Long natural-language memory files (.md, .txt) consume excessive input tokens when loaded into an LLM context. This Skill rewrites prose into terse caveman-style fragments while preserving all technical substance, cutting token usage without losing meaning. ## Core Features & Use Cases - Inline LLM Compression: The agent itself rewrites prose (no external API call), removing articles, filler words, and hedging while keeping code blocks, URLs, paths, and technical terms exactly intact. - Safety Pipeline: Pre-flight detection skips code/config files, a sensitive-path denylist blocks credentials and keys, originals are backed up as <filename>.original.md, and a validator checks headings, code blocks, URLs, and bullet counts after compression. - Benchmarking: A benchmark script measures token savings per file pair using tiktoken (or word count fallback) and reports validation status. - Use Case: You maintain a large CLAUDE.md or project memory file that eats thousands of tokens each session. Run the compress command on it to shrink it by a significant percentage while keeping every command, link, and code snippet verbatim. ## Quick Start Ask the agent to compress a memory file, for example: run /caveman:compress on docs/project-memory.md to shrink its token footprint.
