What problem does it solve? Large project memory files like CLAUDE.md, todo lists, and preference notes are loaded into context on every session start, repeatedly consuming input tokens. This Skill compresses those natural language files into a terse caveman format while preserving code blocks, URLs, headings, and file paths exactly, cutting token usage by roughly 40-60% on benchmarked fixtures. ## Core Features & Use Cases - Token-reducing compression: Rewrites prose by removing articles, filler, hedging, and pleasantries while keeping technical terms, commands, and structure intact. - Safe in-place overwrite with backup: The original is saved byte-for-byte to an out-of-tree data directory as <filename>.original.md, so skill auto-loaders never re-ingest the backup. - Structural validation with targeted repair: Headings, code blocks, URLs, file paths, and inline code are validated after compression; only broken parts are sent back for fixing, never a full recompression. - Use Case: A developer with a 1,100-token CLAUDE.md runs the skill once, gets a ~600-token compressed version loaded every session, and edits the human-readable backup whenever the source needs updating. ## Quick Start Ask the assistant to run /caveman-compress on your CLAUDE.md file to shrink it and keep a readable backup.
