Compresses Markdown and plain-text memory files (CLAUDE.md, todos, preferences) into a concise 'caveman' style to reduce token usage for LLMs. It preserves code blocks, inline code, URLs, file paths, and technical content verbatim, and backs up originals as <filename>.original.md. The process runs via a scripts/ CLI that calls Claude, validates output, and retries targeted fixes when necessary.
Use when you need to reduce token consumption for large text-based memory or documentation files before sending them to an LLM, or to create compact backups for storage. Ideal for agents that frequently re-process long markdown notes or project documentation.
Best suited for Claude-family and Claude Code integrations (the skill explicitly uses Claude in its process), and any agent workflows that need token-efficient memory handling.
Caveman Compress compresses natural-language memory files into a compact 'caveman' format to save LLM tokens, preserving code blocks, URLs, and structure. Scripts are designed as a Python package (run via `python3 -m scripts`) so standalone execution fails with ImportError on relative imports, but the package architecture is sound. Security is strong — explicit sensitive-path detection prevents shipping credentials to the Anthropic API. Requires ANTHROPIC_API_KEY or Claude CLI to function.
anthropictiktokenWell-structured skill with good security awareness (sensitive file detection, file size limits, backup before overwrite). The relative import failures are expected when scripts are extracted and run individually — they're designed as a package. No security concerns.
