Teaching Claude to grunt in fragments saves two-thirds of your tokens

What it does

caveman is a prompt-engineering skill for Claude Code and roughly 30 other coding agents that constrains the model to telegraphic, fragment-heavy output. It strips pleasantries and filler while preserving technical accuracy—turning a 69-token React explanation into 19 tokens. The project ships as a skill file plus session hooks, with extras like a memory-file compressor and terse commit-message generator.

The interesting bit

The author benchmarked against a plain “Answer concisely” baseline rather than the default verbose mode, so the 65% average savings claim is honest rather than cherry-picked. The README even cites a March 2026 arXiv paper suggesting brevity constraints can improve accuracy by 26 points on certain benchmarks—less word, more correct.

Key highlights

• Four compression tiers: lite, full, ultra, and wenyan (classical Chinese).
• Real measured savings vary wildly by task: 87% on a React error boundary, 22% on a callback refactor, averaging 65% across ten prompts.
• caveman-compress rewrites your memory files (e.g. CLAUDE.md) into the same terse dialect, cutting input context by about 46%.
• Companion commands for one-line commit messages, one-line PR reviews, and a running statusline badge of lifetime tokens saved.
• Explicitly does not touch thinking or reasoning tokens—only the final output “mouth” is smaller.

Caveats

• Savings are task-dependent: architecture explanations and async refactors barely shrink, while bug fixes and Docker builds compress dramatically.
• Only output tokens are affected; reasoning tokens remain unchanged, so the wall-clock win depends on how much time you spend reading versus waiting for the model to think.
• Auto-activation is built-in for Claude Code, Codex, and Gemini; Cursor, Windsurf, Cline, and Copilot require --with-init flags, and the remainder of the 30+ supported agents need a manual /caveman trigger each session.

Verdict

Worth a look if you pay per token or simply hate wading through AI-generated pleasantries to get to the useMemo recommendation. Skip it if you need hand-holding prose or your agent bills are already flat-rate.