Agents talk to me tersely by default: no greetings, no restating the question, no hedging, fragments are fine. It cuts reply tokens by roughly three quarters and, more to the point, I read faster. The skill and its two hooks come from JuliusBrussee/caveman (Apache-2.0 / MIT); my copies are in claude/skills/caveman and claude/hooks.
| Level | Example |
|---|---|
| lite | "Your component re-renders because you create a new object reference each render. Wrap it in useMemo." |
| full (my default) | "New object ref each render → re-render. Wrap in useMemo." |
| ultra | "Inline obj prop → new ref → re-render. useMemo." |
Switch with /caveman lite|full|ultra, or say "normal" to turn it off.
What stays normal, always:
- code, commands and commit messages;
- security warnings and anything about data loss;
- text that leaves the session: READMEs, issue comments, release notes. That goes through the humanize rules instead: clear, human, no em dashes;
- teaching. When I'm learning something (studying, a new codebase) the agent switches to a patient mentor voice and walks through it step by step.
The companion skills: caveman-commit (short conventional commits), caveman-review (one-line review comments with exact line references), caveman-compress (rewrites memory files like CLAUDE.md in caveman form to save input tokens, keeping a readable backup).