What it is
Caveman is a Claude Code plugin. A hook intercepts every response and strips the noise: articles (a/an/the), filler (just, really, basically), pleasantries (sure, of course, happy to), and hedging. Fragments are fine. What stays untouched is everything that carries meaning — technical terms, exact error strings, and code blocks. The result reads blunt but loses zero information.
The pattern
Every answer collapses to [thing] [action] [reason]. [next step]. Instead of “Sure! I’d be happy to help — the issue is likely caused by…” you get “Bug in auth middleware. Token expiry uses < not <=. Fix:”
Three levels
- lite — trims the worst filler, keeps sentences.
- full — fragments, dropped articles, hard compression.
- ultra — maximum density, near-telegraphic.
The skill bundle
- cavecrew — subagents: a surgical 1–2 file builder, a read-only code investigator, a diff reviewer.
- caveman-commit — terse, conventional commit messages.
- caveman-compress — squeeze an existing block of text.
- caveman-review — blunt, severity-tagged code review, no praise.
- caveman-stats — how many tokens the mode saved you.
The safety valve
Caveman drops the act exactly where terseness is dangerous: security warnings, irreversible-action confirmations, and multi-step sequences where a clipped fragment could be misread. Commits, PRs and code are always written normally. You get the density without the misfires.