Caveman
Claude Code plugin that forces ultra-compact responses: about half the output tokens at the median on ten questions (repo eval, length only), 8.5% fewer output tokens on 86 real coding tasks (JetBrains, July 2026, quality flat). Three compression levels: lite, full, ultra.
Strengths
- Shorter answers from install, without changing your workflow: 8.5% fewer output tokens on 86 real coding tasks (JetBrains, July 2026, quality flat)
- Three adjustable levels depending on context (lite for reviews, ultra for pure code)
- One-command install, zero configuration after
Limitations
- Compressed responses can lose readability for beginners
- Doesn't affect reasoning tokens, real session gain is ~4-10%
- Its own rules cost about 1,000 to 1,500 input tokens every turn, per its repository: on a session of short questions, the saving can vanish
Best for
- Devs who regularly hit their Claude Code session limits
- Long vibe-coding sessions where narrative responses slow the flow