I benchmarked Claude Code's caveman plugin against "be brief."

5 hours ago

The Caveman compression plugin was benchmarked against the simple prompt 'be brief.' across 24 prompts in six categories.
Both 'be brief.' and Caveman performed similarly in token reduction and quality, with no significant differences in correctness scores.
Caveman's 'Ultra' mode unexpectedly produced longer responses in some categories due to its Auto-Clarity rule, which disables compression for safety warnings and multi-step sequences.
Caveman's main value lies in providing consistent output structure, intensity adjustments, and persistence across sessions, not just token savings.
The benchmark revealed minor issues: one missed required term in 'Lite' mode and tool-use behavior triggered in 'Ultra' mode.
For basic compression, 'be brief.' is sufficient; Caveman is recommended for structured and consistent outputs in extended interactions.

Hasty Briefsbeta