AI Briefingcaveman, performance comparison with 'be brief'
Key point
In a benchmark of 24 dev prompts, 'be brief.' performed almost identically to caveman.
Details
Across 24 dev prompts and 6 categories, baseline, be brief., and caveman lite/full/ultra were compared.
The average scores, graded by a separate Claude using per-prompt rubrics, were as follows.
- baseline: 0.985, 636 tokens
- be brief.: 0.985, 419 tokens
- caveman lite: 0.976, 401 tokens
- caveman full: 0.975, 404 tokens
- caveman ultra: 0.970, 449 tokens
As a result, the two-word be brief. maintained nearly the same quality as the caveman variants while significantly reducing token count. caveman is still useful for output structure consistency, mode switching, and safe escapes from destructive ops, but compression itself was not the key differentiator. Output variance was also high due to the safe-escape branching.
The benchmark harness has been open-sourced.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.