Sounds like a meme. It isn’t.
Caveman is a meta-prompt that strips grammar, hedging, and filler from every reply. The reasoning is untouched — only the output is compressed.
In testing on a routine workflow: ~800 words → ~250 per reply, no useful information lost.
A March 2026 paper, “Brevity Constraints Reverse Performance Hierarchies in Language Models,” found that constraining outputs can actually improve accuracy on certain benchmarks, not just reduce cost.
The angle most write-ups miss:
This isn’t a token-cost optimisation. It’s a signal-to-noise optimisation. In multi-turn agentic workflows, every reply becomes context for the next turn. Compressing output raises the SNR of every downstream prompt — automatically.
Use for: research, analysis, evals, agent chains.
Skip for: creative writing, customer-facing copy, anything tone-dependent.
Repo: <a href="http://github.com/JuliusBrussee/caveman" target="_blank" rel="noopener nofollow">github.com/JuliusBrussee/caveman</a>

Sharing hands-on development experience and insights

Agentic DevOps

Caveman is a meta-prompt (Claude Code skill) that forces the model to respond in stripped-down, grammar-free 'caveman' language, reducing output tokens by ~65% without losing reasoning quality. The key insight is that this isn't just a cost optimization — it's a signal-to-noise ratio improvement for multi-turn agentic workflows, where compressed outputs become cleaner context for downstream prompts. A March 2026 paper supports the idea that brevity constraints can actually improve accuracy on certain benchmarks. Best suited for research, analysis, evals, and agent chains; not recommended for tone-sensitive or customer-facing content.

Caveman: a Claude Code skill that cuts ~65% of output tokens by making the model talk like a caveman.