Caveman-speak AI prompts fail to deliver promised token savings in real coding tasks

Developers increasingly scrutinize the operational costs of AI coding assistants as billing models shift toward usage-based pricing tied to specific models. The search for ways to shrink token consumption has led to experimentation with a technique known as "caveman mode"—stripping articles, formal grammar, and courteous language in favor of terse fragments, theoretically forcing AI systems to cut unnecessary verbosity.

The viral caveman prompt#

A free Claude Code skill on GitHub built by developer Julius Brussee gained significant traction, accumulating tens of thousands of stars and generating high-profile discussions on platforms like Hacker News. The tool's central claim: a dramatic 65% reduction in output tokens. Its marketing pitch, delivered in character, promised that removing conversational filler would deliver comparable results using far fewer words.

Precedent existed for the concept. Elastic had previously developed a similar approach for Elasticsearch queries and reported a 63.6% token reduction across internal test scenarios, demonstrating measurable savings when prompt engineering focused on direct outputs like index names and query syntax rather than explanatory prose.

Testing reveals modest real-world gains#

JetBrains decided to rigorously evaluate the Brussee skill rather than accept its headline figure. Engineer Denis Shiryaev used Harbor, an open-source agent evaluation framework, paired with tasks from SkillsBench, to benchmark the technique across 86 real coding tasks on Claude Code.

The results fell dramatically short of expectations:

• Advertised savings: 65%
• Measured savings: 8.5% (at scale)
• Early testing: approximately 30% reduction on smaller task batches

Why the gap?#

Shiryaev identified a fundamental mismatch between the skill's design and how production agentic systems actually operate. In casual chatbot exchanges, the kind likely used to derive the original 65% claim, conversational padding dominates the token stream. In real coding workflows, however, the bulk of output tokens consist of code, diffs, tool invocations, and exact error messages—material the skill leaves unchanged.