A skill that shortens Claude’s answers — and the claim I could not prove

Claude’s default answers are prose-heavy. Ask a narrow question and you often get several paragraphs where two sentences would do. /caveman is a small Claude Code skill that compresses that prose — same technical content, far fewer words, and funnier than it has any right to be.

What it actually touches

It rewrites the model’s prose. It does not touch code blocks — those come back intact, which is the whole reason it is usable while you work. There are intensity levels; the middle setting is the one worth using, because the most aggressive setting starts costing clarity for laughs.

The claim I expected to make

The natural pitch writes itself: shorter answers mean fewer output tokens, fewer tokens mean a smaller bill. I set out to measure that so the number would be real rather than asserted.

Why it did not survive measurement

It failed, and the reason is structural. In a real session the dominant cost is not the model’s replies — it is everything you send in. Files, prior turns, tool output, the accumulated context. Compressing the answers trims the smaller side of the ledger while the larger side is unchanged. The saving existed but was lost in the noise of base overhead.

• 01True: the answers get materially shorter
• 02True: code blocks are left alone
• 03Not demonstrated: a meaningful reduction in total session tokens

What I published instead

The honest version: this makes answers shorter. It is not a cost-saving tool. That is a weaker headline and a claim that survives someone checking it — which matters more, because the moment a developer measures your number and finds it wrong, nothing else you say gets the benefit of the doubt.