I ran the same test on the two "write less" skills people reach for, Caveman and Ponytail, against the same baseline. Same story. Caveman was +41% over baseline on Opus 4.8, Ponytail +17% on Haiku.

What reliably cut output was removing whole sections. Answer first. Kill the restated question, the narration, the end-of-turn recap, the generic "there may be edge cases." Only claim a test passed…

I told 5 Claude models to "be concise." Two of them generated MORE tokens than if I'd said nothing at all. Opus 4.8 went up 42%. I benchmarked it. Here's what actually cuts AI output 👇