All work
AI experiments · Sep 2026
Did Anthropic fix the writing?
Anthropic said Opus 5.5 puts the point first and follows writing rules. I gave four Claude models the same 80 prompts, with and without rules, and measured it.
opus55-writing.pages.dev
- 640
- answers compared
- 443 → 0
- em dashes, Opus 4.6 vs 5.5
- 39%
- of 5.5 answers kept all five rules
01Method
80 prompts across eight kinds of task, run plain and with five checkable rules. Regex metrics plus Jev judgments for buried leads, padding, and hedging. Every answer is readable side by side.
02Result
The style changed: zero em dashes and much less padding. Rule-following did not. Opus 5.5 kept all five rules less often than Opus 4.6, mostly by running past the word limit.