I've had a lot of success combating this by adding "All summaries need to adhere to ASD-STE100 Simplified Technical English standards" [0] which I discovered from another HN thread [1]

[0] https://www.asd-ste100.org/ [1] https://news.ycombinator.com/item?id=49114639

I actually tried to switch Claude's `output-style` to ASD-STE100 Simplified Technical English for the prose it writes (code comments, commit messages), but it didn't seem to help. Figured I'll just use GPT 5.6 Sol for implementation instead of trying to "socialize" Opus 5. Anthropic will have to address it sooner or later if they want people to keep using it.

Using ASD-STE100 is in my system prompt, not just CLAUDE.md which basically never gets followed.

I actually converted it from the PDF into an explicit skill with the word list inline as well as the main rules.

Works a little bit, but give it a bit of context and the model’s training takes over and it starts talking like a dictionary-huffing crack addict again.

It’s a model problem and no amount of harness hacking seems to fix it.