One way to prevent obvious AI language from sneaking in text destined for other human beings is to ask the model to produce ASD-STE100 Simplified Technical English bullet points.
This will result in a list of sentences that are clear and explanatory, easier to double-check, and convenient for the user to rewrite into a more readable format with a human voice.
I've been thinking about how to avoid having AI-isms creep into my own voice, because I see that way too much in my tech-world friends. It's natural enough: if you spend a lot of time talking to somebody, you usually end up mirroring some of their speech patterns.
One idea I had was to start doing my Claude sessions in German. (That happens to be my strongest second language, no other reason.) Thus any Claude-isms I pick up will presumably not wander into my English, which is what I use for most communications.
This would have the added benefits of getting some much-needed German typing practice, and preventing me from ever lazily copy-pasting AI text into an email or a README.
+1 on this being a great idea. I also need to learn how to write in German (I can speak at B2 and read C1, but I find writing difficult as I've never bothered tbh). The added benefit of avoiding Claudisms is interesting. I think I'll join you on this.
FWIW I did a day of German-language software design with Claude and it was great, a bit strenuous of course but I also wonder if it might have found some bugs it would have overlooked in English.
If you're doing it for writing practice, one of the good/bad things is that it's quite typo-tolerant (in any language). So you can have the chat with less effort, at the risk of practicing bad habits.
Yeah, I tried it for about 2 hours yesterday. It was indeed strenuous when having to think about both the language and the complicated topic being discussed.
I was asking it to correct my wording as we went along, and despite some weird stuff now and then, e.g. "Abfrage" for query, it was quite good.
I think benchmarking some tasks in both languages would be quite cool. I remember about a year or two ago benchmarking tasks in different languages was quite popular.
This will result in a list of sentences that are clear and explanatory, easier to double-check, and convenient for the user to rewrite into a more readable format with a human voice.