Prompt in Your Own Language: Clarity Beats the English Tax

Here’s a question I keep coming back to: when I talk to an LLM, should I write in my native language, or switch to English?

The case for English seems obvious. Most models are trained on more of it. The docs, the benchmarks, the forum threads the model learned from, all overwhelmingly English. So English output should be a little sharper.

And it is. But less than you’d think.

The gap has narrowed. On the current frontier models, German, my native language, is one of the best-covered languages after English. The quality difference is low single digits, not night and day. That advantage is real, but it’s small, and it’s shrinking with every model generation.

Now weigh the other side. When I write in my own language, I think more precisely. I make finer distinctions. I catch my own vagueness. When I force myself into English, I introduce friction exactly where it hurts most: at the input.

And input is where most bad answers come from. Not the model. The prompt. A muddy question in fluent English loses to a sharp question in your mother tongue almost every time. If your native language is where you think clearly, that’s the bigger lever. Use it.

When English still wins

There are exceptions, and they’re worth knowing.

Niche or cutting-edge topics where the sources basically only exist in English. Certain deep technical details, current research. The training density is just higher there.

Smaller or older open-source models, where the non-English gap is still real and sometimes large.

And when the output has to be in English anyway. No reason to round-trip it through a translation in your head.

The pragmatic middle

My default now: write in my native language, but leave technical terms in the original. Don’t translate “context window” or “embedding” or “race condition” into German just to be consistent. You get the clarity of thinking in your own language plus the strong English anchoring the model has for those exact terms.

Which raises the obvious next question: is mixing languages a problem?

Mostly, no. Modern models handle mixed-language input fine, because their internal representation is language-agnostic anyway. A German prompt with English technical terms is closer to the normal case than the edge case.

A few things to keep in mind. The reply language follows the dominant language of your prompt, so if you want the answer in a specific language, just say so. On very small models, frequent switching can tip the output language unintentionally. And when the language itself is the subject, translation, grammar, style, keep your instruction and the text-to-process clearly separated, or the model won’t know which is which.

But as a carrier language with technical terms sprinkled in? Mix freely.

The takeaway is smaller than the debate around it. Write where you think clearest. For most people, most of the time, that’s home.