Why I started writing AI agent tasks in Russian.
Yesterday, I pushed more commits in half a day than I had during the entire previous week. Part of the reason was a new agent loop I had set up. A cycle in which an agent receives a task, implements it, verifies the result, and moves on to the next one.
But an agent loop is only as good as the tasks you feed it. And my Linear tasks suddenly got better when I started writing them in Russian.
This was strange. I started learning English when I was twelve. I have lived in the United States for the past ten years, and I can speak fluently about any subject, including science and philosophy. So why were my tasks noticeably clearer when I wrote them in Russian?
You can speak a second language fluently and still struggle to recall the right word or sense when a sentence is starting to lose precision.
In conversation, this kind of ambiguity often doesn’t matter. The person you are speaking with understands the surrounding context and can ask for clarification when something doesn’t quite add up. An LLM can also infer what you mean or ask a follow-up question. But in an agent loop, a plausible yet incorrect interpretation turns into a faulty assumption, unnecessary code, and an additional iteration.
Before I started writing tasks in Russian, I sometimes asked the agent to confirm that it had understood my request. This helped uncover ambiguities, but it took a lot of time.
Before LLMs, sharing a technical idea often meant first expressing it in English. As a result, many people were translating their thoughts before they had fully finished developing them.
Now you can reverse that order. Think through the task in your native language and give it directly to an LLM. The LLM can still implement the code and write the documentation in English. A Linear task or design document can be translated at the output stage, when it needs to be shared with a colleague or a broader audience. Before LLMs, translating long passages of technical writing was either time-consuming or expensive. With LLMs, it is nearly effortless.
The ability to express your thoughts in your native language may seem like a minor improvement. I’m starting to think it isn’t. Engineering work often begins with defining the problem, and the quality of the thinking at this stage affects almost everything that follows.
English will remain the lingua franca for the foreseeable future. But it may cease to be a language we have to think in and become primarily a language for sharing what we have already thought.
Try describing your next complex task to an LLM in your native language: the context, the constraints, and the expected result. Don’t translate it in advance. See whether the task becomes clearer. Mine did.
P.S. Written in Russian with ❤️. Exported to English by an LLM.