How to Write AI Prompts: hh.ru's Production Practice
hh.ru explained why a production prompt looks more like code than a conversation with ChatGPT. The team advises writing instructions in English, handling few-shot examples carefully, explicitly forbidding risky answers, accounting for the model's knowledge cutoff date, and testing the system on large datasets and real logs rather than a couple of successful cases. The company says a single bad answer in production can reach thousands of users, so a prompt has to be treated as an engineered system.
AI-processed from Habr AI; edited by Hamidun News
hh.ru has shared its practice for writing AI prompts in its production services, where a single bad answer can reach thousands of users.
How do good AI prompts differ from chatting with a chatbot?
Production has none of the forgiveness of an ordinary chat. A single bad answer can reach thousands of users, break a user flow, create reputational risk or hurt conversion. That is why a prompt is a system with constraints, tests and constant debugging.
Why does a neural network in production need special prompts?
In everyday use, you can restart an LLM conversation. In production, a single bad answer can reach thousands of users, break a user flow or create reputational risk.
How should you write AI prompts for business?
According to hh.ru's experience, production prompts are not a conversation with a chatbot but an engineered system with constraints, tests and constant debugging.
How do production prompts differ from ordinary use?
In everyday use, LLMs are forgiving: the user asks a question, clarifies, rewrites and moves on. In a product, a single bad answer can reach thousands of users, break a user flow, create reputational risk or hurt conversion. That is why a production prompt is an engineered system with constraints, tests and constant debugging.
What is the difference between a prompt in a chatbot and in production?
In everyday use, a user can rephrase a question and restart the conversation. A product has no such luxury: a single bad answer can reach thousands of users, break a user flow or create reputational risk.
Key Questions Answered
How do good AI prompts differ from chatting with a chatbot?
Production has none of the forgiveness of an ordinary chat. A single bad answer can reach thousands of users, break a user flow, create reputational risk or hurt conversion. That is why a prompt is a system with constraints, tests and constant debugging.
Why does a neural network in production need special prompts?
In everyday use, you can restart an LLM conversation. In production, a single bad answer can reach thousands of users, break a user flow or create reputational risk.
How should you write AI prompts for business?
According to hh.ru's experience, production prompts are not a conversation with a chatbot but an engineered system with constraints, tests and constant debugging.
How do production prompts differ from ordinary use?
In everyday use, LLMs are forgiving: the user asks a question, clarifies, rewrites and moves on. In a product, a single bad answer can reach thousands of users, break a user flow, create reputational risk or hurt conversion. That is why a production prompt is an engineered system with constraints, tests and constant debugging.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.