Chinese System Prompt Builder

A system prompt is a standing role, not a one-off question

A system prompt — the 人设, or persona — is the standing instruction a Chinese large language model runs under for an entire conversation or an agent's whole lifetime. It is fundamentally different from an ordinary chat prompt: a chat prompt asks one specific question, while a system prompt defines who the assistant is, what it is there to do, how it should behave, and what it must never do. On DeepSeek, Qwen (通义千问), Doubao, Kimi, ERNIE or Zhipu, the system prompt usually lives in a dedicated "System" field and takes priority over every user turn, which is exactly why it is the right place to govern long-term behaviour. This builder keeps that structure for you: fill the fields and it joins them into a clean persona that opens with a "你是…" identity line, followed by clearly headed sections — goals and duties, behaviour rules, voice, boundaries, background — each prefixed with a Markdown-style heading the model can read at a glance.

The single highest-leverage line is the role and identity. "你是公司官方客服助手 小安" steers the model's viewpoint, vocabulary and stance for the rest of the session in one sentence — far more efficiently than a paragraph of adjectives repeated every turn. After the role, the goals and duties describe what the assistant is there to accomplish over time, not what you want from a single message, and the background-knowledge field lets it assume the facts it needs without you re-pasting them each turn. Because all of this is set once and then persists, a few minutes spent making each field concrete pays off across every conversation the persona ever drives.

Boundaries and behaviour rules turn a chatbot into a dependable assistant

The fields that turn a demo into a dependable assistant are behaviour rules and boundaries. Behaviour rules convert vague hopes into executable steps — "confirm the user's need before answering", "quote the policy text", "say 不确定 when unsure" — so the assistant acts consistently instead of improvising. Boundaries and prohibitions are the part most people skip and most regret: "never promise prices", "no medical or legal diagnoses", "hand off to a human when out of scope". In a one-off chat prompt these feel optional, but in a system prompt they are the safety rail that lets you put the assistant in front of real users, because they apply on every turn, not just the one you remembered to type them in.

Because the output is structured plain text, the same persona is portable across every major Chinese model and works just as well as the system message in an agent framework, or on ChatGPT, Claude or Gemini. Write it in Chinese when you want natural, idiomatic Chinese answers; the structure travels regardless of language. And because the whole tool runs locally in your browser, you can iterate freely — tweak one field, copy again, and test the persona in a fresh chat — without anything you type ever leaving your device, being sent to a model, or being stored. Treat the first persona as a draft: deploy it, watch where the assistant drifts or oversteps, and tighten the matching rule or boundary. Two or three rounds usually turn a loose chatbot into an assistant that stays in role, and you keep a clean, reusable system prompt at the end.