Chinese AI Agent Prompt Builder

An agent prompt is a contract, not a request

Building a reliable AI agent has far less to do with the model you pick and far more to do with the system prompt you give it. An ordinary prompt asks a model to answer one question; an agent prompt has to govern a model that acts on its own, calls tools, and takes several steps before it is done. That is a different job, and it needs a different structure. A good agent system prompt reads almost like a contract: it names the agent's role and duties, states the single goal it is responsible for, lists the tools it may use and exactly when to use them, lays out the workflow in order, draws the boundaries it must not cross, and fixes the shape of what it returns. This builder keeps that structure for you. Fill the six fields, and it assembles a clean, ordered system prompt — a leading "你是…" role line, then clearly headed sections for the goal, the tools, the workflow, the constraints and the output format — ready to paste into DeepSeek, Qwen, Kimi, Claude, GPT or the system slot of any agent framework.

The highest-leverage line is still the role, but for an agent it carries more weight, because it also sets the limits of the agent's authority. "你是一个只读的数据分析助手，绝不修改任何记录" tells the model both who it is and what it may never do, in one sentence. After the role comes the goal: a single, testable objective the agent is accountable for, so it does not wander. Then the tools, which for an agent are not optional decoration but the core of how it acts. The mistake most people make is listing tool names and stopping there. The version that works names each tool, says what it is for, and — crucially — says when to call it and when not to. "search_kb：回答事实性问题前必须先调用；不要凭记忆作答" prevents both the agent that never checks and the agent that checks everything twice. Tool discipline written into the prompt is cheaper and more reliable than hoping the model infers it.

Tools, boundaries and format are what make an agent shippable

The fields people skip and regret are the workflow, the boundaries and the output format. The workflow turns a vague goal into an ordered sequence the agent can follow and you can debug: when something breaks, you can see exactly which step it broke at. The boundaries are what make the agent safe to put in front of real users — they tell it when to stop (the task is done, or no further progress is possible), when to hand the problem back to a human, and which actions it must never take on its own. An agent without an escalation rule is an agent that will, sooner or later, confidently do the wrong thing rather than admit it cannot proceed. Spending one extra line on "遇到高风险或无法确定时，停止并交还给人" is consistently the cheapest way to make an agent trustworthy. Constraints do not limit a capable agent; they keep a capable agent from doing damage.

The output format is the field that decides whether your agent can plug into a system at all. A human can read prose; downstream code cannot. If you want the agent's result consumed by another step — a router, a database write, a UI — say so explicitly: "以 JSON 返回，字段为 category、confidence、next_action." A fixed shape turns the agent from a chatbot into a component. Because the whole output is structured plain text, the same prompt is portable across every major Chinese model and works just as well on Claude, GPT or Gemini, and just as well inside whichever agent framework you favour. And because the tool runs entirely in your browser, you can iterate freely — tighten one field, copy again, test the agent, adjust — without anything you type ever leaving your device, being sent to a model, or being stored. Treat the first version as a draft: run the agent, watch where it over-calls a tool or fails to stop, and tighten the matching field. Two or three rounds of that usually turn a flaky agent into a dependable one, and you keep a clean, reusable system prompt at the end.