System Prompt

A Prompt is the instruction text that a user or system inputs to a large model, determining how the model understands the task and the quality of its output.

System prompts are a layer of instructions preset by vendors and invisible to users, injected into the model context before the user's conversation begins, used to shape the model's behavioral boundaries, response style, and safety policies.

A large portion of system prompts is typically dedicated to setting safety guardrails, including refusing to generate harmful content, avoiding dangerous operation instructions, protecting privacy, and preventing jailbreak attacks.

The system prompt is the highest-priority instruction layer in large language model conversation architecture, injected before user messages and influencing how the model processes all subsequent inputs.

Without human guidance, even with strongly configured system prompts, AI agent behavior tends toward two extremes: rapid polarization or polite consensus loops.

A tool schema contains a functional description, parameter names, types, meaning explanations, and a required field indicating whether a parameter is mandatory.

商业AI产品的系统提示词长度从几百token到数万token不等

生产级系统提示词长度可达数千甚至上万token

在企业级应用中，一个生产可用的系统提示词往往需要经过数十次迭代，涵盖角色定义、行为边界、异常处理、输出格式等多个维度

Constraints relying solely on system prompts often prove ineffective when faced with sufficiently sophisticated attacks.

When a Handoff occurs, the receiving agent starts a new run with its own system prompt and inherits the prior user messages and handoff record, but does not inherit the original agent's system prompt (unless explicitly passed).

The system prompt is an input fully controllable by the user; it neither constitutes the model's security boundary nor can it serve as evidence of the model's intent.

A carefully designed system prompt can significantly change output style, because it activates genre conventions already encoded by a large amount of corpus.

The "instruction-conversation genre" is jointly shaped by three key elements: instruction fine-tuning, the system prompt, and a structured input format with special tokens and role assignments.

The system prompt sets the role, boundaries, and tone before the formal conversation begins, and is one of the formal prerequisites for the "instruction-conversation genre" to hold.

In the tool-calling loop, the model is only responsible for decision-making and expression (deciding what to call and what parameters to pass), while actual execution is done by deterministic code, which both avoids the security risk of the model directly operating external systems and makes the calling process auditable and retryable.

Replacing the system prompt directly affects model behavior, so abuse and compatibility issues need to be watched, and compatibility must be verified across differences such as subscription tier and configuration environment.

In terms of load order, Rules and Skills usually take effect before or independently of the system prompt, so replacing the system prompt does not automatically disable them.

Coze prompts are divided into system prompts and user prompts: the system prompt defines the role, skills, constraints and output mode, while the user prompt is the actual question or instruction input.

An overly long system prompt occupies the context window and affects the model's understanding of user input.