AI

System prompt

The persistent set of instructions an AI agent gets at the start of every conversation: who it is, what tone to use, what rules to follow. The character sheet for the agent.

What it means

A system prompt is a special kind of prompt that sits at the top of every conversation an AI agent has. It defines the agent's persona, role, tone, capabilities, and constraints, and stays constant across the conversation regardless of what the user says.

A typical system prompt covers: who the agent is ('You are an assistant for a Singapore aesthetic clinic'), how to behave ('Reply in a calm, professional tone, use British English'), what tools it has access to ('You can check appointment availability and book slots'), and what NOT to do ('Never quote prices not in the price sheet, never give medical advice, escalate to human if asked').

Why it matters

Almost every guardrail in a production AI agent lives in the system prompt. Quality, safety, brand consistency, escalation rules: all of it is system prompt design.

The system prompt is also the most cost-efficient lever. Refining it is fast and free. Switching LLM models is expensive and slow. Most agent improvement happens at the system prompt layer.

Example

A clinic's first agent has a 200-word system prompt. Customers occasionally get strange responses on edge cases. The team expands to a 1,500-word system prompt with explicit examples of edge-case handling, brand voice samples, and a clear escalation rule. Edge-case behaviour stabilises within a day, no model change required.

Where this comes up

← Back to all terms