What is a system prompt?

Every time you talk to a chatbot like ChatGPT, Claude or a personal assistant, the model is reading more than your message. Before your words ever reach it, a separate block of text called the system prompt has already told the model who it is, how to behave and what it is allowed to do. It is the difference between a raw language model and a coherent, well-behaved assistant.

What is a system prompt, exactly?

A system prompt is a special instruction placed at the very start of an AI conversation that defines the assistant's identity, behaviour and constraints. Chat models process messages in roles: a 'system' role for the standing instructions, a 'user' role for what you type, and an 'assistant' role for the model's replies. The system prompt occupies that first system role and applies to every turn that follows, even though you usually never see it.

Think of it as a job briefing handed to a new employee on day one. It might say: 'You are a friendly support agent for an airline. Be concise. Never invent flight times. If you are unsure, say so and offer to escalate.' The model then interprets each customer question through that lens. Change the briefing and the same underlying model behaves like a completely different assistant.

A typical system prompt covers several things at once:

• Identity and persona, such as the assistant's name, role and personality.
• Tone and style, for example formal, playful, terse or thorough.
• Rules and guardrails, including what topics to avoid and how to handle uncertainty.
• Available tools, so the model knows it can search the web, send email or create files.
• Context and memory, like the user's preferences or facts worth remembering.
• Output format, such as 'answer in plain language' or 'return valid JSON'.

How does a system prompt shape behaviour and persona?

A system prompt shapes behaviour because language models predict their next response based on everything in their context window, and the system role sits at the top with the most steering weight. Instructions there act like persistent defaults: the model carries 'be concise' or 'never reveal internal data' through dozens of turns without being reminded. This is why two assistants built on the identical underlying model can feel worlds apart, one warm and chatty, the other clipped and clinical.

Persona is the most visible effect. A line such as 'You are Miyo, a calm and capable personal assistant' gives the model a stable voice. Behavioural rules are subtler but just as important: telling the model how to handle requests it cannot fulfil, when to ask a clarifying question, or how to treat information that came from an untrusted source. A well-written system prompt is the single biggest lever over how trustworthy and useful an AI feels in practice.

System prompt vs user prompt

It helps to see the two side by side. The user prompt is what you actively type each turn; the system prompt is the fixed briefing that frames all of it.

How does MiyoMind use a per-user system prompt?

MiyoMind builds a fresh system prompt for every individual user rather than shipping one generic instruction set to everyone. When you set up your assistant, your chosen persona, tone, role and any long-term memories MiyoMind has learned about you are assembled into a personalised system prompt. That instruction is what your assistant, Miyo, reads before each reply across WhatsApp, Telegram, Discord and the web dashboard, so it stays consistent everywhere you talk to it.

This per-user approach is why MiyoMind can feel genuinely tailored. The system prompt tells Miyo which connectors you have linked, like Gmail or Google Calendar, what you have asked it to remember, and how you prefer it to communicate. As your memories and preferences change, MiyoMind refreshes the prompt so the assistant keeps up without you re-explaining yourself.

Because the system prompt can carry sensitive context, MiyoMind treats it as a security boundary. User-supplied text is sanitised and wrapped in clearly marked trust-boundary tags so that content from documents, search results or messages cannot quietly overwrite the standing instructions, part of a layered prompt-injection defence that runs on every message. Memories and connector credentials are encrypted at rest, and on paid plans your assistant runs inside a dedicated, sandboxed container.

In short, the system prompt is the quiet contract between you and the model. MiyoMind makes that contract personal: it turns your preferences, persona and memory into the instructions Miyo follows, so the assistant behaves like yours, not like a stranger's, on whatever app you reach for.