A system prompt is the first message in a conversation with an AI model, written by whoever built the tool and invisible to whoever uses it. Your question is the second message. Everything the model says back is shaped by the one it read first.
System prompt vs user prompt: what the model reads first
| System prompt | User prompt | |
|---|---|---|
| Position in the chat | First message | Second message |
| Written by | Whoever built the tool | You |
| Shown to you | No | Yes |
| Holds | Role, tone, rules, and the tools the model is allowed to use | The thing you want done |
A model reads from the top down. First it takes in “you are a helpful customer service support agent for AMC Incorporation, be friendly, ask about pricing”. Then it reaches the message the customer typed: “how do I cancel my subscription?” The reply you see is the model answering both at once. Your question never arrives on its own.
Can you see the system prompt of a tool you use?
No. Type into ChatGPT and OpenAI has already put its own instructions in front of your message before the model sees it: be friendly, answer concisely. Most people using it have no idea that step happens. Every AI product you have ever used has one and none of them show it to you.
System prompt examples: what goes inside one
| Product | What its system prompt establishes |
|---|---|
| ChatGPT | You are a useful assistant. Be friendly, answer concisely. |
| Cursor | The model is a coding agent, plus a long list of coding and engineering best practices. |
| Perplexity | The model is a search agent, and it always cites its sources. |
| Nairon’s AI CRM | Company-specific detail about the people the CRM is used on. |
That last row is the one worth copying, and it sounds almost too small to matter when you write it out:
Jane doesn’t like long documents. Adam likes long documents.
Luka Eric, 2:45
Lines like that are cheap to write and nobody else has them. Whatever your dispatcher knows about a customer that you would have to explain to a new hire on day one belongs in the system prompt.
Tool access lives inside the system prompt
Giving a model the ability to use your Gmail or your calendar happens in the same place. The tools are nested inside the system prompt and described to the model before you send a word, so one block of text decides both how the thing talks and what it is allowed to touch. Personality and permissions come out of the same file. Tool calls get their own lesson.
Why the same model gives two different answers
One system prompt says answer like a legal assistant. The other says answer like a customer service rep. Same model, same question, two different outputs.
For a property management company that gap is the whole product. The agent handling an after-hours maintenance call and the agent drafting a lease renewal are the same model. What separates them is a paragraph of text nobody outside the company ever reads.
Does a system prompt only affect the first reply?
No. It still governs message 30 and message 40 of the same conversation: the tone, the questions the model asks back, the things it will not say. So write it for the whole conversation. The first answer is the least of what it decides.
How many system prompts does an AI product need?
Hundreds, one per job. The agent that enters data into your CRM runs on a prompt written for that task and nothing else. Add a second feature to the same CRM and it gets a prompt of its own. Stretching one set of instructions across both makes both worse, because every line you add to cover the second job is a line of noise during the first.
So when you are shopping for an AI tool for your solar or HVAC business, the model it runs on tells you little. Most vendors are running the same handful of models. Ask what they put in the system prompt, and how much of your own company you are allowed to put in there yourself.
