System Prompt

Before you type a single word, the AI has already been told who to be.

There's a hidden paragraph that loads first, every single time, sitting at the very top of the room before your message ever shows up. You don't see it. It does. And it shapes everything that comes back.

Think of it as the handbook you hand a new employee before their first shift. Same person, wildly different behavior depending on what the handbook says:

• "You work at a hospital. Be calm, be careful, never guess."
• "You work at a comedy club. Loosen up, crack jokes, keep it moving."

Same hire. Two completely different employees. The handbook did that, not the person.

The system prompt is that handbook for AI. It's usually three kinds of instruction stacked together:

1. Who to be. "You are a helpful assistant." Tone, personality, attitude.

2. What to do. "Only answer questions about our products." The actual job.

3. What never to do. "Never give medical advice. Never make up a price." The guardrails.

This is why two AI tools built on the exact same model feel like different creatures. A customer support bot that politely refuses to talk about anything but your return policy, and a chaotic meme-writer that swears at you for fun, can be running the identical engine underneath. The only thing that changed is the handbook somebody wrote at the top.

Here's where it touches you directly. When you set "custom instructions" in ChatGPT (the box where you tell it your name, your job, how you like answers), you're writing your own little system prompt. You're handing every future chat a handbook before it starts. That's why it suddenly stops over-explaining, or starts giving you the short version you asked for, without you nagging it each time.

So here's the move most people miss. When the AI keeps doing something you hate, the instinct is to argue with it. You correct it, it apologizes, and three messages later it's right back to the same thing. Of course it is. You yelled at the employee on the floor but never touched the handbook they read this morning.

Want different behavior from AI? Don't argue with it turn after turn. Change the handbook.