Before you ever type a word, the app secretly hands the bot a set of hidden instructions — telling it how to act, how long to talk, whether to be a pirate. Those secret orders are the system prompt. Let's build one and watch it change the bot live.
The bot always answers the same question — "what's the weather?" — but how it answers is up to these knobs. Every knob you flip secretly rewrites the hidden instructions at the bottom. Try Pirate. 🏴☠️
The system prompt says "You are a weather bot." But a sneaky user types an order trying to override it. Who wins? Flip the shield and hit Send to find out.
💬 What the user types (edit it!)
🤖 Bot reply
You only get to see the bot's reply — the system prompt is hidden. Read the clues (caps? emojis? pirate talk? a question at the end?) and pick which secret prompt produced it. Crack 3 to win the badge!
Our mixer writes one hidden note. Real AI stacks up a whole ladder of instructions, sends them as structured data, and trains the model to obey them. Here's the grown-up version:
Glue a personality line in front of the chat. ✅ Dead simple, instant vibe change. ⚠️ A pushy user can talk the bot out of it.
Rank the voices: system beats developer beats user. ✅ Higher rules win fights. ⚠️ The ranking still isn't bulletproof (see Day 17).
The model is taught to follow instructions and prefer helpful, safe answers. ✅ Follows rules by default. ⚠️ Costs a huge amount of training to get right.
You learned the bot's best-kept secret: a system prompt of hidden instructions, set before you ever type, decides its whole personality. You saw the user-vs-rules tug-of-war, and you can now read a reply and reverse-engineer its secret orders like a detective.