Prompts: how to talk to an AI

A prompt is just… what you say

Your prompt is the words you type. The AI’s reply is its best guess at what would come next if a smart, helpful person wrote a response.

Because the AI is a “next-word guesser,” small changes in how you ask change what you get. A lot.

The three layers of a prompt

When you chat with an AI, you’re usually not the only voice it’s hearing. There are three prompts stacked on top of each other:

1. System prompt (the AI’s hidden instructions). The company that built the chat app writes this. You never see it. It says things like “You are a helpful assistant. Be polite. Don’t help with anything dangerous. Today’s date is…” It’s why ChatGPT, Claude, and Gemini feel a little different out of the box, even when you ask the same question.

2. Custom instructions (your saved preferences). Most chat apps let you save a short note about yourself that gets quietly added to every conversation. Things like “I’m a 5th-grade teacher. Always explain at that level.” or “Reply in Spanish unless I ask otherwise.” You write this once; it sticks.

3. User prompt (what you type right now). This is the message you just sent.

The model reads all three together, then writes its reply. That’s why two people asking “what’s a good dinner?” in different apps can get wildly different answers.

Three rules of thumb

1. Be specific. “Write something” → blah. “Write a two-sentence bedtime story about a lonely cloud” → cute.
2. Give it a job. “You are a patient math tutor explaining to a curious kid” → way better answers. (This is basically writing your own little system prompt.)
3. Show, don’t just tell. Give an example of what you want.

Getting good at this set of skills, finding the right way to phrase things so the AI gives you what you actually want, is called prompt engineering. Despite the fancy name, it’s mostly: be clear, give context, give examples, and try again.

Try it

Type your own prompt below. The reply streams in word-by-word, that’s literally the model picking the next word, over and over.

Experiments to try

• Same question, but ask for “in three bullet points.”
• Ask it to “explain to me like I’m five.”
• Ask it to “respond as a friendly pirate.” (Then as a grumpy pirate.)

The model doesn’t change. Only your prompt does. And yet the answer is totally different. That’s how powerful the words are.