Roles: System, User, Assistant

Why this matters

Every serious chat API has three roles. If you use them right, your prompt gets dramatically more reliable. If you cram everything into 'user', you fight the API.

Demo

System — durable instructions. Tone, persona, what the model is for, what it never does. Set once, persists.

User — the actual turn-by-turn input. What the human just said or pasted.

Assistant — the model's previous responses (the API sees these as context).

Example structure:

[
  { "role": "system", "content": "You are a careful code reviewer. Never approve code with unhandled errors." },
  { "role": "user", "content": "Please review:\n```js\nconst x = JSON.parse(input)\n```" }
]

Putting the reviewer persona in the system message means every subsequent user turn inherits it — you don't have to re-specify.

Try it yourself

In your playground, write a system message: "You answer only in rhyming couplets." Then in user, ask a normal business question. Notice how the persona holds across turns.