Two kinds of message

Most chat tools — and every API — separate a system prompt from your user messages. The system prompt sets the standing rules: who the assistant is, how it should behave, what it must always or never do. User messages are the individual requests. The system prompt is sticky: it applies to the whole conversation, not just one turn.

In a chat app this often hides behind "custom instructions" or a "persona." In the API it's an explicit field you send with every call.

Why personas work

Telling the model who it is changes more than tone — it shifts the whole distribution of likely answers. "You are a meticulous tax accountant" pulls toward caution, caveats, and citations; "you are a punchy copywriter" pulls toward short, vivid lines. You're not adding knowledge; you're selecting the slice of the model's range you want.

What belongs in a system prompt

Put the durable stuff here — the rules you'd otherwise repeat every message:

• The role and its standards ("senior Python engineer; write tested, readable code").
• Hard constraints ("never invent figures; if unsure, say so").
• A default format ("answer concisely; use bullets for lists of three or more").
• The audience ("the reader is a non-technical founder").

Leave the per-task specifics — the actual question, this document, today's deadline — in the user message.

A reusable pattern

A dependable system prompt has three parts: identity (who you are), rules (what you always and never do), and style (how you sound). Write one for each recurring job and you stop re-explaining yourself every session.

The user message is what you want this time; the system prompt is who the assistant always is. Move the repeating rules up into the system prompt and every later message gets shorter and more consistent.

Try this: Write a six-line system prompt for a task you do often — identity, two rules, and a style line. Paste it into your tool's custom-instructions box and run a normal request. Same question, a noticeably more "you" answer.