What is a system prompt? the hidden instruction shaping every AI answer

TL;DR: A system prompt is a standing instruction that sits above your conversation and defines the model’s role, rules and format. It is the highest-leverage thing you can write, and most people never write one.

The two layers

Every AI interaction has at least two layers of instruction:

• The system prompt — set once, applies to everything after. “You are a technical editor. You never rewrite; you only critique. You always give three specific notes.”
• The user message — your actual request, sent per turn.

The system prompt has more weight. When the two conflict, the model generally leans toward the system layer. That is why a good one changes the character of every subsequent answer rather than just one.

Where you actually set one

You may be using them without knowing:

• Custom instructions in ChatGPT settings — a system prompt applied to every chat.
• Projects in ChatGPT and Claude — a system prompt scoped to one workspace. Compared in the Projects comparison.
• Custom GPTs / Claude Styles — packaged system prompts you can reuse and share.
• The API — an explicit system field, where this is most visible.

What belongs in one

1. Role. Not “you are a helpful assistant” — that is the default and adds nothing. “You are a sceptical editor who has published in trade press for fifteen years.”
2. Rules. Especially negative ones. “Never use the words ‘delve’, ‘landscape’ or ‘furthermore’.” “Never invent statistics; if you do not know, say so.”
3. Format. “Answer in under 200 words. Bullets only when listing three or more things.”
4. Context about you. What you do, who you write for, what you already know. This stops the model explaining basics every time.

What does not work

Vagueness. “Be creative”, “write engagingly”, “be professional” — these do essentially nothing because they have no operational meaning. Every constraint should be something you could check compliance with. If you cannot verify it, the model cannot target it.

Also: length. Very long system prompts dilute themselves. Ten sharp rules beat forty soft ones, and they cost fewer tokens on every single turn.

A template worth stealing

Paste this into custom instructions and adapt:

“I am a [role] writing for [audience]. Assume I already understand [domain basics] — do not explain them. Default to answers under 200 words unless I ask for depth. Lead with the direct answer, then the reasoning. If a question has a real trade-off, name both sides rather than picking the safe one. If you are uncertain, say so explicitly instead of hedging with vague language. Never use bullet points for fewer than three items.”

Why this beats prompt tricks

Individual clever prompts help one answer. A system prompt improves every answer for months, including the ones where you were lazy and typed six words. It is the compounding version of prompt engineering — and if you want ready-made examples to adapt, the ten prompts I reuse weekly are all built on this structure.