What is a system prompt?

By Dylan Bailey, Certified Claude Expert · Updated 4 August 2026 · 6 min read

Most teams using AI have the same experience: it's genuinely useful, but they find themselves typing the same preamble every time. "Write in Australian English." "We're a B2B business." "Don't use exclamation marks." "Keep it under 200 words."

A system prompt is where those instructions go so you stop repeating them.

How it differs from what you type

Your prompt is the request. The system prompt is the standing brief that sits behind every request — read by the model before it sees your message at all.

It's the difference between telling a new starter the house rules on their first day and re-explaining them at the start of every task. The rules haven't changed; you've just stopped paying the cost of repeating them.

What belongs in one

• Who it's helping and at what. "You are assisting the operations team at an Australian offshore staffing business." Context shapes every answer that follows.
• Tone and language. Australian English, plain language, no marketing filler. Be specific — "professional" means nothing.
• Hard rules. Never invent figures. Never give legal or financial advice. Always flag uncertainty rather than guessing.
• Default format. If most outputs should be short paragraphs rather than bullet lists, say so once here.
• Words to avoid. Jargon you've banned, competitor names, phrasing that clashes with your brand.

What doesn't belong

• Facts that change. Prices, staff names, current promotions. These go stale silently and the model states them with total confidence.
• Task-specific detail. If it only applies to one job, it belongs in that request.
• Long documents. Pasting a whole brand guide costs context on every conversation. Summarise the rules that actually bite.

An Australian small-business example

A Perth building supplies company has four staff using AI for customer emails. Each writes differently, so the outputs vary — some too casual, one prone to American spelling, another producing walls of text.

They write a shared system prompt: Australian English, short paragraphs, no exclamation marks, always offer a phone call as the next step, never quote a price without checking, sign off with the team name.

The variation collapses immediately. Not because anyone got better at prompting, but because the standing rules stopped depending on whether each person remembered them at 4pm.

Writing a good one

Start from output you didn't like. Every time the AI produces something slightly wrong — too formal, wrong spelling, too long — that's a candidate rule. Add it once and the correction becomes permanent.

Be concrete. "Be professional" gives the model nothing. "Write like you're emailing a busy tradesperson: short sentences, no jargon, get to the point in the first line" gives it something to act on.

Then review it every few months. System prompts accumulate rules written for problems that no longer occur, and each one costs a little clarity on everything else.

Why this is the highest-leverage change

Better prompting is a skill each person has to learn individually. A system prompt is a fix written once that raises the floor for everybody — including the people who will never get particularly good at prompting and shouldn't have to.

If a team is getting inconsistent AI output, this is almost always the first thing to fix, and almost always cheaper than whatever else is being considered.