What is a system prompt in plain English?

A system prompt is a standing instruction that sits above a conversation and tells the AI who it is, what rules to follow, and how to format its answers. You write it once, and it applies to every message in that chat or app without you retyping it. Think of it as the job description and house rules you would give a new staff member on day one, rather than re-briefing them on every single task. The regular messages you type are the day-to-day requests; the system prompt is the standing context they all sit inside.

Where does the system prompt live?

It depends on the tool. In Claude, it is the instructions field on a Claude Project, which apply to every chat started inside that Project. In ChatGPT, it is the instructions field when you build a Custom GPT. If you are wiring AI into your own software, it is the system parameter on the API, set separately from the user message. Same idea in all three: a standing instruction set apart from the back-and-forth of the conversation. The chat box you type into day to day is the user message, not the system prompt.

Why does a business owner care about system prompts?

Because it is how you make AI consistent across a team. Without one, every person writes their own ad-hoc instructions, so the same request produces a different tone, format and quality depending on who asked. A shared Project system prompt fixes your brand voice, your formatting rules, your do-not-do list and your standard context in one place. Everyone draws from the same source of truth, and output stops being a lottery. It is the difference between ten people each inventing a process and ten people running the same one.

What makes a good system prompt?

Be specific and concrete, the same as any good prompt. Give the AI a clear role, the context about your business it needs, explicit rules including what not to do, and the default output format you expect. Keep it tight and readable - a wall of vague aspiration helps nobody. Use plain headings or bullets so the rules are easy to scan and easy to edit later. And treat it as a living document: when you spot the AI making the same mistake repeatedly, add a line to the system prompt rather than correcting it by hand every time.

What are common mistakes with system prompts?

Three are common. First, being vague - 'be professional' tells the model almost nothing, while 'use Australian spelling, no jargon, answers under 150 words' is actionable. Second, cramming so much in that it crowds the context window and buries the important rules; keep it focused. Third, treating it as set-and-forget. Models update, your business changes, and instructions drift out of date. The fix is to review shared system prompts periodically and keep one owner responsible for each, rather than letting them quietly rot.

← All insights

Explainers·7 min read

What is a system prompt? Setting standing rules for your AI

A system prompt is a standing instruction that sits above a conversation and tells the AI who it is, what rules to follow, and how to format its answers. You write it once. It applies to every message in that chat or app, without you retyping it.

If a regular prompt is a single request - “draft this email”, “summarise this report” - the system prompt is the house rules every request lands inside. It is the closest thing AI has to an induction pack for a new hire: the role, the standards, the things we do not do here.

A standing instruction, not a one-off

Every conversation with AI has two layers, even if you only see one.

The user message is what you type into the chat box: your actual question or task. It changes every time.
The system prompt is the instruction set behind it: who the AI is acting as, the rules it must follow, and the default format for its answers. It stays the same across every message.

Most people only ever touch the first layer. They retype context into every message - “remember we are a plumbing business, keep it friendly, Australian spelling” - because nobody told them the second layer exists. The system prompt is where that standing context belongs, so you set it once and stop repeating yourself.

Where it lives

The idea is the same everywhere, but the field has a different name depending on the tool.

Claude Projects. A Project has an instructions field that applies to every chat started inside it. Set up a “Client proposals” Project with your voice, structure and standard inclusions, and every proposal chat inherits them.
Custom GPTs. When you build a Custom GPT in ChatGPT, the instructions field plays the same role and defines how that assistant behaves for everyone who uses it.
The API. If you are wiring AI into your own software, the system prompt is a dedicated system parameter, set apart from the user message. This is where a custom build encodes its standing behaviour.

Three names, one concept: a standing instruction held separately from the conversation.

Why an owner should care

Here is the part that matters for a business rather than a single user. A system prompt is how you make AI consistent across a team.

Without one, every person writes their own ad-hoc instructions in the chat box. So the same task - replying to a customer complaint, say - comes out formal from one staff member, chatty from another and a single curt line from a third. The quality is a lottery that depends on who asked and how good their prompting was that day.

A shared Project system prompt ends the lottery. You fix the brand voice, the formatting rules, the standard context and the do-not-do list in one place, and everyone draws from the same source of truth. The output stops depending on individual skill and starts reflecting your standards: not ten people each inventing a process, but ten people running the same one. It is also the cheapest quality lever you have - you are not buying a better model, just writing down once the rules you already hold in your head.

How it works in practice

A weak system prompt and a strong one for the same assistant look very different.

Weak:

You are a helpful assistant for our business. Be professional and friendly.

That tells the model almost nothing it did not already assume. “Professional and friendly” could describe a million different replies.

Strong:

You are the customer support voice for Brightwater Plumbing, a commercial plumbing business in Brisbane. Use Australian English. Keep replies under 150 words, warm but direct, no jargon. Always acknowledge the customer’s issue before offering a solution. Never quote a price - say a team member will confirm it. End with a clear next step.

Same model. The second one produces something close to on-brand on the first attempt, because it carries the role, the rules, the constraints and the format. This is the same skill as everyday prompting - be specific, give context, state the format - just written down once instead of retyped constantly.

What to do

Write one for any task you repeat. If a workflow comes up weekly, it deserves a Project with a proper system prompt rather than fresh instructions every time.
Be concrete. Role, context about your business, explicit rules including what not to do, and the default format. Swap “be professional” for testable rules like “answers under 150 words, Australian spelling”.
Keep it tight and scannable. Use plain headings or bullets so the rules are easy to read and edit later. A focused half-page beats a rambling essay.
Treat it as living. When the AI keeps making the same mistake, add a line to the system prompt so the fix sticks for everyone, rather than correcting it by hand each time.

Honest limits

A system prompt is powerful, but it is not a force field. The model can still drift from your instructions, especially in long conversations, and it will occasionally ignore a rule. For anything that drives money or goes out under your name, keep a human checking.

It also has trade-offs. The system prompt takes up room in the context window - the AI’s working memory for that conversation - so an enormous one leaves less space for the actual work and can bury your important rules in noise. Tight beats exhaustive.

And it ages. Models update and your business changes, so instructions drift out of date. Give each shared system prompt one owner and a periodic review, like any other process document. Set it and forget it, and it rots.

The one-line version

A system prompt is the standing instruction - who the AI is, what rules to follow, how to format answers - that applies to every conversation, so you do not repeat yourself. It lives in Claude Project instructions, Custom GPT instructions or the API system parameter, and for a team it is the simplest way to make AI consistent and on-brand. Write it concretely, keep it tight, and review it so it does not go stale.

Frequently asked questions

What is a system prompt in plain English?: A system prompt is a standing instruction that sits above a conversation and tells the AI who it is, what rules to follow, and how to format its answers. You write it once, and it applies to every message in that chat or app without you retyping it. Think of it as the job description and house rules you would give a new staff member on day one, rather than re-briefing them on every single task. The regular messages you type are the day-to-day requests; the system prompt is the standing context they all sit inside.
Where does the system prompt live?: It depends on the tool. In Claude, it is the instructions field on a Claude Project, which apply to every chat started inside that Project. In ChatGPT, it is the instructions field when you build a Custom GPT. If you are wiring AI into your own software, it is the system parameter on the API, set separately from the user message. Same idea in all three: a standing instruction set apart from the back-and-forth of the conversation. The chat box you type into day to day is the user message, not the system prompt.
Why does a business owner care about system prompts?: Because it is how you make AI consistent across a team. Without one, every person writes their own ad-hoc instructions, so the same request produces a different tone, format and quality depending on who asked. A shared Project system prompt fixes your brand voice, your formatting rules, your do-not-do list and your standard context in one place. Everyone draws from the same source of truth, and output stops being a lottery. It is the difference between ten people each inventing a process and ten people running the same one.
What makes a good system prompt?: Be specific and concrete, the same as any good prompt. Give the AI a clear role, the context about your business it needs, explicit rules including what not to do, and the default output format you expect. Keep it tight and readable - a wall of vague aspiration helps nobody. Use plain headings or bullets so the rules are easy to scan and easy to edit later. And treat it as a living document: when you spot the AI making the same mistake repeatedly, add a line to the system prompt rather than correcting it by hand every time.
What are common mistakes with system prompts?: Three are common. First, being vague - 'be professional' tells the model almost nothing, while 'use Australian spelling, no jargon, answers under 150 words' is actionable. Second, cramming so much in that it crowds the context window and buries the important rules; keep it focused. Third, treating it as set-and-forget. Models update, your business changes, and instructions drift out of date. The fix is to review shared system prompts periodically and keep one owner responsible for each, rather than letting them quietly rot.

Where this fits

Claude Implementation

Install Claude properly across your team - Claude Code, Claude.ai projects and skills, custom Anthropic SDK builds.

Explore Claude Implementation →Book a free 30-min discovery call