How to Write a System Prompt for Your Business

A system prompt is the standing instruction your AI reads before every task — who it works for, what it knows about your business, and how it should behave. Learning how to write a system prompt is the single highest-leverage hour a business owner can spend on AI, because it's the difference between generic output you rewrite every time and output that sounds like it came from someone who already works at your company. This guide covers what goes in one, where to put it in Claude and ChatGPT, real examples you can adapt, and how to test it against actual work.

TL;DR

A system prompt has four jobs: assign a role, state the facts of your business, set the hard rules, and define your voice. Write each as its own labeled section, use specifics an employee would need rather than brochure language, and test it by giving the AI a real task and fixing whatever comes back wrong. Once you know how to write a system prompt, every conversation starts from "I work here" instead of "tell me about your business." The free Claude Cowork course teaches this hands-on with real business workflows, and the examples below give you a working starting point today.

What a system prompt is (and what it isn't)

The message you type into the chat box is one request. The system prompt is the context that sits underneath every request, so you never have to re-explain your business.

Where it lives depends on the tool:

The name varies — custom instructions, project instructions, CLAUDE.md — but the job is identical: turn a general-purpose model into one that answers as if it's on your payroll.

Most owners either never write one or write a single vague line like "you are a helpful assistant for a landscaping company." That produces generic output because the AI still has to guess at everything specific: your prices, your service area, your policies, how you talk to customers. A good system prompt removes the guessing.

The four jobs of a business system prompt

The most common mistake is cramming everything into one paragraph. When role, facts, rules, and tone are blended together, the AI can't tell a hard constraint from a nice-to-have. Keep the four jobs separate and labeled.

1. Role — who the AI is and who it serves

One or two sentences. Name the business, the person it's helping, and the kinds of work it does.

You are the operations assistant for Ridgeline Plumbing, a two-van
plumbing company in Grand Rapids. You help the owner, Dave, write
quotes, reply to customer messages, and draft job notes.

2. Facts — the specifics it would otherwise invent

This is the section that makes output usable instead of merely plausible. Services, prices, service area, hours, booking process, warranty terms. Anything a new hire would need on day one.

FACTS
- Services: residential plumbing repair, water heater replacement,
  drain cleaning. We do NOT do gas line work or new construction.
- Pricing: $95 call-out fee (covers first hour), $70/hr after.
  Water heater replacement: $1,400-$1,900 installed depending on unit.
- Service area: 25-mile radius of Grand Rapids. We do not serve
  Lansing.
- Hours: Mon-Fri 7am-5pm. No weekend calls except active leaks
  (emergency rate: $175 first hour).
- Booking: customers call or text; we confirm a 2-hour arrival window
  the evening before.

3. Rules — the decisions it must never get wrong

State these as absolutes, separate from the facts. The AI treats a labeled rules section far more strictly than the same information buried in prose.

RULES
- Never quote below the $95 call-out fee.
- Never promise same-day service. Say "usually within 1-2 business
  days" instead.
- Always confirm the full address before drafting a quote.
- Never diagnose a problem definitively from a description alone —
  frame it as "most likely" and note we confirm on site.

4. Voice — how you actually sound

Without this, every reply reads like a corporate template. Describe your tone in plain terms, and if you can, show it.

VOICE
- Short, plain sentences. No jargon, no corporate filler.
- Friendly but direct. We sound like a neighbor who knows plumbing,
  not a call center.
- Sign customer messages as "Dave, Ridgeline Plumbing."
Example of our tone: "Hey Sarah — that sounds like a worn flapper
valve, which is a quick fix. We can get out there Thursday morning.
Call-out is $95 and that covers the first hour."

Put those four blocks together and you have a complete, working system prompt in under a page.

Write the facts as decisions, not adjectives

The facts section is where most business system prompts fail, and the fix is simple: write the specifics an employee would need, not the words a brochure would use.

"We offer premium, reliable service at competitive prices" gives the AI nothing to act on. "Call-out is $95, first hour included, $70 an hour after, 25-mile radius, no gas work" tells it exactly how to write a quote.

Here's the test: read your facts section and ask whether a brand-new hire could answer a customer question using only what's written there. If they'd have to come ask you something — what's the warranty, do we serve that town, what's the deposit — the AI will hit the same gap. The difference is the new hire asks; the AI guesses, confidently, and gets it wrong.

If you run a trade business, the sections write themselves from the questions customers already ask. The trades posts in this series — ChatGPT for HVAC businesses, ChatGPT for general contractors — show what those recurring questions look like per industry.

Structure beats length: use labeled sections

A long wall of text performs worse than a shorter prompt with clear headings. As your system prompt grows past a few paragraphs, structure is what keeps it usable — the model navigates labeled sections instead of reading a blur. This mirrors the pattern Anthropic publishes for structuring context files like CLAUDE.md: a short orientation at the top, labeled sections underneath.

And there's a natural ceiling. A system prompt should stay around one page. If your business has more detail than fits — full price lists, process documents, policy details — the system prompt becomes the orientation layer that points to a set of reference files, and the depth lives in those files. That larger structure is an AI knowledge base, and it's how a short prompt can sit on top of a lot of company knowledge without bloating. We cover how to build one in our guide to building an AI knowledge base for your business, and the related problem of getting AI to retain your context between conversations in how to make AI remember your business.

A reasonable rule of thumb: under one page, put it all in the system prompt. Over one page, split it — prompt on top, files underneath.

Test it against a real task, then tighten

You cannot evaluate a system prompt by reading it. You evaluate it by giving the AI a real job and looking at where the output is wrong.

The loop:

1. Write the first draft using the four sections above. Don't polish it.
2. Run a real task. Not a test question — actual work. "Draft a reply to this customer text: 'my water heater is leaking, how much to replace it?'"
3. Mark every miss. Generic phrasing, a wrong price, a promise you'd never make, a tone that isn't yours. Each miss points to a fact, rule, or voice note you left out.
4. Add the missing line and rerun. Two or three rounds turns a rough draft into a prompt whose output you can send as-is.

This is much faster than trying to write the perfect prompt from scratch, because you're not guessing what the AI needs — you're watching it fail and filling the exact gap. It's the same start-small approach we recommend across how to use AI for your business: one real rep beats an afternoon of theorizing.

Where system prompts fail (and what to verify)

A system prompt makes AI output specific to your business. It does not make it correct. Three failure modes to stay honest about:

A stale prompt is worse than no prompt. The moment your prices, services, or policies change, the prompt is wrong — and the AI will state the old facts with total confidence. If you raised your call-out fee to $110 and the prompt still says $95, every quote it drafts underprices you. Update the relevant line the day something changes, and reread the whole prompt every couple of months. The first draft is the hard part; upkeep is a two-minute edit.

The AI will still fill gaps with invented detail. Anything not covered in your facts, the model guesses at — plausibly and confidently. If a customer asks about financing and your prompt says nothing about financing, you may get a fluent answer describing a payment plan you don't offer. When output surprises you, treat it as a missing fact, not a smart AI.

No live data, no math you should trust. A system prompt doesn't give the AI access to your calendar, your inventory, current permit rules, or reliable arithmetic. Numbers that end up in a quote, dates you commit to, and anything regulatory get verified by you before they reach a customer. The system prompt handles the words and the context; the judgment stays yours.

Frequently asked questions

What is a system prompt in simple terms?

A system prompt is the standing instruction an AI reads before every conversation — who it works for, what it knows, and how it should behave. It's separate from the messages you type: those are individual requests, while the system prompt is the permanent context underneath all of them. In ChatGPT it's called custom instructions; in Claude it lives in a Project's instructions or a CLAUDE.md file.

How long should a business system prompt be?

Aim for half a page to one page, organized into labeled sections for role, facts, rules, and voice. Shorter than that usually means the facts are too thin to be useful; much longer and it should be split into a brief system prompt on top of separate reference documents, with the prompt pointing to the files.

What is the difference between a system prompt and custom instructions?

They're the same idea under different names. "Custom instructions" is ChatGPT's label for its system-prompt field; Claude uses project instructions or a CLAUDE.md file; developers using the API set a literal system parameter. Whatever the field is called, the content that belongs in it is identical: role, business facts, hard rules, and voice.

Can I use the same system prompt for Claude and ChatGPT?

Yes. A well-written business system prompt is plain text describing your company, and both tools read it the same way. Write it once, keep it in a document you control, and paste it into Claude's project instructions and ChatGPT's custom instructions. When something about the business changes, update the master copy and re-paste.

What should you not put in a system prompt?

Leave out marketing adjectives ("premium, best-in-class service") because the AI can't act on them, and leave out anything that changes weekly, like current availability, since it will go stale and be stated as fact. Also skip passwords, full customer records, and sensitive personal data — a system prompt is context, not a vault. Detail that runs past a page belongs in separate knowledge-base files the prompt points to.

Write yours this week

Draft the four sections for your business — role, facts, rules, voice — then give the AI one real task and fix what comes back wrong. Two or three rounds and you'll have a system prompt that makes every future conversation start from a running head start.

The free Claude Cowork course walks through this hands-on, including how to set up a CLAUDE.md file and put your business context to work on real files and workflows. It's built for owners and operators, not developers. And if you'd rather have your system prompt and business context written and installed for you in one pass, Get Set Up on Claude does exactly that.