A System Prompt Is the Standing Instruction That Tells Your AI Who It Works For and How to Act
A system prompt is the persistent set of instructions an AI reads before every task, defining who it is working for, what it knows, and how it should behave. It is different from the message you type into the chat box. That message is one request. The system prompt is the context that sits underneath every request, so you do not have to re-explain your business each time. In ChatGPT it lives in custom instructions or a project's instructions field. In Claude it lives in a project's custom instructions or a CLAUDE.md file. Wherever it lives, its job is the same: turn a general-purpose model into one that answers as if it already works at your company.
Most people never write one, or they write a single vague line like "you are a helpful assistant for a small business." That produces generic output because the AI still has to guess at everything specific. A good system prompt removes the guessing.
A Business System Prompt Has Four Jobs, Not One
The best system prompts do four distinct things: they assign a role, state the facts, set the rules, and define the voice. Cramming everything into one paragraph is the most common mistake, because the AI cannot tell a hard rule from a nice-to-have when they are all mixed together. Keep the four jobs visible and separate.
- Role. Tell the AI who it is and who it serves. "You are the operations assistant for a two-van plumbing company in Leeds. You help the owner write quotes, reply to customers, and draft job notes." One or two sentences.
- Facts. Give it the specifics it will otherwise invent: your services, your prices, your service area, your hours, your booking process. This is the part that makes output usable instead of plausible.
- Rules. State the decisions you never want it to get wrong. "Never quote below the 90-dollar call-out fee. Never promise same-day service. Always confirm the address before drafting a quote."
- Voice. Describe how you sound so replies do not read like a template. "Write in short, plain sentences. No jargon. Friendly but direct. Sign off as Dave."
When those four are clearly labeled, the AI can act on all of them at once instead of blending them into a single blurry impression of your business.
Write the Facts as Decisions, Not Adjectives
The facts section is where most system prompts fail, and the fix is to write specifics an employee would need, not words a brochure would use. "We offer premium, reliable service" tells the AI nothing it can act on. "First hour is 90 dollars, 60 for each hour after, we serve a 20-mile radius, and we do not take gas work" tells it exactly how to write a quote. This is the difference between context that describes your business and context that lets an AI operate it, which we cover in depth in what operational AI context actually is.
A quick test: read your facts section and ask whether a new hire could answer a customer with only that information. If they would still have to ask you a question, the AI will have to guess at the same thing, and it will guess confidently and wrong.
Structure the Prompt in Layers So the AI Can Find the Right Detail
A long wall of text works worse than a short prompt with clear sections and headings. As a system prompt grows past a few paragraphs, structure becomes the thing that keeps it usable. Use headings for the role, the facts, the rules, and the voice, and keep each section scannable. This mirrors the pattern Anthropic publishes for structuring AI context: a short orientation up top, then labeled sections underneath that the model can navigate instead of reading straight through.
If your business has more detail than fits comfortably in one prompt, the prompt itself becomes a short orientation layer that points to a larger knowledge base. That larger structure, a brief system prompt on top of organized reference files, is what we call an AI business brain, and it is how a small prompt can sit on top of a lot of knowledge without becoming a wall of text. For the trade-offs between a single instructions field and a structured set of files, see a folder of docs versus custom instructions.
Test the Prompt Against a Real Task, Then Tighten It
The only way to know a system prompt works is to give the AI a real job and see where the output is wrong. Write your first draft, then ask it to do the thing you actually need: draft a quote, reply to a specific customer message, write a job summary. Every place the answer is generic or wrong points to a fact or rule you left out. Add it, and run the task again. Two or three rounds of this turns a rough prompt into one that produces output you can use as-is.
This is faster than trying to write the perfect prompt from scratch. You are not guessing at what the AI needs; you are watching it fail and filling the exact gap.
Keep It Current or It Will Confidently Give Wrong Answers
A system prompt goes stale the moment your prices, services, or policies change, and a stale prompt is worse than none because the AI states the old facts with full confidence. If you raised your call-out fee to 100 dollars but the prompt still says 90, every quote it drafts will be wrong. Treat the prompt as a living document: update the relevant line when something changes, and reread the whole thing every few months. Because it is plain text, editing is quick. The hard part is the first draft, not the upkeep, and the same problem of an AI forgetting or misremembering your business is covered in why your AI keeps forgetting your business. For a step-by-step on where to paste any of this in ChatGPT specifically, see our ChatGPT instructions guide.
How AI Brain Docs Fits In
AI Brain Docs writes your system prompt for you, along with the knowledge base it sits on top of, so you skip the blank-page problem entirely. You answer a short set of questions about your business, around six of them, and it generates a system-prompt.md file with the role, facts, rules, and voice already filled in from your answers, plus a structured knowledge base covering your services, pricing, customers, and processes. It is the layered version described above, built and organized rather than drafted from scratch.
You drop the files into ChatGPT, Claude, or Claude Code once, and from then on your AI starts every task already knowing who it works for and how you want it to act. You can generate yours at aibraindocs.com/start.