Anatomy of a good system prompt

If you search for system prompt examples you will find thousands of them, from three-line notes to ten-page documents with headings and numbered rules. They look wildly different. But strip away the formatting and nearly all of the good ones contain the same six ingredients. Once you can name them, you can write one, and just as important, you can look at one that is not working and see what is missing.

The six are: role, audience, goals, format rules, boundaries, and examples. Let's take them one at a time, then build one.

Role and audience: who is talking to whom

Role answers "who is the AI in this conversation?" Not a costume, a job. "You are the marketing assistant for a window tint shop" tells the model what kind of knowledge to reach for and what kind of person to sound like. Compare that to "You are a helpful assistant," which tells it nothing it did not already assume.

A good role line includes the business and the place. "A gutter company in Washington County, Utah" is better than "a gutter company," because the model now knows about desert monsoons, flat roofs, and the fact that half your customers moved here from somewhere colder and do not know what a scupper is.

Audience answers "who reads what the AI writes?" This is the part people skip most, and it is the part that does the most work. The same information gets written completely differently for a retired couple in Sun River, a property manager with forty units, or a contractor who already knows the trade. Tell the model who is on the other end and it will adjust reading level, assumed knowledge, and tone without you asking each time.

Goals: what the AI is for

Goals are the two or three jobs you actually want done. Not everything the AI could do, the things you will really use it for. "Help me write review replies, Facebook posts, and quote follow-up emails" is a goal list. It lets the model bring the right instincts to a vague request: if you type "Dana was unhappy," an assistant with that goal list guesses you want a review reply, not a diary entry.

Goals also shape what the AI volunteers. If one of your goals is "bring people into the shop," the model will naturally end posts with a reason to visit. If you never say that, it will end them with whatever seems generically nice.

Keep the list short. Three to five goals. If you find yourself writing twelve, you probably need two different system prompts, one per kind of work. Lesson 5 shows how to keep several.

Format rules and boundaries: the shape and the fence

Format rules are the mechanical preferences you would otherwise repeat in every chat: length, structure, punctuation, whether lists are allowed. These are the highest-return sentences in the whole prompt because they fix the most common complaints about AI writing in one stroke. "Under 80 words. Plain sentences. No bullet lists in anything a customer will read. No hashtags, no emojis." Four rules, and the output stops looking like it came from a robot.

Boundaries are the lines the AI must not cross. "Never quote a price. Never promise a date. Never make up a review." They are different from format rules because breaking them costs you money or trust, not just style points. Boundaries also include the single most useful instruction in any business system prompt: "If you are missing a detail, ask me one question instead of guessing." That one line turns the model's worst habit, confident invention, into its best one, asking. Lesson 4 is entirely about boundaries, so we will keep it brief here.

Examples: show, do not just tell

You met this trick in AI Basics lesson 5: paste something you like and say "match this." In a system prompt, an example is a short sample of your real voice that the model can imitate. One or two real sentences you actually wrote, or would happily have written, does more than a paragraph of adjectives.

"Warm but professional" means twenty different things to twenty different people. "Thanks for coming by, Maria. Glad the ceramic tint is keeping the truck cooler." means exactly one thing. The model is a pattern matcher. Give it the pattern.

Now build one. Tap the blocks below and watch the prompt assemble. Notice how it reads with only a role, and how different it feels once boundaries and an example are in.

You are the marketing and customer-service assistant for Red Rock Window Tint, a window tinting shop in St. George, Utah.

Facts, and keeping the whole thing short

There is an optional seventh part that many businesses end up adding: a short facts block. It is a plain list of the things the model cannot possibly know and will otherwise invent: your hours, your service area, the towns you cover, the services you do and do not offer, your phone number, how customers book. Five to ten lines, no prose.

The test for whether a fact belongs here is whether a customer might ask about it. "We serve St. George, Washington, Hurricane, Ivins, and Santa Clara. We do not go to Cedar City." That one line stops the AI from cheerfully agreeing to a job ninety minutes away. Keep it current: a stale facts block is worse than none, because the model will state the old hours with total confidence. If you have more than a dozen facts, or they change often, they belong in a file attached to the project instead, which the Documents and Data module covers.

How long should the whole prompt be? Shorter than you think. The six parts above fit comfortably in 150 to 300 words for most small businesses. Remember from AI Basics that everything in the chat, including the system prompt, is measured in tokens and takes up room in the model's memory. A tight briefing is read carefully. A rambling one gets skimmed, and the instructions in the middle are the ones that get lost.

A useful test: read each sentence and ask, "would the model behave differently if I deleted this?" If not, delete it. "Be helpful" fails that test. "Never promise a date" passes it.

Order matters a little, too. Put role and audience first, because they frame everything. Put boundaries near the end so they are the last thing the model reads before your message. Examples can go anywhere, but they are easiest to maintain at the bottom, where you can swap them out as your voice evolves.

Try this yourself

Fill in this template for your own business. Write it in a notes app, not in the AI yet. You will install it in lesson 5 after you have tuned the voice in lesson 3 and the boundaries in lesson 4.

ROLE
You are the assistant for [business name], a [what you do] in [city], Utah.

AUDIENCE
Everything you write is for [who reads it: homeowners, tourists, contractors, tenants]. Assume they [do / do not] know the trade.

GOALS
Help me with: [job 1], [job 2], [job 3].

FORMAT RULES
[Length]. [Sentences or lists]. No hashtags, no emojis, no exclamation points. [Anything else you always fix by hand].

BOUNDARIES
Never [thing that costs money]. Never [thing that costs trust]. If you are missing a detail, ask me one question instead of guessing.

EXAMPLE OF MY VOICE
"[One or two real sentences you have actually sent to a customer.]"

Keep the headings. They cost a few tokens and make the prompt far easier for you to edit later, which matters more than you would guess. The one you will actually maintain beats the one that was perfect on day one.