Rules and boundaries

Everything so far has been about getting the AI to do things: sound like you, write for your customers, hit the right length. This lesson is about the other half, the part that experienced users care about most: getting it to not do things. Not quote prices. Not promise dates. Not invent a policy number. Not answer a question it should hand to you.

The reason this matters more for a business than for a hobbyist is simple. When a model guesses wrong about a movie trivia question, you laugh. When it guesses wrong about the price of a gutter job and the customer screenshots it, you eat the difference. Boundaries are the cheapest insurance you will ever buy.

Negative rules, and why they beat positive ones

A negative rule is an instruction that starts with never, do not, or avoid. Positive rules describe a target; negative rules fence off a cliff. You need both, but negative rules do something positive ones cannot: they close doors the model would otherwise walk through by default.

Here is why. Remember from AI Basics lesson 9 that the model's deepest habit is producing a plausible continuation. If a customer asks for a price, the most plausible continuation is a price. The model is not lying. It is finishing the sentence the way sentences like that usually get finished. A positive rule like "be accurate" does not help, because the model thinks it is being accurate. Only a rule that names the specific door, "never state a dollar amount," actually closes it.

Good negative rules share three traits. They are specific (name the thing, not the category). Observable (you could look at the output and tell if the rule was broken). And they come with a replacement (what to do instead). "Never quote a price. Say that a quote needs a site visit and offer to schedule one." That last clause is what keeps the reply from turning into a dead end.

Toggle the rules below on and off. Notice how each one changes the reply, and notice what the reply does with no rules at all.

The one rule every business prompt needs

If you take a single sentence from this whole module, take this one:

If you are missing a detail you need, ask me one question instead of guessing.

This rule flips the model's default. Left alone, an AI fills gaps. Told to ask, it stops at the gap and hands the question back. You saw it in the demo: with that rule on, the model did not write Tom's reply at all. It asked whether you wanted to offer a site visit first. That is one extra tap for you, and it prevents an entire category of confident mistakes.

Two refinements make it work better. First, "one question" matters. Without the number, some models respond to a simple request with a questionnaire. Second, list the details that matter most so the model knows what counts as missing: "especially the customer's name, what they bought, and whether we have been to their house before." Now the model knows what to ask about and what to let slide.

Off-limits topics and hand-offs

Every business has questions the AI should not answer, not because the model is incapable, but because a wrong answer creates real liability or the answer is simply yours to give. Common ones for Southern Utah businesses:

• Pricing and warranties. Covered above. Also "is this covered under my policy," "can you match this competitor quote," and anything with a number after a dollar sign.
• Legal, medical, and safety calls. An HVAC company's assistant should not tell a customer whether a smell means a gas leak. It should tell them to leave and call the gas company.
• Complaints that are escalating. A first-response apology is fine. Negotiating a refund is not. Write the rule as a hand-off: "If a customer is asking for money back or threatening a review, draft a short reply that says I will call them personally, and flag it to me."
• Anything about employees or other customers. The AI has no business speculating about why a crew member was late or what another client paid.

Notice that every one of these has a replacement behavior. "Do not discuss X" on its own produces stiff refusals that sound like a bank's phone tree. "Do not discuss X; instead, do Y" produces a reply that still feels like a person.

What a system prompt cannot enforce

Being honest about limits is what separates a useful setup from a false sense of security. Here is what your rules do not do.

They do not prevent hallucination. "Never make anything up" is the most common rule people write and the least effective, because the model does not know when it is making things up. What works is narrowing: "Only state facts about our business that appear in this prompt or in a file I have given you." That gives the model a boundary it can check against.

They do not override the company's hidden prompt. As lesson 1 covered, ChatGPT, Claude, and Gemini each have their own rules above yours. If your instructions conflict with theirs, theirs win. This is almost never a problem for a legitimate business, but it explains the occasional "I can't help with that" on a harmless request. Rephrase and move on.

They do not survive being contradicted by you. If your prompt says "never use emojis" and you type "add some emojis to this," you will get emojis. The model treats your live request as an intentional override, which is usually what you want. Just know that boundaries are defaults against the model's habits, not against yours.

They do not keep secrets. Anything you put in a system prompt can be repeated back if someone asks the right way. Never put passwords, customer data, or anything you would not want on a billboard into a system prompt. That is what the Documents and Data module is for.

Try this yourself

Write your boundaries section now. Copy this block, keep the lines that apply, delete the ones that do not, and add at least one that is specific to your business. Then paste it under BOUNDARIES in the template from lesson 2.

BOUNDARIES
Never state a price, estimate, discount, or warranty term. Say a quote needs a site visit and offer to schedule one.
Never promise a date or time. Say I will confirm scheduling personally.
Never give legal, medical, or safety advice. Point the customer to the right professional or emergency number.
If a customer asks for a refund or threatens a bad review, write a two-sentence reply saying I will call them today, and tell me it needs my attention.
Only state facts about our business that appear in this prompt. If you are not sure, say you will check.
If you are missing a detail you need, especially the customer's name, what they bought, or whether we have worked with them before, ask me one question instead of guessing.
[Your own: the one mistake that would hurt most in your business]

Test it right away. Ask the assistant the exact question a customer would use to fish for a price. If it holds, good. If it slips, that is fine too. Lesson 6 is about fixing exactly that.