Writing System Rules and a Style Guide for Your AI

Almost every business writes its first AI instructions the same way: "You are a helpful assistant for a dental clinic. Be professional, friendly and concise." Then they are surprised when the AI promises a same-day appointment. Our guide to teaching an AI to sound like your business separates knowledge from behaviour for exactly this reason — the documents say what is true, the rules say what to do about it.

Adjectives do almost nothing

"Professional" is a word that means something different to a law firm and a surf school. "Friendly" is already the model's default. "Concise" is the only one of the three that reliably changes anything, and even then "keep replies under three sentences unless the customer asked for detail" works far better.

The test for any line in your rules is whether you could tell, from a transcript, that the line existed. If removing it would change nothing observable, it is taking up space and diluting the instructions that matter.

Write situations, not values

Useful rules name a situation and prescribe a response. They read like training notes for a new hire, because that is what they are.

• If someone asks for a discount, say that pricing is fixed and offer to note their interest for a manager — never negotiate.
• If someone describes symptoms, do not give clinical advice; offer the earliest appointment and say a clinician will discuss it.
• If you do not find the answer in the knowledge base, say you will check with a colleague rather than guessing.
• Never confirm a delivery date. Confirm dispatch and give the tracking link instead.
• Reply in the language the customer wrote in, including dialect, and do not switch languages mid-conversation.

That last rule matters more than people expect in mixed markets — the practicalities are covered in the industry-specific approach described in why Gulf businesses run on WhatsApp.

“

A rule you could not detect in a transcript is not a rule. It is a wish.

— be digital ai team

The three things every rule set needs

Beyond your specifics, three instructions do disproportionate work. First, an identity line: what the business is, what it sells, and that the AI is an assistant for it, not a general chatbot. Second, an honesty clause: when uncertain, say so and escalate rather than improvise. Third, a hard list of things never to do — commitments about refunds, legal or medical advice, anything involving competitor comparisons.

Keep that never-list short. Twenty prohibitions dilute each other; five specific ones are remembered. If the list keeps growing, the real problem is usually scope — the AI is being asked to handle conversations that should go straight to a person, which is a question for when the AI should stop and fetch a human.

Style is mostly length and formality

The style guide has less to do than people think, because most of the tone comes from the examples you give. Specify message length, whether to use the customer's first name, whether emoji are acceptable, and how formal the address should be — a genuine decision in German and Arabic, where the formal and informal forms carry real weight.

Then give three or four real examples of good replies, taken from your own past conversations. Two examples do more than a paragraph of description, because the model matches patterns better than it follows adjectives.

Keep it short enough to be read

Long instruction sets degrade. Somewhere past a page, rules start contradicting each other, and the model resolves contradictions in ways you did not choose. If your rules run past a page, look for the ones describing facts — those belong in the knowledge base described in building a knowledge base your AI can actually use, not in the system prompt.

Change one thing at a time

When something goes wrong, the temptation is to rewrite the whole prompt. Resist it. Add or change one rule, then watch the next twenty conversations. Rewriting everything means you learn nothing about which line was responsible, and you will be back in a fortnight with the same problem in a new shape. Rules earn their place by evidence, one at a time, and the set you end up with after three months of that discipline is usually half the length of the one you started with.