Why your AI assistant needs a name and a job description

Five of the assistants we run have names: Cruz, Drew, Kai, Mario, and Blair. None of them got a name for fun. The name came after the job description, the same way it would for a new hire.

Why a name at all

A chat bubble that says "Chat with us" is anonymous. Visitors treat it like a form. A bubble that says "Hi, I'm Kai. Ask me about your project" gets talked to like a person at a front desk, which is the behavior we want: real questions, in the visitor's own words, with enough detail to make a good lead.

The name also does something for the business. Once the assistant is "Drew," the owner can say "Drew got that one wrong" and we all know what is being discussed. Problems get specific. A nameless feature gets vague complaints; a named one gets fixable ones.

We pick short, easy names that fit the brand and do not sound like an employee's full name. Nobody should wonder if Mario is a person they can ask for. And the first thing each assistant says when asked is that it is an AI assistant, so the name never becomes a disguise.

The job description comes first

Before we write a line of instructions, we write a job description with the owner, the same way they would for a receptionist. It is one page and it answers plain questions:

• Who does this assistant talk to? (Homeowners planning a project. Riders asking about a shuttle. People applying for a job.)
• What does it know? (This website, nothing else.)
• What is it for? (Answer, qualify, and get a callback booked.)
• What tone does it use? (Friendly, brief, no exclamation points, no sales pressure.)
• What may it do on its own? (Answer from the site, collect contact details, create a lead.)
• What must it hand to a person? (Prices not on the site, complaints, anything legal or medical, anything it is unsure about.)
• What must it never do? (Pretend to be human, invent facts, discuss competitors, collect sensitive data.)

The owner reads it and edits it in their own words. "Do not let people think we do pools." "If someone mentions a leak, tell them to call the office number, do not take a message." Those edits are gold, because they are the things only the owner knows.

How the job description becomes instructions

The model behind the assistant runs on a set of written instructions that it reads before every conversation. Builders call this the system prompt. Ours is the job description, rewritten into direct sentences the model follows well, plus the mechanics: which page the visitor is on, what tap options to offer, how to format a callback request.

That translation is mostly literal. "Friendly, brief, no sales pressure" becomes a few lines about sentence length and about never asking twice for a phone number. "Hand off anything legal" becomes a specific trigger and a specific action. The owner's edit about leaks becomes its own rule with the office number in it.

When an owner later asks "why did Blair say that," we open the job description, find the line, and fix it there. The instructions are regenerated from it. We do not patch the prompt directly, because then the document and the behavior drift apart and nobody can tell what the assistant is supposed to do anymore.

Personality limits

The temptation is to give the assistant a big personality. We go the other way. A little warmth, a consistent voice, and then it gets out of the way. Jokes get old by the third visit. Enthusiasm reads as sales pressure. An assistant that sounds like a calm, competent person at a front desk is the one visitors trust with their phone number.

Each of the five has a slightly different voice because each business has a different voice. The one on a transit site is quick and practical. The one on a cabinetry site is patient and detail oriented. None of them are chatty.

Scope is a feature

A narrow job description is not a weakness. An assistant that only knows its own website and only does three things does those three things reliably, and the owner can explain what it does in one sentence. That is harder to achieve than it sounds, and it is the difference between an assistant customers use and one they close.

Cruz, Drew, Kai, Mario, and Blair each have a one page job description on file, and each one is the actual source of how they behave. That is the method we use for every assistant we build, and it is the same one we would use for a person.

Questions people ask