Let the rule say hello. Let the model answer when asked.

“Good morning! Here are three insights about your work today.”

Three insights. Written fresh by a model. Every time you open the page.

Worded a little differently each time, too. 🤔

I’ve been building an assistant called Claris for ClarisTXM, my independent concept for running BPO work. Early on I had to decide something: what should she say before anyone asks her anything?

The easy answer is to let the model write it.

Send the page to a language model, ask for a friendly summary, show the result. It feels smart. It demos well.

Then you look closer.

The greeting takes a few seconds to appear. The wording changes between visits. Now and then it could mention a number that isn’t on the page. And every page load costs a model call that nobody asked for.

That’s a lot of risk for “good morning”.

So Claris speaks first, by rule.

Her bar at the top of every page holds a fixed sentence with real fields filled in. “Hi, first name, good morning.” Then one line about the page you’re on, from a template written for that page.

The first name comes from the person who signed in. Morning, afternoon or evening comes from their own clock. No model touches any of it.

Next to that sits one button: Ask Claris.

Press it, and the panel opens already asking a question about that page. Now a model answers. And every answer carries the same label: “Generated. Check before you act.”

Two voices. One rule for which comes first.

Deterministic first. AI forward. Human in the lead.

That line became ClarisTXM’s core line in October. Here’s what it means on a screen.

1. 1Deterministic first. If a template and a few read fields can say it, a template says it. Greetings, counts, page names, what to do next.
2. 2AI forward. The model is one click away, on the page, already holding the question. Not hidden in a menu.
3. 3Human in the lead. A rule decides what’s on the list. A person decides what to do. The model helps, and says so.

Here’s what surprised me. The rule voice doesn’t make Claris feel less helpful. It makes her feel more reliable.

She says the same thing the same way. People learn to trust it. And they notice when something changes, because the change is real.

When someone does want a conversation, the model gets a better start too. It answers a real question, from the page’s facts, for a person who asked.

To be clear, ClarisTXM is a concept I’m building, not a deployed product, and nothing here is a measured result. It’s a design rule. So far it holds up on every page I’ve tried it on.

So next time someone wants the AI to “just say something friendly” at the top of the page, ask what it would say.

If you can write it down, you probably don’t need a model to write it for you. 😅