SMEme adds a reasoning layer to the AI chatbot you already use. A human expert builds a decision tree and writes the questions, branches, and outcomes in plain language. SMEme compiles that tree into logic and hands it to a constraint solver. When a case comes in, the chatbot finds the facts in your files and apps. A person approves each answer. Then the solver works out what the rules require. °FAI: Rules Don’t Belong in Prompts.

Arista Labs built it. Over three parts, I’ll walk through how it works and why it matches the way I build.

Talking to a prompt isn’t the same as teaching it.

We might talk to our prompts about our tribal knowledge. How we decide on an underbase. But the AI doesn’t retain that in a consistent way.

Every time you ask, the model rereads your instructions and decides how to apply them. Ask twice and it may weigh things differently, skip a step, or fill a gap with a guess.

Language models predict. That’s how they work. It’s also why a production call needs more than a prompt.

The challenge in a shop.

In our industry, the knowledge that matters most usually lives in people, not on paper. Often it’s the person who’s been at the press for a long time. Hand that to an AI through a prompt and you run into three problems.

The answer can change from one day to the next. You can’t see why it decided what it decided. And when it doesn’t know, it doesn’t tell you. It guesses.

That’s the equivalent of inexperience.

Taking the rules out of the prompt.

SMEme moves the rules somewhere they can’t drift, a reasoning layer. Same facts, same outcome. When the evidence conflicts or runs out, SMEme leaves the result open instead of filling it in.

I built my pipeline for screen print factories on the same principle. Discrete tools. A human at every gate. Everything stays read-only until it earns more.

What this really shows: the AI was never the place to keep your knowledge. It’s good at finding things and talking about them. The rules need their own home.

Part 2 shows you a decision tree, the reasoning layer.

Tools used:

• Claude (read the SMEme docs, drafted this article)
• SMEme (the tool under discussion)