Lab Notes 1 min read

Lab note 02: Never let the model choose its own tone.

Tone is a decision about the person, not the message. Make it before the model writes a word.

Early on, I let the model decide how to speak. Sometimes it was brisk, sometimes chatty, sometimes oddly formal. Worse, it drifted towards flattery. Agreeable is what models do when no one tells them otherwise.

Now the tone is chosen first, by simple rules about the situation: a busy afternoon gets brisk, an evening gets warm, something urgent gets direct, something sensitive gets careful. The model is then asked to write in that register. It doesn’t get a vote.

A second rule sits under every tone: lead with meaning, never a count. “A couple need you, and one’s been sitting two days” instead of “3 emails need reply.”

The result reads less like software printing a report and more like a person who has read everything and is telling you what matters.

Rupesh Pasupuleti writes about where money, payments and AI are heading, and runs a small lab of experiments in personalised agents and agentic teams.

The pathMore thinking