LLMs for microcopy: Why a custom system beats a good prompt

Say we ask a model for a button label, an error message, or an empty state. What comes back is just … fine. Generic. So we write a better prompt. Give it a role, sharpen the wording. Try again, then try again. And when the copy still doesn’t sound like our product, we assume we asked wrong.

But we didn’t, though.

The prompt is the smallest and least stable part of what a model is calculating. Everything that makes our writing ours ... the glossary, the brand voice, the things only our team knows. It all has to come from somewhere. That somewhere is us.

This talk is about what this system could look like, how to build it, when to use it, and why.

In this session, you’ll learn:

• How LLMs work: tokens, context windows, memory, and reasoning.
• When to reach for LLMs: guidelines for the team.
• Real experiments: ... and how they failed.
• Building and enabling: two sides of the same coin.