For chatbot users scared of falling behind, the expensive mistake is thinking this starts with better prompts.
You see google-labs-code / stitch-skills, assume the magic is in the wording, and almost keep scrolling. That is how time, budget, and attention get spent on the wrong step.
The real shift is simpler: in Stitch, the look gets written into DESIGN.md first.
The public setup notes point the same way in 3 places. One says to check whether a project already has a design system, basically the saved rules for how things should look, and if it does, the prompt should stop repeating color, font, and theme. Another treats DESIGN.md as the project's source of truth, meaning the one place those style rules should live, so they stop getting retyped every time. A third says to extract style from screens and code you already have, then write that into the same file.
So the lesson is not "write prettier prompts." It's "lock the look once, then stop paying the repetition tax."
A tool update is worth watching not because of how many features it lists, but because it changes your next decision.
Boundary: this is a doc read, not a live team test. I only reviewed the public GitHub setup notes on July 12, 2026, so don't stretch this into "every quick sketch needs a full design system." But if you want more consistent output, the leverage starts before the prompt, not inside it.
Share this with the friend who still thinks prettier prompts are the whole trick. What would you lock once first: colors, fonts, or page mood?