先说结论

If you mostly use chat-style AI tools and you are trying not to fall behind on new AI products, this is the kind of detail that is easy to scroll past and expensive to miss.

The scene is familiar: a new UI tool shows up, you open the chat box, and you start iterating on adjectives like “cleaner,” “bolder,” or “more premium.” That is the wrong first move here. In google-labs-code/stitch-skills, visual taste gets written into DESIGN.md before it gets written into a prompt.[S002]

Why this matters: if you miss that shift, you keep optimizing the noisy layer and ignore the artifact that is meant to carry consistency. You burn time on wording, budget on repeated retries, and attention on surface-level output while the real lever sits one layer lower.

为什么这次值得看

A tool update is worth reading not by how many features it lists, but by whether it changes your next decision. This one does.

The proof is unusually consistent across three docs. generate-design says to check whether a design system already exists first, and if it does, do not keep stuffing colors, fonts, or theme instructions into the prompt.[S001] manage-design-system defines DESIGN.md as the project's source of truth and says the tokens live at project level, so you do not have to restate them each time.[S002] design-md exists to extract a semantic design system from existing screens and code and write it into DESIGN.md.[S003]

关键证据

That does not mean every quick sketch needs a full design system. This is a docs read, not a benchmark: the current main-branch skill docs are teaching a different default. Taste is being moved into a reusable file, not endlessly re-described in chat.

If you use AI UI tools, the next question is not “How do I write a better style prompt?” It is “What is the source of truth for taste in this workflow?”

If that reframes the way you evaluate AI design tools, share it with the person who is still tuning prompts before defining the system.

适合谁 / 下一步怎么用

最后落到动作：share