Developers discuss how to precisely convey UI design demands to AI to prevent style drift

On November 3, 2026, at 18:02, a developer posted a discussion thread on V2EX titled "Discussion: How to communicate UI design requirements precisely to AI to avoid style drift." The thread addresses a problem many people hit when using AI for interface work: the layout and functionality are already settled, yet the visual result keeps falling short. When the developer simply asked AI to make the interface "a bit more beautiful," the system tended to return a completely fresh style instead of refining the existing one, so the result never matched the original expectation.

The concrete case is a root certificate installation wizard for Sniffy, an open-source packet capture tool that the developer is building. The screen is meant to let users pick their own device and follow on-page instructions and screenshots to complete the installation. The developer ran several versions of the page through gpt image generation and also tried design-related Skills, including frontend-design and impeccable, feeding them images of the existing page. Each generated version had some usable elements, but the overall impression was that everything competed for attention at once: the title and icon, the buttons, and the content area were all trying to stand out. That over-emphasized look did not fit the software's needs, where the developer wanted a cleaner overall appearance, clearer information hierarchy, and visual consistency with the rest of the tool.

The lesson the developer draws is that vague feedback such as "I don't think it looks good" does not give AI enough to work with, and there remains considerable uncertainty about how to make AI understand and deliver what is actually wanted. The practical route is to spell out concrete requirements: font family and size, spacing, and the ratio of height to width. With only a loose brief, the outcome is largely a matter of luck. In the end, the developer chose to work from a rough draft drawn by hand and used that layout as the basis for the final screen, then shared the comparison images to invite discussion from the community on the best ways to specify design intent when generating UI with AI.