# AI设计攻略,在 VibeCoding 时代做 UI 判断的一个小技巧 - ID: 97 - URL: https://www.stoyard.com/viewpoint/97 - English URL: https://www.stoyard.com/en/viewpoint/97 - Category: 设计思维 - English Category: Design Thinking - English Title: AI design strategy, a small skill to make UI judgment in the VibeCoding age ## 中文主体 > 一张UI界面做得好不好,其实不用打开 Figma 去量间距审查细节,可能就是把眼睛眯起来看三秒就够了。字全都糊成一片的时候,你还能一眼看出这页是干什么的、先看哪儿、后看哪儿,层级就对了 ## 一张 UI 界面做得好不好,其实不用打开 Figma 去量间距、审查细节,可能就是把眼睛眯起来看三秒就够了。字全都糊成一片的时候,你还能一眼看出这页是干什么的、先看哪儿、后看哪儿,层级就对了;如果糊完之后只剩一团平铺的灰,那就说明你只是把元素摆上去了,没有排过队。  眯眼测试示意:左侧为清晰界面,右侧为模糊后只留下大块面与明暗结构 ## 还需要设计师判断:AI 在界面上暂时不懂“谁更重要” 用自然语言让 AI 生成界面,效率确实高,但你会发现它特别容易交出一张“平”的页面:标题 24px、正文 16px,按钮和说明文字一样抢眼,所有卡片长一个样,间距整整齐齐,视觉上却没有任何一处“先说话”。 原因不复杂:AI 接收的是“把哪些元素放上去”,而不是“谁最该被看见”。在它的理解里,一个标题和一个辅助说明的差别只是标签不同,权重是一样的。你不给优先级,它就平均对待。而用户的注意力从来不会平均分配,一屏之内,人只愿意先看一个东西。 所以层级不只是美化,它更是把信息翻译出来:**把“业务上最重要的信息”翻译成“眼睛最先看到的东西”。** ## 设计师的眯眼测试:把细节抽掉,只留结构 眯眼的原理很简单:视觉细节,比如小字、细线、浅色图标,在模糊后第一个消失,剩下的只有大块面、强对比和明暗关系。也就是说,眯眼测的是页面的骨架,不是外表的装饰。 具体怎么用呢?退后一步,把眼睛眯到字看不清,然后问自己三个问题: - 还能看出这页在干什么吗? - 第一眼落在哪个区域? - 如果需要用户做一件事,比如下单、注册、提交,那个入口还在吗? 三个都答得上来,说明层级合格;有一个答不上来,就说明该大的不够大、该弱的没弱下去。 ## 攻略之给 AI 的三句提示词 实践中最有价值的,是三个可以直接复制的提示词,它们对应三种不同的修改场景。 **做完页面,先让它自查一遍。** 复制这句给 AI:优化 UI 视觉层级。它会重新过一遍整页的权重,通常会调整标题与正文的字号差、主按钮与次要按钮的对比、卡片之间的一致性和留白节奏,适合新页面刚成型的时候用。 **改组件改不动,先别急着调样式,先核对一下。** 复制:核对一下视觉层级。很多时候问题不在组件本身,而在它和周围元素的关系,一个卡片单独看没问题,放进页面里就淹没了。这句话会把注意力拉回到“它在整页里排第几”。 **只想动字、不想动布局。** 在后面再补一句:只改字号、粗细和颜色,不动布局和间距。这是最实用的一句,因为 AI 改层级时最常见的副作用就是顺手重排布局,把本来对的结构一起弄乱。加上这个约束,你就能拿到“只换皮肤、不动骨架”的安全修改。 ## 四个真正有效的层级手段 提示词只是入口,判断还得靠自己。能拉开层级的手段其实只有四类,按性价比从高到低排: - **字号**:最直接。标题与正文之间至少要有 1.5 倍以上的差,才叫“标题”。 - **字重**:粗体是廉价的强调,但要用得克制,一页里三处粗体,等于没有粗体。 - **颜色与对比**:主信息用最高对比,辅助信息降到六成左右。同色系拉开深浅,比换颜色更高级。 - **留白**:最容易被忽略,也最有效。给主内容更多呼吸空间,它自己就会浮起来。 位置、动效、阴影都排在后面。很多页面做不出层级,问题不在缺手段,而在用得太平均:**每个元素都在喊,结果谁也没被听见。** ## 一句话的工作流 把上面的东西收成一条每天都能用的流程: 先想清楚这一页只让用户干一件事,把那件事的主信息拎出来;再用字号、字重、颜色、留白把它推到第一顺位;剩下的信息一律降权;最后眯起眼验收。 STOYARD Insight** 在 AI 辅助创作和 VibeCoding 常态化的阶段,工具负责生成方案,判断依然来自设计师。STOYARD 在品牌与数字体验项目中,把视觉层级、信息优先级和界面验收沉淀为可复用的工作方法,帮助团队在高效率产出的同时守住界面的清晰度与品牌质感。 **关键词**: UI 视觉层级、AI 设计、VibeCoding、UI 判断、眯眼测试、提示词、界面设计、设计层级、生成式 AI、设计师能力 **分类**: AI 趋势 / 设计思考 / UI 设计 / 工作方法 **资料来源**: STOYARD 团队公开实践观点;文章发布于 2026 年 11 月 22 日。 ## English Content > Whether a UI interface is well done or not, in fact, it is not necessary to open Figma to measure the spacing and review the details. It may be enough to narrow your eyes for three seconds. When all the words are pasted into one piece, you can still see a ## Is a UI interface well done or not, in fact, there is no need to open Figma to measure the spacing and review the details, which may beSquinting for three seconds is enough.. When all the text blurs into a solid mass, if you can still instantly tell what the page is about and where to look first and then next, the hierarchy is right.If, after applying the blur, all that remains is a flat expanse of gray, it means you’ve merely placed the elements there without arranging them in any order.  Schematic for squinting test: clear interface on the left side, and only large surface and light and dark structure are left after blurring on the right side. ## Designers are also required to judge: AI do not understand "who is more important" on the interface for the time being" Using natural language to have AI generate interfaces is indeed efficient, but you’ll find it tends to produce overly “flat” pages: headings at 24px, body text at 16px, buttons as eye-catching as the accompanying labels, all cards looking identical, and spacing that’s perfectly uniform—yet visually, nothing stands out or grabs your attention first. The reason is simple: AI receives the instruction “which elements to place,” not “who should be seen the most.” In its view, the difference between a title and a subtitle is merely one of labeling; their weights are the same. If you don’t assign priorities, it treats everything equally. And users’ attention is never evenly distributed; within a single screen, people tend to focus on just one thing first. So the hierarchy doesn't just beautify, it translates the message:**Translate "the most important information in business" as "the first thing the eye sees".** ## The designer's squinting test: take out the details and leave only the structure. The principle behind squinting is simple: visual details—such as small text, thin lines, and light-colored icons—are the first to disappear when blurred, leaving only large areas of color, strong contrasts, and variations in light and shadow. In other words, squinting tests the page’s skeletal structure, not its decorative appearance. How exactly do you use it? Step back, squint until the text is barely legible, and then ask yourself three questions: - Can you still tell what this page is doing? - Which area do you notice first? - If you need users to perform an action—such as placing an order, registering, or submitting—is that entry point still available? If you can answer all three, it shows your hierarchical thinking is up to standard; if you can’t answer one, it means the strong aren’t strong enough and the weak haven’t been weakened sufficiently. ## A Guide: Three Prompt Instructions for AI In practice, the most valuable are three directly replicable prompts, each corresponding to a different modification scenario. **After finishing the page, let it run a self-check first.** Copy this sentence to AI:优化 UI 视觉层级. It recalculates the page’s overall weight, typically adjusting the size contrast between headings and body text, the contrast between primary and secondary buttons, consistency across cards, and the rhythm of whitespace—making it ideal for use when a new page is just taking shape. **Can’t modify the component? Don’t rush to adjust the styles—let’s double-check first.** Copy:核对一下视觉层级. Often, the problem isn’t with the component itself but with how it relates to its surrounding elements; a card may look fine on its own, but it gets lost when placed within a page. This sentence shifts the focus back to “what rank it holds on the entire page.” **I only want to move the text, not the layout.** One more word at the end:只改字号、粗细和颜色,不动布局和间距. This is the most practical sentence, because the most common side effect when AI changes the hierarchy is that it casually rearranges the layout, messing up even the originally correct structure. With this constraint in place, you can safely modify the skin without affecting the skeleton. ## Four truly effective hierarchical measures Prompts are just a starting point; the final judgment still rests with you. There are actually only four types of methods that can create a hierarchy, ranked from highest to lowest cost-effectiveness: - **Font size**:Most direct. There must be at least a 1.5-line space between the title and the body text for it to qualify as a “title.” - **Font weight**Bold text is an inexpensive way to emphasize, but use it sparingly; three instances of bold on a single page are effectively the same as using none at all. - **Color and Contrast**:Use maximum contrast for the main information, and reduce the auxiliary information to about 60%. Varying the shades within a single color family feels more sophisticated than switching to different colors. - **Leaving blank space**: The most easily overlooked, yet the most effective. Give the main content more breathing room, and it will naturally stand out. Position, animations, and shadows all come last. Many pages can't make hierarchy, the problem is not the lack of means, but the use of peace and justice:**Every element was shouting, and no one was heard.** ## One-line workflow Turn the above into a daily workflow you can use every day: First, make sure this page guides users to do just one thing, and surface the core information for that task. Then, use font size, weight, color, and white space to prioritize it above all else; demote all other information. Finally, step back and take a critical look. In the stage of AI assisted creation and VibeCoding normalization, the tool is responsible for generating the scheme, and the judgment still comes from the designer. STOYARD in the brand and digital experience project, the visual hierarchy, information priority and interface acceptance are deposited into reusable working methods to help the team maintain the clarity and brand texture of the interface while producing efficient output. **Keywords**: UI visual level, AI design, VibeCoding, UI judgment, squinting test, prompt words, interface design, design level, generative AI, designer ability **Category**: AI Trends/Design Thinking/UI Design/Working Methods **Source of Information**: STOYARD team open practice views; article published November 22, 2026.