Screenshot To Code

作者 onewave-aifc5b7851a6c7無授權條款325 個星標收錄於 2026年10月8日更新於 2026年10月8日儲存庫6 天前更新

Turns a screenshot, mockup, Figma export, or photo of a UI into working front-end code - React + Tailwind CSS v4 by default, or Next.js, Vue, or plain HTML/CSS - matching layout, spacing, colors, and typography, then renders the result and compares it against the image to close visual gaps. Use whenever the user shares an image of a website, app screen, dashboard, component, or wireframe and wants it built, cloned, recreated, or "made real", or says "code this up", "build this design", or "match this screenshot".

AI 產生的概覽

將介面截圖、設計稿或 Figma 匯出轉換為前端程式碼,並把渲染結果與圖片進行視覺比對。

功能
把使用者介面圖片轉換成可用的前端程式碼,預設使用 React 搭配 Tailwind CSS v4,也支援 Next.js、Vue 或純 HTML/CSS。它引導代理先撰寫簡短的設計規格,再用可重複使用的元件與設計權杖建立響應式且無障礙的標記,接著渲染結果並與原圖比對。最後交付檔案、執行說明,以及一份假設清單,例如推測的字型或自行補充的狀態。
適用情境
當使用者提供網站、應用程式畫面、儀表板、元件或線框圖圖片,並希望將其建置、複製或重現時使用。也適用於「把這個寫出來」「照這個設計做」「照這張截圖還原」這類請求。
執行需求
不包含指令碼,僅為說明文件。渲染與比對步驟需要瀏覽器工具,例如 Playwright 或無頭瀏覽器;目標技術堆疊可能需要 Node.js、Vite、React、Next.js、Vue 或 Tailwind CSS。

Screenshot to Code

Recreate the UI in the image as clean, responsive, accessible code, then check the render against the image.

Workflow

  1. Pick the target stack. If the user is inside a project, match it: read package.json for React/Next/Vue and the Tailwind version, and reuse existing components, tokens, and icon libraries. Do not ask when the repo already answers it. With no project:

    • Single screen or quick prototype: one self-contained index.html (Tailwind via the browser build, see references/stack-setup.md [blocked]).
    • App or multi-screen: React + TypeScript + Tailwind v4 on Vite.
    • User mentions Next.js, SSR, or routing: Next.js App Router.
  2. Write a short spec before coding. Looking at the image, note:

    • Layout regions top to bottom (nav, hero, sidebar, grid, footer) and the grid or flex structure of each
    • Design tokens: 3-6 colors as hex, font family guess and the type scale, spacing rhythm (usually multiples of 4px), corner radius, shadow style
    • Repeated components (cards, list rows, buttons) that should be one component with props
    • What the image cannot show: hover and focus states, mobile layout, real data, content below the fold

    Estimate sizes from the image resolution. If the screenshot is 2x (a Retina capture 2880px wide is a 1440px layout), halve the pixel measurements.

  3. Build. Use semantic elements (header, nav, main, section, button, a), extract repeated pieces into components, and put tokens in one place (Tailwind v4 @theme or CSS custom properties) rather than scattering hex codes. Use real text from the image, not lorem ipsum. For images and logos you cannot extract, use sized placeholders with descriptive alt text. Use an icon library the project already has (default: lucide-react), not hand-drawn SVGs or emoji.

  4. Make it responsive. Build for the screenshot's viewport first, then define how it collapses: multi-column grids stack, navs become a menu button, and type scales down with clamp() or responsive utilities. If the screenshot is mobile, go the other direction.

  5. Render and compare. If you can run a browser (Playwright, a headless browser, or a browser tool), screenshot the result at the source image's viewport width and compare side by side. See references/visual-check.md [blocked]. Fix the largest differences first: layout and alignment, then spacing, then type size and weight, then color. Two or three passes is usually enough; stop when remaining differences are at the level of font rendering.

  6. Deliver. Provide the files, how to run them, and a short list of assumptions (fonts guessed, states invented, content inferred). Name anything that needs real assets.

Stack notes

  • React 19: function components with TypeScript prop types. propTypes checks were removed from React 19; do not add them. forwardRef is not needed for passing ref to function components.
  • Tailwind CSS v4: configuration lives in CSS. Use @import "tailwindcss"; and @theme { --color-brand: #...; }, not a tailwind.config.js and not the v3 @tailwind base/components/utilities directives. Custom tokens become utilities automatically (bg-brand).
  • Next.js (App Router): pages in app/, components are Server Components by default. Add "use client" only to components that use state, effects, or event handlers.
  • Create React App is deprecated. Use Vite for plain React.

Setup commands for each stack are in references/stack-setup.md [blocked].

Worked example

Input: a 1440px-wide screenshot of a pricing section - centered heading, three plan cards with the middle one highlighted, feature checklists, and a button per card.

Spec:

  • Regions: heading block, then a 3-column card grid, max width about 1100px, 24px gap
  • Tokens: background #0B1220, card #111A2E, accent #3B82F6, text #E5E7EB, muted #94A3B8; Inter-like sans; radius 12px
  • Components: PlanCard with name, price, features[], highlighted, cta
  • Unknowns: hover states, monthly/annual toggle behavior, mobile layout

Build: PricingSection.tsx mapping a plans array into PlanCard; highlighted adds an accent border and a "Most popular" label; grid-cols-1 md:grid-cols-3. Check icons from lucide-react.

Compare: the first render has cards 40px too tall because of button padding, and the heading weight is 600 against a 700 in the image. Fix both, re-render, and deliver with an assumptions list: toggle not built, font assumed Inter.

Failure modes

  • Guessing the stack when the repo already defines it, or mixing Tailwind v3 config into a v4 project.
  • Absolute positioning to force a pixel match. It breaks on the first resize. Use flex and grid, and accept small differences.
  • One giant component. Anything that repeats three times is a component with props.
  • Invisible states. Buttons and links need hover and focus-visible styles even though the screenshot cannot show them. Inputs need labels.
  • Color drift. Sample colors from flat areas of the image, not anti-aliased edges or gradients, and check text contrast meets WCAG AA (4.5:1 for body text).
  • Claiming a match without looking. If you could not render the result, say the comparison was not done.

來源與署名

來源:onewave-ai/claude-skills位於screenshot-to-code提交fc5b785

授權條款: 無授權條款

內容歸原作者所有。SourceWeft 從公開儲存庫中收錄這些內容。

檢舉或申請下架