image-to-code

作者: openai

将选定的图像、截图、模型或图像生成参考实现为忠实的响应式前端,在产品设计获取上下文确认后…

npx skills add https://github.com/openai/role-specific-plugins --skill image-to-code

Image to Code

You're tasked with translating the visual target image into a high-quality, interactive website or web app.

Critical Overrides

User Context

Before starting, load $user-context and run its preflight script when local shell access is available.

Use saved product URLs, Figma files, screenshots, reference images, codebase paths, Storybook, tokens, design systems, brand assets, component refs, browser preferences, and share targets as grounding material when relevant.

Do not inspect every saved reference. Inspect only what the current task needs.

[IMPORTANT] Previewing prototypes in ChatGPT Work Mode

Starting sites-preview is not verification. Verification requires opening http://terminal.local:4173/ in the cloud browser, inspecting the rendered page, testing primary interactions, checking browser console errors, and passing design QA.

Do not substitute HTTP health, build success, preview-service status, or deployment success for browser verification. If the cloud browser cannot be used, report verification as blocked.

For local prototype verification and design QA in ChatGPT Work Mode:

  1. Install dependencies if needed. The project must have an npm dev script.
  2. sites-preview runs npm run dev -- --host 0.0.0.0 --port 4173 --strictPort. The dev script must accept those flags.
  3. For Vite or Vinext projects, use "dev": "vite". Do not use vinext dev. Configure Vite with server.host: "0.0.0.0" and server.allowedHosts: ["terminal.local"].
  4. From the site root, run sites-preview start "$PWD".
  5. Open http://terminal.local:4173/ in the cloud browser. Do not use localhost, 127.0.0.1, 0.0.0.0, HTTPS, or another port.
  6. Verify the rendered site and its primary interactions before reporting completion.
  7. Never mention terminal.local or output that URL to the user. Deploy a checkpoint with @Sites and give the user only the deployed checkpoint URL.

Mobile prototypes

For a mobile app or phone prototype, build and verify at 390 × 844 unless the user names a device. Use the starter's .mobile-prototype wrapper. Do not use min-height: 100vh on the app shell. The app surface must end at the 390 × 844 frame. Every screen must fit 390px wide with no horizontal scroll, clipped text, clipped controls, or off-screen primary actions.

Workflow

CRITICAL: THIS IS NOT GUIDANCE. THIS IS A CHECKLIST TO COMPLETE.

  1. Do not start unless you have a selected image, screenshot, mockup, or Image Gen result to recreate. A written brief is not enough.

  2. Resolve the exact selected visual target before building.

    • If the user selected a numbered $ideate option, use the Nth displayed generated-image result from the most recent ideation set. Do not use the original concept planning order or Image Gen prompt submission order.
    • Use the concept-name list from $ideate only when it was explicitly written in the same displayed-image order.
    • A generated-image result ID, selected image attachment, screenshot, mockup, or Figma frame is stronger than a bare ordinal. Prefer that exact reference when available.
    • If the selected result cannot be resolved unambiguously, stop before implementation and ask the user to name the concept or reattach/select the image. Never guess and build a nearby option.
  3. Treat the resolved image as the design to recreate.

  4. If the provided design is a mobile viewport, build a mobile app. If it's unclear, default to desktop.

  5. Review the reference design, catalog every image asset in the design, and use the Image Gen tool to create individual images for each one. Zoom in so you can catch every asset that needs to be generated.

    Examples include:

    • Hero images including full bleed image backgrounds
    • Featured article imagery
    • Thumbnails
    • Decorative illustrations
    • Textures and background motifs
    • Logos
    • Product images
    • Avatars

    Rules:

    • CRITICAL RULE: Do not create custom div art, CSS art, inline SVGs, handcrafted SVGs, HTML element drawings, div/span shapes, CSS drawings, gradients, emoji, or text glyphs instead of real icons and image assets ever. Use the built-in Image Gen tool for images and the closest matching icon library for icons.
    • If text is part of an image asset, keep it in the image asset. Examples include full bleed hero images, signs, posters, packaging, storefronts, article art, and illustrations where the type belongs to the visual itself. Do not crop the background image and recreate that text with transparent text boxes, HTML, CSS, or separate overlay layers unless the source clearly shows editable UI text sitting on top of the image.
    • Do not use generic placeholders where the reference implies custom visual content.
    • Generated assets must share the same art direction, palette, rendering style, and design language as the reference mockup.
    • The built-in Image Gen tool does not support transparent images; post-process generated assets when transparency is required.

Parallel asset production

After cataloging and measuring the reference assets, spawn up to three asset subagents while the main agent builds the app structure.

Give each subagent one raster asset task at a time with its reference crop, exact dimensions, focal point, style, output path, and consuming component. Asset subagents generate, inspect, save, and report the asset path only. They must not edit source code, run the browser, or deploy.

Prioritize critical above-the-fold assets first, then reuse agents for supporting assets. Do not delegate standard UI icons or supplied brand logos.

  1. Define all sections of the page. For each section, meticulously measure the layout, spacing between elements, and the size and space of the elements themselves.

  2. Find freely available fonts that match the target design.

  3. Find a freely available icon library that matches the target design. Do not default to Lucide icons. Search for the best match.

    Rules:

    • CRITICAL RULE: Do not create custom inline SVGs, handcrafted SVGs, HTML element drawings, div/span shapes, CSS drawings, gradients, emoji, or text glyphs. Use the built-in Image Gen tool to generate assets and use the closest matching icon library for icons.
  4. Build the app starting with local-prototype-preflight. Unless the user asks for a static mock, full production behavior, or a different scope, bring the app or website to life with:

    • Working navigation, links, tabs, menus, and primary CTAs.
    • Functional inputs, filters, toggles, selections, and forms shown in the main experience.
    • Visible UI states: hover, focus, selected, open/closed, loading, empty, and success where relevant.
    • The main task, conversion path, or user journey working from start to finish when the product has one.

    Controls outside the core experience may be visual-only. Do not build auth, persistence, backend/API calls, integrations, or exhaustive edge cases unless requested.

    Rules:

    • Place every image asset you generated into its position before proceeding. I repeat, replace all placeholders, including CSS/SVG placeholders, before proceeding.
    • Do not leave controls in the core experience as static chrome. Do not create new pages or routes unless the user asks for them.
  5. Run the local app.

  6. Capture the local app using the Browser Choice rule in $index.

  7. Run design-qa as the blocking build gate.

    Steps:

    • Open the reference image and the latest prototype screenshot before writing the QA report.
    • Compare the same viewport and the same interaction state. If they do not match, capture the missing view first.
    • Save the QA report as design-qa.md in the project root.
    • Fix P0/P1/P2 issues, capture the app again, and repeat until the QA report says final result: passed.
    • Do not keep looping on P3 polish. Include any remaining P3s as follow-up iteration notes.
    • If source capture, prototype capture, or visual comparison is blocked, stop. design-qa.md must say final result: blocked.
    • Do not hand off unless design-qa.md exists and says final result: passed.
  8. Handoff the app or website.

    • Only hand off after design-qa passes.
    • Keep the prototype running locally.
    • In Codex Desktop, provide the clickable local URL. In ChatGPT Work Mode, deploy a checkpoint with @Sites and provide only its deployed URL; never output terminal.local.
    • After the prototype link, use the shared build handoff from critical-overrides.md. Do not add a different completion message.
    • Include the post-build iteration and share nudge from critical-overrides.

来自 openai 的更多技能

user-context
openai
加载或管理数据分析插件的持久化源路由偏好、引导逻辑、设置进度及语义层注册表。
official
notion-research-documentation
openai
研究Notion内容,并将其综合成带有引用的结构化简报、报告或对比。通过定向查询搜索并获取Notion页面,然后按主题组织发现,附带内联来源引用和参考文献部分。根据范围和用户目标,从四种输出格式(快速简报、研究摘要、对比、综合报告)中选择。使用内置模板创建和更新Notion页面;直接链接来源,并在新信息到达时跟踪变更...
official
rcsb-pdb-skill
openai
提交紧凑的RCSB PDB请求以获取核心元数据、Search API查询和FASTA下载。当用户需要简洁的RCSB摘要时使用;保存原始JSON或…
official
pdf
openai
PDF的读取、创建与验证,支持可视化渲染与程序化生成。使用Poppler(pdftoppm)将PDF页面渲染为PNG,以便在交付前直观检查布局、间距与排版;通过reportlab程序化生成PDF,确保格式可靠;利用pdfplumber或pypdf提取文本与元数据。执行质量标准:无文本裁剪、元素重叠、表格损坏或渲染伪影;仅使用ASCII连字符,引用内容需可读。使用...
official
test-coverage-improver
openai
改进OpenAI Agents JS mon
official
playwright
openai
基于终端驱动的浏览器自动化,支持元素快照与交互式UI工作流。通过playwright-cli包装脚本运行(需npx),支持无头模式与有头模式进行可视化调试。核心工作流:打开页面、获取快照以稳定元素引用、使用引用进行交互、在导航或DOM变更后重新快照。包含表单填写、点击、输入、多标签页管理、截图/PDF捕获及用于流程调试的追踪记录。元素引用(如e3、e15)...
official
ukb-topmed-phewas-skill
openai
通过接受rsID、GRCh37或GRCh38输入并解析为所需的GRCh38查询,获取单个变体的紧凑型UKB-TOPMed PheWAS摘要。当需要…时使用。
official
code-review-context
openai
模型可见上下文
official