img-to-frontend

name	img-to-frontend
description	Use when the user wants an end-to-end premium visual design workflow for any type of website or web page created directly from images.

Img to Frontend

Overview

Use this skill to run an image-first design-to-code workflow for visually appealing awwward winning websites. It must adapt to the user's brief, audience, content model, interaction needs, and brand intent rather than forcing a fixed category or visual trope.

The workflow is intentionally staged:

1. Invoke $imagegen to create four distinct website/page design images.
2. Stop so the user can select the image or images to use as the design reference.
3. Translate only the selected image into a precise implementation prompt.
4. Build the real page and iterate against screenshots until structure, hierarchy, typography, spacing, color, media treatment, component geometry, interactions, and responsive behavior are close to the reference.

Load references/visual-iteration-checklist.md for design variety, implementation, close visual replication, premium website work, complex UI sections, or repeated visual refinement.

When To Use

Use this skill when the user asks to:

• Create several distinct website or page concepts before coding.
• Turn an image or generated design into a real frontend page.
• Produce a detailed prompt another coding agent can use to build a design.
• Replicate a screenshot as real HTML/CSS/React rather than as an image.
• Iterate on visual details like spacing, scale, typography, media treatment, visual systems, component alignment, interaction states, or responsive breakpoints.

Do not use this skill for ordinary business-logic changes, backend features, simple copy edits, or design critique that will not lead to generated images, a detailed implementation prompt, or frontend code.

Non-Negotiable Image Gate

The first deliverable is always four generated images. Do not jump straight to text prompts, implementation prompts, or code.

Required behavior:

• Load and use the $imagegen skill for phase 1.
• Generate four unique website/page images with the built-in image generation tool by default.
• Make one image-generation call per concept unless the image tool supports equivalent multi-output generation in the current environment.
• Render or present all four images to the user, one at a time.
• Stop after presenting the images and ask which one the user wants to explore.
• Do not write the detailed build prompt until the user selects a specific image.
• Do not implement frontend code until the selected-image prompt phase is complete or the user explicitly skips it.

If image generation is unavailable or fails, report that blocker and ask whether the user wants text-only fallbacks. Do not silently substitute text briefs for the image stage.

Core Workflow

1. Design Exploration

Start by identifying the site context, audience, goal, style constraints, viewport, and must-have sections. If the user already gave enough context, do not pause for more input.

Generate four distinct website/page images through $imagegen. Each image must have its own visual language, not minor palette swaps. Vary structure, hierarchy, typography strategy, interaction model, information architecture, content emphasis, density, and brand behavior.

Choose directions that fit the site's purpose and audience. Different website types should develop from their own communication needs, not from a shared default template.

Use ui-mockup as the image-generation use case unless the user asks for a different raster style. The images should be polished enough for the user to choose between real visual directions, not rough wireframes.

For each direction, briefly label:

• Overall art direction and emotional tone.
• Page layout and content hierarchy.
• Typography treatment and scale relationship.
• Color palette with exact roles.
• Visual centerpiece appropriate to the site's purpose, content, and user task.
• Credibility, context, or trust treatment when relevant.
• Motion idea, if appropriate.
• What makes it meaningfully different from the other three.

2. Selection Pause

After the four images are created, stop and wait for the user to pick one or request revisions. Do not write the detailed build prompt and do not begin implementation until the user selects an image or explicitly asks to skip selection.

When the user references a selected image, verify which image they mean from the conversation context. If there is ambiguity, ask a short clarification before writing the prompt or coding.

3. Deep Build Prompt

Before coding, convert the selected image into a granular implementation prompt. The prompt should be specific enough that another frontend agent can build a near exact replica without guessing.

Include:

• Canvas and viewport assumptions, including desktop target size and responsive requirements.
• Exact page structure, grid proportions, maximum widths, gutters, and alignment rules.
• Typography hierarchy, font choices, letter spacing, line-height, weights, italics, uppercase treatments, and no-wrap constraints.
• Color tokens for background, surfaces, borders, primary accents, muted text, active states, status states, shadows, and media treatments.
• Component inventory covering only the modules that actually appear in the selected design, described by their role, hierarchy, layout behavior, and content type.
• Detailed behavior for any complex visual systems present in the selected image, including their structure, anchors, scale rules, states, responsive behavior, and fallback behavior.
• Animation rules that are restrained and purposeful.
• Explicit "do not" constraints that prevent fake implementation, generic styling, broken geometry, unreadable content, and overlapping responsive modules.
• Acceptance criteria and screenshot comparison checklist.

4. Codebase Discovery

Inspect the project before editing. Determine the framework, routing model, styling system, package manager, existing design tokens, component patterns, image/font handling, lint/test/build commands, and primary codepath.

Respect existing architecture and design-system conventions. If the user is asking about library or framework docs, use the appropriate documentation skill before relying on memory.

If a reference image includes browser chrome, operating system UI, or mock browser decorations, treat that chrome as presentation context only. Build the actual page content, not the browser wrapper, unless the user explicitly asks for a browser mockup component.

5. Implementation

Build the selected design as real UI code.

Requirements:

• Use semantic structure and accessible interactive elements.
• Define design tokens for colors, spacing, typography, borders, shadows, media treatment, and breakpoints.
• Prefer real CSS/SVG/HTML for visual marks, controls, structured visuals, content modules, and other interface details.
• Do not flatten the design into a static image.
• If the design includes linked, anchored, or spatially related elements, keep them in a shared coordinate or layout system so they resize together.
• Use responsive layout rules that shrink complex UI first, then stack or simplify before overlap.
• Preserve premium whitespace while avoiding unusable dead zones on large monitors.
• Keep copy readable and avoid truncation unless the design explicitly calls for it.
• Add only meaningful motion: page-load reveals, subtle hover states, media transitions, scroll reveals, or interface ambience when it supports the design.

6. Visual Iteration Loop

Run the page locally and capture screenshots at the reference viewport plus at least one wide desktop, one constrained desktop/tablet, and one mobile width when feasible.

Loop until the implementation has no obvious visual mismatch:

1. Compare the screenshot against the selected reference image.
2. Identify the largest visible mismatch first: composition, scale, alignment, typography, spacing, color, imagery, component geometry, complex-widget behavior, or responsive behavior.
3. Patch the implementation.
4. Re-run the page and capture a new screenshot.
5. Repeat until additional changes are minor refinements rather than structural corrections.

Pay special attention to failures that commonly appear in screenshot replication:

• Content is bunched together because component scale is too large.
• Large monitors have excessive deadspace because the layout max-width is too narrow.
• Orientation elements are offset from the main content system.
• First-read text elements do not share the alignment, rhythm, or hierarchy shown in the reference.
• Action labels wrap, icons drift, or control geometry breaks.
• Repeated content, media, input, or trust modules collide with adjacent sections.
• Dense component labels truncate because internal spacing is too tight.
• Anchored or connected visual elements separate from the objects they are meant to describe.
• Complex widgets overlap before the layout stacks or simplifies.
• Mobile layout waits too long to stack.

Ensure the accuracy of your final output, especially before yielding back to the user. The mock generated image and the final output should look nearly identical. Focus on the details and precision.

7. Verification And Closeout

Run the most relevant checks available for the project: lint, typecheck, build, unit tests, and visual browser verification. If a check cannot run, state why.

Final response should include:

• What was built or changed.
• Files changed.
• Verification commands and results.
• Any residual visual risks, especially if screenshot comparison was limited by tooling or missing references.

Quality Bar

The output should feel like a premium website, not a generated template. Favor confident hierarchy, intentional palette, sharp spacing, precise alignment, brief-appropriate content modules, and a single strong visual thesis.

When matching a reference, optimize for visible fidelity over code cleverness. Typography, spacing, alignment, imagery, component geometry, and responsive behavior are first-class deliverables, not polish.

When the user says something looks wrong, treat it as direct visual evidence. Inspect the current implementation and reference, then patch the specific mismatch rather than explaining it away.