Design image to code

Design — Image To Code

Reproduce a reference image as a real, runnable frontend — verified by screenshot diff, not "looks roughly like it". The discipline that separates this from generic codegen: you measure the result against the reference and iterate until the gap is minor and explainable.

When to use

• You have a mockup/screenshot (from a designer, design-ideate, or generated) to implement.
• "Build this design", "code this screenshot", "make this real".
• Recreating a UI you can see but don't have code for.

When NOT to use

• The source is a live site (you can hit its URL/HTML) → design-url-to-code (it can extract real CSS).
• No image yet, just an idea → design-ideate.
• You need interactivity/flow wiring, not a static match → design-prototype (run this first to build screens).

Inputs

• The reference image(s) (path).
• Target stack: ask; default to matching the repo, else React + Vite + Tailwind + TS.
• ./design/context.md constraints if present.

Workflow

1. Inspect the image with a vision/read tool. Inventory every meaningful element. Extract the design system: layout grid, type scale, spacing rhythm, palette, components, radii, shadows.
2. Reuse before inventing: if adapting an existing project, read its components/tokens/fonts first and use them.
3. First pass: build to match the primary viewport as closely as possible. Don't chase responsive yet.
4. Capture: run the app, screenshot at the reference's exact dimensions via a browser-automation tool.
5. Compare & list mismatches in priority order: layout → type → spacing → color → icons.
6. Iterate until differences are minor and explainable. Then handle responsive (tablet/mobile).
7. Missing assets: use supplied files; generate raster assets with an image tool; or use clearly-marked placeholders + a note.

Worked example

reference: ./mocks/pricing.png  (1440px wide)
stack: matches repo → Next.js + Tailwind + Aurora DS
pass 1: built 3-tier pricing grid, header, CTA
capture: screenshot @1440 → ./design/checks/pricing-impl.png
mismatches:
 - [layout] cards 3-col but gap too tight (16 vs ~32 in ref)
 - [type] tier price 32px, ref ~40px
 - [color] "popular" badge wrong accent (#7C3 vs ref #16A34A)
 - [spacing] section pad 48 vs ref ~80
iterate: fix gap=32, price=40px, badge=#16A34A, section pad=80
re-capture: gap now minor (sub-pixel font rendering only) → DONE
responsive: collapse to 1-col < 768px, stack CTA
verify: build clean, typecheck clean, runs at :3000

Quality bar / Definition of done

• Implementation screenshot ≈ reference at the target viewport; remaining diffs are minor and explained.
• Existing design-system components/tokens reused, not duplicated.
• Build + typecheck clean; runs locally with one command.
• Responsive handled after the reference-size match is solid.

Common pitfalls

• Claiming parity without comparing — always capture + diff. "Looks close" isn't verification.
• "Improving" the design — match the reference unless explicitly asked to deviate; note any forced deviations (e.g. for a11y).
• Inventing new tokens when the project already has a scale — creates drift.
• Chasing responsive too early — nail the reference viewport first, or you iterate on a moving target.
• Eyeballing spacing/type — measure against the reference; "looks like 16" is often 24.

Handoff

Static screens done → design-prototype (wire interactivity/flow) → design-qa (conformance) → design-share.

Tooling

Needs a vision/read tool (inspect the image) and a browser-automation tool (screenshot the build for comparison). Optional image tool for missing assets. Without a browser tool you can still build, but flag that visual parity is unverified.