one-shot-ui

Catch what the eye can't.

Deterministic screenshot diffing for AI coding agents. Turn a reference screenshot into structured data, diff any build against it — pixel, layout, color, and type — then get the exact CSS to fix. Or copy a UI's design language onto a brand-new screen.

The problem

AI agents get UI ~90% of the way there — then stall. The layout looks right, but a card is 8px too tall, a panel is the wrong shade of gray, a shadow is flat, a gap is off by 24px. Asking the model to "look at the screenshot again and fix it" is slow, and you get a different answer every time.

one-shot-ui closes that last 10% deterministically. It extracts structured data from a reference screenshot — layout regions, colors, typography, spacing, design tokens — diffs your implementation against it, and returns specific, ranked fixes, not "make it look more like this."

Set width to 616px (currently 640px)
Change the fill color to #303040.
Set box-shadow to -3px 0px 24px 0px rgba(28, 29, 38, 0.32).
gap: 176px; /* currently ~152px */

Copy-paste CSS, ranked by visual impact — every example above is real output from the run below.

Watch it converge

The run command loops extract → capture → compare → fix until the heatmap goes quiet. In the run above, the agent's first build looked identical to the eye — one-shot-ui flagged 15 concrete deltas (position, size, color, shadow, spacing) and the loop drove the build to ~2.5% pixel mismatch, within ~0.5% of the tool's own estimated irreducible floor (≈2%, sub-pixel font rendering) for this design.

Why one-shot-ui

• Deterministic, not vibes. Stable pixel + structural diff scores — same input, same numbers — so you can gate CI on "is this pixel-close enough?"
• Exact fixes, not nudges. It returns concrete CSS (width: 616px, #303040, gap: 176px), grouped by component and ranked by visual impact.
• Structural, not just pixels. Detects missing/extra elements, position & size shifts, color, shadow, spacing, and typography — and labels which differences are irreducible (anti-aliasing, photographic content) so agents don't chase ghosts.
• Not gameable. converge scores structural fidelity — is the reference's content present, placed, and legible? — not just pixel mismatch, and it refuses to win pixels by hiding, overlapping, or recoloring text into its background. It even flags reference text your build renders but clips out of view (a fixed-height overflow:hidden box) — the failure a pixel diff rewards, because cropped text lowers the mismatch.
• Agent-native. Ships an AGENTS.md (auto-discovered by Claude Code, Cursor, Codex, …) plus a Claude Code skill, so your agent drives it without hand-holding.
• Local & private. Pixel diffing, OCR, and layout extraction all run on your machine. No images leave your box, no API keys.

Copy a UI's style, not just match it

Matching a screenshot pixel-for-pixel is one job. The other: building a different screen that feels like it belongs to the same design system. one-shot-ui extracts the design language from a reference and verifies a new UI conforms to it — deterministically,

... View full README on GitHub