UI Craft gives AI coding agents something they lack: design taste. Agents write code fast, but they don't know what looks good. The result is those recognizably "AI-made" interfaces — purple gradients, bounce animations, ALL CAPS everywhere. UI Craft attacks that problem before the code is even written.
You install it as a skill (or CLI) in your agent, and it works across five harnesses: Claude Code, Cursor, Codex, Gemini and OpenCode.
The problem it solves
You ask an agent for a dashboard and get back something that "works" but that no designer would ship. It's not a templates problem: the agent is missing the craft knowledge that separates a considered interface from a generic one. UI Craft encodes that knowledge and applies it systematically.
How it works
UI Craft relies on three layers that work together:
- Anti-slop detector: checks every interface against 33 rules — gradient glows, bounce easing,
transition: all, filler ALL CAPS — and flags them before they ship. - Scored quality gate: an MCP gate scores every surface 0–100 across anti-slop, tokens and accessibility. Below the bar, it blocks.
- Convergence loop:
/finalize --convergedoesn't just report — it fixes and re-runs until the gate is green.
What's inside
- 31 design domains across five clusters: visual layer, behavior, surfaces (dashboards, forms, AI chat, data viz), critique, and a "decision spine."
- 22 slash commands, each a focused pass (
/craft,/audit,/polish,/animate,/harden...). - 4 MCP gates and 2 review agents (design-reviewer and a11y-auditor).
- Three dials — CRAFT_LEVEL, MOTION_INTENSITY and VISUAL_DENSITY — that change the agent's behavior per project.
The core test
Every interface is measured against a single question: "Would someone believe AI made this?" If the answer is yes, it starts over. That binary bar is what turns UI Craft into a design quality gate you can even wire into your CI.
Results
- A deterministic design score you can use to fail a CI build.
- Compatible with 5 agents from a single install.
- Free and open source, built on the experience of SmoothUI.
Project link: skills.smoothui.dev
