Impeccable

The missing design vocabulary for agents.

Impeccable — the missing design vocabulary for agents

Website: impeccable.style GitHub: github.com/pbakaus/impeccable Stars: 64k | Forks: 3.4k | License: Apache 2.0 · (star_count 64,289 as of 2026-08-31)

What it is

A design language and agent skill from Paul Bakaus that strips the "slop" from AI-generated interfaces. It gives coding agents (Claude Code, Cursor, GitHub Copilot, Gemini CLI, Codex CLI, Grok Build, Antigravity, OpenCode, Pi) precise commands to steer design output and iterates visual variants live in the product.

The pitch: before/after demo after demo — generic AI beige, purple gradients, ghost cards, side-tab borders, vague headlines, too many fields — replaced with intentional design: sharp typography, clear hierarchy, deliberate color, real content.

How it works

Hooks inspect UI edits in real time. Each edit gets run through slop detectors; findings come back to the agent; the agent corrects on the second pass. The same detector can fail a PR in CI.

Per-harness builds. Same skill, recompiled per target — builds for models with known tells carry extra slop rules banning that model's habits (Gemini build kills image-on-hover motion; Codex build refuses ghost-cards and over-rounding).

Command vocabulary. /polish, /distill, /clarify — and the more thorough mode pushes through multiple iterations.

DESIGN.md integration. Impeccable reads a project's DESIGN.md (the Google Stitch-compatible spec — YAML design tokens + markdown rationale) so the agent knows what a color is for, can validate against WCAG, and doesn't guess intent.

Install

Claude Code:

/plugin marketplace add pbakaus/impeccable

Then open /plugin and install Impeccable from the list.

npm / other harnesses:

npx impeccable install

Requires Node 22.12+.

GitHub Copilot: Built in — enable under Settings → Experimental.

Tools in the ecosystem

Why interesting

This is the most-watched design-for-agents project right now (61k stars, 3.7k forks). It sits at the intersection of several things the repo already tracks: the slop-problem in AI-generated UI (cf. variant, frontend-slides), design-system-as-context-for-agents (cf. design-md), and the emerging category of agent-native design tools (cf. claude-design, stitch-google, aura-build).

The hook-based detector model is the interesting technical piece — rather than asking the agent to "be less slop" (which doesn't work), it intercepts edits and enforces rules deterministically, then hands the findings back. Same idea as Vercel's agent lint rules vercel-product-design-agents, but productized and broad.

By

Paul Bakaus (@pbakaus)

Built with LogoFlowershow