Impeccable Makes Claude, Codex and Kimi 10X Better

Julian Goldie explains that while AI coding models like Claude, Codex, and Kimi excel at generating working code quickly, they all produce visually similar designs due to training on startup templates. This results in what he calls 'AI slop'—identical fonts, purple gradients, rounded icon squares, and nested cards that make every AI-built site indistinguishable. Impeccable, created by Paul Bakaus from Anthropic's design framework, solves this by providing a design layer that gives AI actual design intelligence before code generation.

The setup process involves running '/impeccable init' which creates two key files: product.md (defining purpose and audience) and design.md (specifying colors, fonts, and design rules). These files ensure the AI understands context before designing, eliminating guesswork. Impeccable offers 23 distinct commands like '/impeccable bolder', '/impeccable quieter', '/impeccable distill', and '/impeccable typeset' that let users communicate design adjustments in plain language. The tool recognizes four design modes—persuade, operate, read, and experience—and tailors designs to each page's specific function.

The detector component identifies approximately 60 specific design flaws including side tab borders, bounce animations, dark glows, cramped spacing, long text lines, and tiny mobile tap targets. This detector operates three ways: as a command-line tool, as a Chrome extension for analyzing any website, and as a git hook that runs during development to catch problems before they ship. The live mode feature allows real-time visual iteration where users click elements to see three alternative designs and save preferred versions directly to source code. Goldie emphasizes this is not about finding the smartest model but providing direction and standards to the model already in use.