pbakaus/impeccable spots AI UI cliches with 62 rules

impeccable targets a familiar problem with AI coding agents such as Claude Code, Codex and Cursor: their front-end output tends to look alike, with heavy use of the Inter font, purple-to-blue gradients, and nested cards without purpose. The project offers two responses—slash commands like /impeccable audit, /polish and /critique for agents, and the detect CLI that checks HTML, CSS or URLs against 62 rules without an LLM.

The author's test used the detect command, which runs via npx and needs no browser for static files. The 94 primary findings and 46 advisories looked alarming at first, but nearly all came from two color variables in a shared template: the eyebrow text color #626d80 on a dark background measured 3.3:1 contrast against the 4.5:1 WCAG AA threshold. The advisory notes concerned terminal text at 4.2:1, just under the bar, and the two overused-font flags for Fraunces were treated as an intentional design choice rather than a bug.

Because the template is shared across the 100-day series, the author edited only a duplicate file, scaling the color to #76839a to reach 4.51:1. The before-and-after runs on that copy showed the detector is deterministic, with the same input producing the same result each time—relevant for teams that want reproducible checks in a pipeline.