GitHub's trending page on Thursday was dominated, once again, by "agent skills" — small plug-ins that change how coding assistants write code — and near the top sat a repository making a pointed promise: teach your AI harness to design properly.
Impeccable, by developer Paul Bakaus, ships as a single skill with 24 commands and 61 deterministic detector rules. The target is a well-documented failure mode. Most coding models were trained on overlapping SaaS templates, so left alone they converge on the same handful of "tells": Inter for everything, purple-to-blue gradients, cards nested inside cards, grey text on coloured backgrounds, and a rounded-square icon tile above every heading.
The project's answer is a shared vocabulary. After a one-time `/impeccable init` writes a PRODUCT.md recording durable product facts — audience, purpose, constraints, voice — commands such as `polish`, `audit`, `critique`, `distill`, `animate`, `bolder` and `quieter` let a developer ask for specific design moves rather than hoping the model guesses. A `harden` command targets error handling, internationalisation and text overflow; `onboard` concentrates on first-run flows and empty states.
Unlike most skill packs, Impeccable also works without a model at all. Its engine ships as a standalone binary: `npx impeccable detect src/` scans a directory, a file or a live URL and returns machine-readable JSON, with exit codes tuned for CI (0 clean, 2 findings, 1 scan failure). The project is candid that this is a floor, not a ceiling — a clean detector run, its own documentation warns, is "evidence, not proof of visual or accessibility quality".
The repository is Apache 2.0 and installs across Cursor, Claude Code, GitHub Copilot, Codex CLI, Gemini CLI, Grok Build and a dozen other harnesses, with provider-native hooks that surface findings as an agent edits UI files. On GitHub's daily trend list it showed roughly 74,000 stars with 717 added in the past day.
The subtext is a maturing market. For two years the question for AI coding tools was whether the code runs. With agents now shipping interfaces at volume, the next differentiator is not whether the code is correct but whether the result is distinguishable from everything else the model has ever produced.
