jakubkrehel/skills

A set of eleven agent skills for interface work, each scoped tightly enough to be useful without being so narrow it only fires once. The collection is organized around a hub-and-spoke pattern: better-interface acts as a coordinator that calls every other skill in sequence, while the remaining ten skills can be invoked independently or composed as needed.

The domain coverage is genuinely broad. There are skills for the craft details that tend to slip through code review — border radius consistency, shadow layering, micro-interaction timing — alongside skills for the things that are even harder to catch: whether button labels are actually clear, whether empty states say anything useful, whether error messages help the user or just report a failure. Accessibility gets its own skill rather than being folded into a generic "polish" pass, covering focus states, ARIA, keyboard navigation, hit area sizing, and motion preferences as distinct concerns.

Two skills stand out as particularly practical for iterative work. variant builds multiple genuinely different versions of a UI component behind a picker so you can compare them in the live page rather than in a diff. break renders a component in every realistic state — loading, error, empty, overflow, disabled — and produces a visual report of what failed. Both are user-invoked rather than running automatically, which is the right call: they're expensive to run and you want to choose when they fire.

explain-interface is the most unusual entry. Point it at a URL or a screenshot, name a visual effect you're curious about, and it traces the implementation layers behind it. That's a different use case from the others — less about improving your own work, more about learning from someone else's.

Installation works through either a CLI command or a Claude Code plugin. The CLI path keeps skill names flat (/interface-review), while the plugin namespaces them under the plugin identifier (/interfaces:interface-review). Both install the same underlying skills; the difference is purely in invocation syntax. The CLI path supports selective installation if you only want a subset.

The collection is clearly built for frontend engineers who are already working in agent-assisted environments and want the agent to hold opinions about interface quality rather than just generate code. Whether those opinions are well-calibrated depends entirely on what's inside each SKILL.md file, which the README doesn't expose — so the actual quality of the guidance is opaque from here. But the architecture is sound: narrow skills, clear invocation patterns, a coordinator for when you want the full sweep.