jakubkrehel/skills
github.com/jakubkrehel/skills- Category
- AI Agents
- Rank
- No. 1522Tools index
Previous survey · No. 1558 ·
- Pricing
- Open Source
- Type
- TOOL
- Use case
- Design & Media
- Interfaces
- Agent Skill / Plugin
- Builder
- @jakubkrehel
- GitHub
- 7.8k stars
- Date
About
A collection of agent skills designed to improve various parts of a product, covering animations, typography, layout, and colors. Install via a simple CLI command and let AI agents apply design improvements automatically. Built for developers and designers who want AI-assisted UI polish.
What it does
A set of Markdown instruction files that teach a coding assistant how to judge and build web interfaces. Some hold knowledge about one discipline, such as accessibility, color or type. Others hold a procedure: review a diff, render one component under worst-case content, or build three alternate versions behind a switcher on the real page. An orchestrating reviewer runs the discipline files in a fixed order, starting with accessibility, and merges their results into one ranked table with a Block or Approve verdict.
Why it's ranked here
The rules are specific enough to check. Hit targets cite the WCAG 24 by 24 pixel baseline, focus rings need a 2px perimeter, reflow must hold at 320px width and 200% zoom. The review logic is disciplined too: every finding must cite a file and line, a check that could not run is reported as not verified rather than as a failure, and reports stop at fifteen findings with blockers always listed first. It is MIT licensed and ships to Claude Code, Codex and opencode users from one repository.
What's good
Each rule lives in exactly one discipline, so the reviewers do not contradict each other, and a missing discipline is marked as not reviewed instead of improvised from memory. The change reviewer reads removed lines to catch regressions, separates new problems from ones that predate the change, and refuses to check out a branch because that would overwrite open work. Fixes are ranked cheapest first: delete, use the native element, reuse an existing token, then correct a value, and only then add something new.
Tradeoffs
This is prose for an assistant, not software. The authors state there is no build, lint or test tooling, so whether a rule is followed depends entirely on the model reading it. The guidance targets web stacks: CSS, ARIA and browser behavior, with nothing for native mobile or desktop toolkits. Four procedures only run when a person asks, and two of them write throwaway pages into the project that stay until deleted. Plugin users receive updates only when the maintainer bumps a version number.
How to use it well
Fits a team shipping a web front end that wants an assistant to audit a pull request or a screen against concrete accessibility and design rules before merge. Install the whole set rather than one file, since the orchestrated review marks any absent discipline as not reviewed. Run the change reviewer on a branch, the stress tester on a component before real data arrives, and the variant builder when a layout decision is still open. It does not cover correctness, security or performance; it defers those to a normal code review.
Technical notes+
README.md lists eleven skills, each a directory with a SKILL.md entry point. .claude-plugin/plugin.json and .claude-plugin/marketplace.json both name the plugin interfaces, so plugin users call /interfaces:better-interface while skills CLI users call /better-interface. AGENTS.md defines the ownership table and says interface-review, variant, break and explain-interface set disable-model-invocation: true plus a matching Codex policy flag; better-interface therefore cannot start interface-review and asks the user instead. skills/better-interface/SKILL.md holds the severity ladder, thirteen escalation triggers and the 15-finding cap. skills/interface-review/SKILL.md resolves scope from git merge-base before the dirty tree and allows git fetch but never git checkout or git stash.
Observed
- License
- MIT
- Format
- Markdown skill files only; the repository states it has no build, lint or test tooling
- Install surface
- Skills CLI via npx, and a Claude Code plugin marketplace served from the same repository
- Harness support
- Claude Code, Codex and opencode
- Invocation
- Four skills are user-invoked only; the rest can be invoked by the model
- Scope
- Web interfaces: HTML semantics, CSS, ARIA and browser rendering
Read from README.md, LICENSE, .claude-plugin/plugin.json, .claude-plugin/marketplace.json, skills/better-interface/SKILL.md, skills/interface-review/SKILL.md, skills/better-colors/SKILL.md, skills/better-accessibility/SKILL.md, skills/break/SKILL.md, skills/variant/SKILL.md, AGENTS.md.
What it can do
Apply animation improvements to UI components
Existing UI codebase → UI components with enhanced animations
Improve typography styling across a product
Existing UI codebase → Updated typography styles and font configurations
Optimize UI layout structure
Existing UI layout code → Refactored and improved layout code
Enhance color usage and palette consistency
Existing UI codebase → Updated color definitions and applied color improvements
Install agent skills via CLI
CLI command with skill package reference → Installed AI agent skills ready for use in the project
Automatically apply AI-driven UI design polish
Existing product UI codebase → AI-improved UI with refined design across multiple dimensions
Tags
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.