GitHub radar
One consistent animated icon style for your project — no designer
Kubai Icons teaches your AI coding assistant to draw small animated icons in one consistent style instead of a mismatched grab-bag. Ask the agent and get a matching icon set for your project, plus an open library of 2,800+ ready-made icons.
It's a skill (instruction set) for Claude Code, Codex and similar AI assistants, plus an open library of 2,800+ animated icons (SVG, React, Vue).
When it helps
Useful if you're building a site or app with an AI agent and want consistent icons without hiring a designer. Skip it if you already have a finished icon set or a designer.
Pros
- Icons come out in one consistent style, not mismatched
- Icons animate on hover — a folder opens, a lock springs
- Free to use even in commercial products, MIT licensed
- No designer needed — the agent draws them itself
Cons
- Needs a model that can read images, e.g. Claude Opus 5.5
- Drawing a new icon takes several agent self-check rounds
- The export script's names and palettes are in Chinese
How to set it up — step by step
- 1Open your AI agent (Claude Code or Codex) inside your project folder.
- 2Send the agent the text from the block below.
- 3Grant the agent extra permissions if it asks for them.
- 4Check that a SKILL.md file appeared in the skill folder.
- 5Restart the agent and ask it to draw icons for your project's main features.
- 6For reliable results, use a model that can read images, such as Claude Opus 5.5.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the Kubai Icons skill from https://github.com/kubai087/kubai-icons: clone the repository into your skills folder (~/.claude/skills in Claude Code, ~/.agents/skills in Codex), make sure SKILL.md is there, and draw icons for this project's main features in one consistent style. Ask me if you need any extra permissions.
Runs on a regular laptop — the coding agent you already use does the drawing, no extra software to install. For reliable results the author recommends a model that can read images, such as Claude Opus 5.5.
Open on GitHub▌ More finds