What is luban-skill?

Luban is a skill for AI coding agents that helps you improve other AI agent skills you have written. The name comes from a legendary Chinese master craftsman, and the tool is framed around the idea of taking a rough but working piece to a workshop for proper refinement rather than just asking an AI to "make it look better." The core problem it addresses is a common one: you build an AI skill that works for yourself, publish it on GitHub, and nobody installs it because the documentation is unclear, there is no evidence it works, and you have no structured way to figure out what to fix first. Luban responds to this with a five-step process. First it questions whether the skill is worth polishing at all, rather than starting changes immediately. Then it searches for comparable skills online to establish a benchmark with real URLs. Next it scores the skill against three measurement axes: structure, real-world test results, and whether the skill is actively maintained. After that it makes changes one at a time, each passing a validation gate before being kept. Finally it leaves a post-publish observation checklist for the next improvement round. What Luban delivers is a 13-section polishing report covering benchmark comparisons, an ecosystem position assessment, a scoring rubric, three specific improvement directions with one recommendation, and ready-to-use rewritten sections. It also produces a shareable summary card showing before-and-after scores, and it converts any one-off validation scripts into permanent repository tools with written-down standards. Installation is one command: npx skills add LearnPrompt/luban-skill -g. Claude Code users can also install it through the plugin marketplace. After installing, you activate it by telling your agent "let Luban review this skill" and pointing it at a skill directory, a GitHub repository link, or the contents of a SKILL.md file. The repository includes a detailed case study showing Luban used on the ai-news-radar project (about 1,000 stars), where the workflow found an 8-day data pipeline outage hidden behind green CI, removed 327 false AI-generated entries from 83,725 historical records with zero errors, and cut the number of cards rendered on the first screen by 30 percent. Every number in the case study links to the actual pull request.