A structured PR review workflow that combines language-agnostic quality checks with framework-specific references. It walks through five steps: understanding changes, security and logic validation, language-specific best practices (TypeScript internal, React/Next.js via Vercel's external skill), and an approval decision based on a point system where Critical issues are automatic rejections. The interesting piece is the CI optimization: a triage phase using Haiku generates a .pr-triage.json that tells the main review which references to load and which surface issues to skip, cutting token costs significantly. It also supports incremental reviews on PR updates by caching previous state. Built for GitHub Actions but works standalone, with special handling for authorization patterns and PostgreSQL RLS when detected.
npx skills add https://github.com/xtone/ai_development_tools --skill code-review