Refine

Run a single refinement pass over the current application. Evaluate the codebase using every applicable skill available in this session, then produce a prioritized list of improvements.

How It Works

1. Detect the stack — Identify the framework, language, and rendering strategy from config files (package.json, next.config.*, app.json, tsconfig.json, etc.)
2. Select applicable skills — From the skills available in this session, pick every one that is relevant to the detected stack. Typical candidates:
   • next-best-practices — file conventions, RSC boundaries, data patterns, metadata
   • next-cache-components — PPR, use cache, cacheLife/cacheTag
   • vercel-react-best-practices — React & Next.js performance optimization
   • vercel-composition-patterns — component API design, compound components
   • vercel-react-native-skills — React Native / Expo performance (only if applicable)
   • web-design-guidelines — accessibility, UI/UX audit
   • agent-best-practices — general engineering quality
3. Invoke each selected skill against the codebase to gather findings
4. Deduplicate and rank — Merge overlapping findings, then sort by impact

Scope

If arguments are provided ($ARGUMENTS), limit the analysis to that path or area. Otherwise evaluate the full application.

Output Format

Present findings as a single table sorted by impact (highest first):

Priority levels

• Critical — Major performance, accessibility, or correctness issue
• High — Meaningful UX or performance win with low effort
• Medium — Noticeable improvement, moderate effort
• Low — Polish or minor optimization

Categories

• Perf — Performance (bundle size, rendering, caching, data fetching)
• UI — Visual design, layout, responsiveness
• UX — Interaction design, navigation, feedback, loading states
• A11y — Accessibility (WCAG compliance, keyboard nav, screen readers)
• Arch — Architecture, component structure, code organization
• DX — Developer experience (types, conventions, maintainability)

Rules

• One pass only — Do not start implementing fixes. This is an evaluation.
• Be specific — Reference actual files, components, and line ranges. No vague advice.
• Show before and after — The “Current” and “Suggested” columns should contain concrete descriptions or brief code snippets, not abstract statements.
• No duplicates — If multiple skills flag the same issue, list it once under the most relevant category.
• Stay honest — If the codebase is already well-optimized in an area, say so. Don’t manufacture findings.
• Cap the list — Aim for 10–20 actionable findings. Quality over quantity.

After the Table

Provide a brief summary:

1. Top 3 wins — The three changes that would deliver the most impact for the least effort
2. Stack health — One sentence overall assessment (e.g., “Solid foundation with room for caching improvements”)