The Code Review Skill is a production-ready, modular skill designed for Claude Code, transforming AI-assisted code review into a structured, consistent, and expert-level process. It covers over 11 programming languages and frameworks, encompassing more than 9,500 lines of curated review guidelines that are progressively loaded to minimize context window usage during operation.

The core methodology of this skill revolves around a context-aware, multi-phase review workflow coupled with on-demand, granular guideline retrieval to optimize Large Language Model (LLM) interaction and ensure comprehensive, actionable feedback.

Key Features and Methodology Breakdown:

1. Progressive Disclosure: This is a crucial technical optimization for LLMs. The system operates with a lean core skill (SKILL.md - ~190 lines). Language-specific and domain-specific guides (e.g., react.md, java.md, security-review-guide.md) ranging from 200 to 1,000 lines each are loaded dynamically only when explicitly requested by the user's prompt or inferred from the review context. This intelligent loading strategy prevents context window overflow and focuses the LLM on relevant review criteria.

2. Four-Phase Review Process: The skill guides the AI through a structured, sequential workflow, ensuring a systematic and thorough review:
   • Phase 1: Context Gathering: The AI first analyzes the Pull Request (PR) scope, linked issues, and the developer's intent. This initial phase establishes a foundational understanding for the subsequent analysis.
   • Phase 2: High-Level Review: This phase focuses on architectural implications, potential performance impacts, and the robustness of the testing strategy. It aims to identify macroscopic issues before diving into code specifics.
   • Phase 3: Line-by-Line Analysis: This is the detailed inspection phase, where the AI evaluates code for logic correctness, potential security vulnerabilities (leveraging dedicated security checklists), maintainability, and handling of edge cases.
   • Phase 4: Summary & Decision: The final phase involves synthesizing findings into structured feedback, determining the approval status, and outlining clear action items for the author.

3. Severity Labeling: Every finding identified by the AI is meticulously categorized using a predefined set of severity labels, allowing for clear prioritization and communication:
   • 🔴 blocking: Must be fixed before merge.
   • 🟠 important: Should be fixed; may block depending on context.
   • 🟡 nit: Minor style or preference issue.
   • 🔵 suggestion: Optional improvement worth considering.
   • 📚 learning: Educational note for the author.
   • 🌟 praise: Explicitly highlights excellent work.

4. Specialized Guidelines and Domains: The skill's knowledge base is extensive and highly specialized, stored in a modular reference/ directory. It includes:
   • Frontend: React 19 / Next.js / TanStack Query v5, Vue 3.5 + Composition API, CSS / Less / Sass, TypeScript. Highlights include specific patterns for React Hooks (e.g., useActionState, useOptimistic), Server Components, and strict TypeScript configurations.
   • Backend: Java 17/21 + Spring Boot 3, Python, Go, Rust. Advanced concepts covered include Java Virtual Threads, Spring Boot ProblemDetail, Go goroutine management and context.Context propagation, and Rust ownership/borrowing, async/await safety, and unsafe code review requirements.
   • Systems: C, C++, Qt Framework. Specific checks involve C pointer/buffer safety, C++ RAII and move semantics, and Qt object model and signal/slot threading.
   • Architecture & Performance: Dedicated guides for Architecture Design Review (SOLID, anti-patterns, coupling/cohesion) and Performance Review (Core Web Vitals, N+1 problem, memory leaks).
   • Security-First Approach: Integrates dedicated security checklists (security-review-guide.md) across all language ecosystems.
   • Best Practices: Includes code-review-best-practices.md for communication and process guidelines, promoting a collaborative tone (questions over commands).

5. Automation Awareness: The guidelines are designed to differentiate between issues that human review should catch (e.g., architectural flaws, security logic) versus those typically handled by automated linters or formatters, ensuring efficient use of the AI's capabilities.

The skill is activated within Claude Code sessions by commands like "Use code-review-skill to review this PR," intelligently loading relevant knowledge modules based on the prompt's context (e.g., "Review this React component" loads react.md, "Security review of this Go service" loads go.md and security-review-guide.md). This modular design and structured execution make the AI-assisted code review process highly effective and consistent.