Smelter: A TDD-driven Development Framework for Claude Code, New Practice of Multi-Agent Code Review

Smelter is a test-driven development (TDD) framework designed specifically for Claude Code. It addresses quality control challenges in AI-assisted programming and improves code quality through a 10-step structured workflow, file-based memory system, and multi-agent code review mechanism. Its core lies in combining TDD with multi-agent review, refining AI-generated code into high-quality products like a smelter.

With the popularity of AI programming assistants like Claude Code and GitHub Copilot, developers' work styles have changed, but new problems have also emerged: How to ensure the quality of AI-generated code? How to comply with team norms? Traditional code review struggles with large volumes of AI-generated code—manual review is time-consuming, and pure AI self-review lacks effective quality control. A new development model is needed to balance efficiency and quality.

Smelter's design philosophy is to combine TDD with multi-agent review to solve quality issues in AI-assisted programming. The name 'Smelter' implies refining rough AI code into maintainable, testable high-quality software through strict processes and multi-level reviews.

Smelter defines a 10-step development workflow, each with a clear goal:

1. Requirement Clarification: Clarify requirement boundaries through dialogue with AI;
2. Test Design: Design test cases based on requirements (core of TDD);
3. Skeleton Code: Write basic structure to make tests runnable;
4. AI-Generated Implementation: Generate code guided by tests;
5. Unit Test Execution: Verify the implementation, and feed back failures to AI for correction;
6. Review Preparation: Organize context;
7. Multi-Agent Review: Multiple AIs review from perspectives like functionality, style, performance, etc.;
8. Issue Summary & Fix: Handle review comments by priority;
9. Regression Testing: Ensure fixes do not introduce new issues;
10. Documentation Update: Record decisions and changes.

File-Based Memory System: Persist key information to files, including requirement documents, design decision records, review history, and knowledge bases. Its advantages are persistence (no loss) and auditability (traceable).

Multi-Agent Review Mechanism: Multiple AIs review from different angles: functionality (correctness), style (norms), performance (bottlenecks), security (vulnerabilities), and maintainability (readability). Parallel review ensures comprehensiveness.

Application scenarios of Smelter:

• Enterprise-level projects: Require strict quality and norm reviews;
• Open-source projects: Maintain a unified style and automate checks for external contributions;
• Teaching scenarios: Help understand the importance of TDD and review;
• Rapid prototyping: Iterate quickly while maintaining quality.

Future outlook for Smelter:

• Deep integration with CI/CD pipelines to implement automated quality gates;
• Learn from project history to optimize review strategies;
• Support more AI programming assistants;
• Establish a review knowledge base to accumulate project-specific rules. Provide a new paradigm for developers pursuing high-quality code.