Claude Dev Workflow: A Layered Workflow Practice for AI-Assisted Development

With the popularity of AI programming assistants like Claude Code, how to maintain efficiency without losing engineering rigor has become a key issue. Claude Dev Workflow is a practical solution designed to address this problem. It introduces engineering discipline to AI-assisted programming through XS/M/L three-level task classification, RFC gating, multi-agent code review, and progressive security levels. The core is to use AI to enhance rather than replace traditional engineering practices.

This project is essentially the author's encapsulation of daily development workflows into a Skill that Claude Code can automatically load and execute. It is a set of conventions and processes rather than a framework. It applies to all code projects such as CLI tools, MCP servers, SaaS prototypes, etc., and has been open-sourced to provide a reference for developers who want to maintain discipline in AI-assisted development.

Three-level Task Classification: XS (extra small) quick fixes require smoke testing; M (medium) feature development requires result images, RFC documents, code review sub-agents, and user approval; L (large) architecture changes require additional security reviews, phased plans, and comprehensive checks. Pre-flight Check: Sync branches, list PRs, and read backlogs to avoid outdated code and duplicate work. Result Image: Mandatory description of requirements from the user's perspective; RFC Gating: Design first before implementation to prevent architectural debt.

Multi-agent Code Review: Launch sub-agents to check readability, boundary cases, over-engineering, and key leaks. Progressive Security Levels: S1 (local tools) runs security reviews and key detection; S2 (pre-production SaaS) performs SAST static analysis, OWASP checks, and threat modeling; S3 (production environment) conducts penetration testing and external audits. No-blame Incident Log: Record events, root causes, and improvement measures to learn from mistakes.

Installation: Clone the repository to a persistent location and create a symbolic link to the Claude skills directory. Supporting Projects: Complementary to the claude-scaffold-project to generate the required directory structure. Typical Examples: XS tasks (fixing typos) are completed quickly; M tasks (adding adapters) require result images, RFCs, and user approval; L tasks (architecture migration) require additional security reviews and rollback plans.

Limitations: Not a framework, not a productivity tool, not for PM tasks; mainly targeted at JS/Node and Python but the logic is universal. Summary: This workflow injects traditional software engineering discipline into AI programming assistants, demonstrating how to integrate AI into rigorous processes rather than letting tools dominate. It is the direction of disciplined AI-assisted development that the industry needs.