GitHub Podcast: RAG, MCP, and Skills Coexist, Not Compete — The Essence of AI Development is Risk Management and Judgment

The GitHub Podcast analyzed recent claims in the AI development community, such as 'RAG is dead' and 'Skills killed MCP,' stating they are factually incorrect and that each technology represents a coexistence relationship solving different problems.

Roles and Relationships of RAG, MCP, and Skills

• RAG (Retrieval-Augmented Generation) remains a valid core technology that enhances answer accuracy by retrieving information outside the model's training data. Good retrieval helps the model start closer to the correct answer and reduces token waste.
• MCP (Model Context Protocol) provides a standardized interface for agents to access tools and data.
• Skills represent packaged expertise, including team workflows, project conventions, and tool usage.
• These three are not in competition. Agents work together in a single workflow by accessing tools via MCP, following project guidelines via Skills, and retrieving context via RAG.

Developer Competencies and Code Review in the AI Era

• The claim that 'there is no need to read AI-generated code' is false. Developers are responsible for the code, and thorough review is essential for high-risk tasks like production environment authentication refactoring. AI shifts the location of effort but does not eliminate the work.
• Whether AI is used during hiring is not an absolute criterion. What matters is Judgment. Fluency in explaining when to use AI, when to work manually, and how to review generated code is evaluated as a competency.

Codebase Maintainability and AI

• Regarding the claim that 'if fine-tuning is needed for a codebase, the code is bad,' it was pointed out that since modern AI models learn general patterns and architectures, a codebase that AI cannot understand is also difficult for humans to maintain. AI acts as a new stress test for verifying maintainability alongside code review, testing, and onboarding.
• In conclusion, rather than taking permanent sides in AI tool debates, it is important to experiment by presenting practical evidence and trade-offs. The correct approach is to use MCP when standard interfaces are needed, Skills when context management is needed, and RAG when an information base is needed.