Should you read the code, is RAG dead, and did Skills kill MCP?

The GitHub Podcast's latest episode — "Should you read the code, is RAG dead, and did Skills kill MCP?" — is a masterclass in asking the questions the AI industry would rather you not think too hard about. As reported at https://github.blog/ai-and-ml/should-you-read-the-code-is-rag-dead-and-did-skills-kill-mcp/, the episode taps into three of the most contentious debates currently rattling around the open source AI ecosystem. None of these are settled science, and the fact that a platform like GitHub is willing to host the conversation publicly says something about how far we've come — and how far we still have to go. Let's start with "should you read the code?" On its surface, this sounds like a question with an obvious answer — yes, obviously, read the damn code. But the reality is more uncomfortable. AI-assisted development has normalized a workflow where code is generated, reviewed at a high level, and merged with a confidence that would have been unthinkable a decade ago. The deeper question isn't whether you *can* read every line, but whether your mental model of the system survives when you stop. And for open source maintainers, that's existential: if your contributors can't explain what their PRs do, you don't have a community, you have a cargo cult. Then there's the RAG debate. RAG was supposed to be the pragmatic middle ground between fine-tuning your own model and just vibing with a general-purpose LLM. But as the conversation signals, the architecture is getting squeezed from both sides — long-context windows are making naive retrieval look increasingly redundant, while tool-use and agentic patterns are absorbing the "go fetch information" use case that RAG was built for. That doesn't mean RAG is dead, but it does mean "RAG" as a buzzword is losing its explanatory power. What remains is the underlying problem: how do you ground a statistical parrot in your actual codebase without it hallucinating your API contracts into oblivion? And finally, the Skills vs. MCP question is where things get genuinely spicy. The Model Context Protocol was supposed to be the USB-C of AI tooling — one standard connector to rule them all. But standards only win if they're simple enough to adopt, and MCP's spec, while elegant, demands a certain amount of ceremony. Skills, by contrast, are the "just put the instructions in a folder" school of thought — and in open source, the simpler option usually wins the adoption war. That doesn't mean MCP is dead either; it just means the protocol layer is becoming a commodity while the actual value moves up the stack into how skills are authored, versioned, and shared. The real takeaway from this episode is that the AI tooling landscape is in a Cambrian explosion, and we're still arguing over taxonomy while the ecosystem eats itself. Whether you're a maintainer, a contributor, or just someone trying to ship software without losing your mind, these are the questions worth chewing on. Read the code. Question the architecture. And for the love of all that is holy, keep an eye on your dependencies.