I’d go even further and add that these distinct tools don’t belong in the same toolbox.

Im not OP but as someone who did one of these recently, I think about it kind of like sharing an essay. I had an idea, I clauded a working thing, used and refined until it was useful for my team then wanted to share the kernel that developed. OP is not claiming it’s like a panacea or a masterpiece and idk I appreciate the thoughts it’s making me think.

My other more general thought here is that the nature of LLMs is so generic and context specific that just learning about a workflow can unlock a ton of potential. This is evident in how the null state of every chat UI has a bunch of examples to teach you to use it. MCP’s are a portable way to share a workflow. To get concrete: I was pulling a lot of business data sources into my Claude code sessions to great effect but it was a force multiplier for my team to figure out an MCP version that would tell their Claude’s how to do the same. None of it was deeply technical engineering but it was immediately useful and most of the team using it are not engineers.