Hmm... I didn't expect the skill to kick-off its own software development rabbit-hole but that's what seems to be going on. I'll see what it can code-up for itself to give me int insights. Funny how it keeps editing the code it wrote. So I guess everyone will get different outcomes.
Oh, got it! You meant that "skill" is generating code to accomplish its task. Yes, that's what it's doing. I didn't even prompt for that; I just defined the steps, and it decided to do it. Later, when I was building, I asked Claude if that's the default behavior of Claude Code skills, and it said yes. Claude basically decides that if it's a big data processing job, it creates a temporary script and does that. Maybe it doesn't do this for all skills because, in this case, I was processing big files. Instead of loading the entire text into the context and extracting it, it writes the logic and does the job. I don't think it's specific to skills, to be honest, but I didn't explicitly instruct it. I was thinking about it, but then I realized it figures it out on its own.
Regarding the whole skills thing, I think they're currently suggesting building skills for this kind of task and using MCPs mainly for pulling data from external sources. The main idea is not to overload the context with MCP tool lists and their signatures. Basically, if you register 5 MCP servers, and if each has about 5 tools on average, you end up loading 25 function signatures into the context. That's why they recommend skills. But honestly, I haven't checked how they optimize skills loading. They might be looking at the YML front matter, only loading it to the context and deciding to use that skill if the front matter defines the intention and that intention matches the user's. Actually, that's something on my list to check it out and see how it discovers skills and what I need to consider in my skill definitions to ensure they are recognized. But I believe, at worst, you just write: "Use this skill," and it loads it by the file name.
Hmm... I didn't expect the skill to kick-off its own software development rabbit-hole but that's what seems to be going on. I'll see what it can code-up for itself to give me int insights. Funny how it keeps editing the code it wrote. So I guess everyone will get different outcomes.
Oh, got it! You meant that "skill" is generating code to accomplish its task. Yes, that's what it's doing. I didn't even prompt for that; I just defined the steps, and it decided to do it. Later, when I was building, I asked Claude if that's the default behavior of Claude Code skills, and it said yes. Claude basically decides that if it's a big data processing job, it creates a temporary script and does that. Maybe it doesn't do this for all skills because, in this case, I was processing big files. Instead of loading the entire text into the context and extracting it, it writes the logic and does the job. I don't think it's specific to skills, to be honest, but I didn't explicitly instruct it. I was thinking about it, but then I realized it figures it out on its own.
Regarding the whole skills thing, I think they're currently suggesting building skills for this kind of task and using MCPs mainly for pulling data from external sources. The main idea is not to overload the context with MCP tool lists and their signatures. Basically, if you register 5 MCP servers, and if each has about 5 tools on average, you end up loading 25 function signatures into the context. That's why they recommend skills. But honestly, I haven't checked how they optimize skills loading. They might be looking at the YML front matter, only loading it to the context and deciding to use that skill if the front matter defines the intention and that intention matches the user's. Actually, that's something on my list to check it out and see how it discovers skills and what I need to consider in my skill definitions to ensure they are recognized. But I believe, at worst, you just write: "Use this skill," and it loads it by the file name.
Read this article if you want to learn more about the recently announced Agent Skills for Claude. https://www.anthropic.com/engineering/equipping-agents-for-the-real-world-with-agent-skills