Building Agent Skills with Claude research mode

You can use Claude Desktop to generate Skills packages according to the spec. Here’s how it works:

• If you have a docs MCP relevant to the skill, enable that in Claude Desktop
• Make sure Code Execution and File Creation are enabled in Settings > Capabilities
• Feed the Agent Skills spec to Claude Desktop
• Trigger a Research run to populate the context with whatever information you think will help
• When research is complete, tell Claude Desktop to generate a zip with a skills package per the spec

It’ll spit one out.

From here you can drop it into Claude Code and enhance it further using additional context:

• project files
• example code/projects downloaded from GitHub
• existing documentation, research and other artifacts you have locally

Context is the difference between useful inference and waste.

Skills create tidy context packaging that agents can select as needed to improve correctness and efficiency. This is supremely valuable if you’re working with newer technology that’s just not represented in the model, either because it’s new or because it’s obscure.

Skills are designed for progressive disclosure: they’re a very simple artifact that’s structured such that the agent can strategically load just what’s needed, when it’s needed.

You can accomplish similar effects by jamming raw markdown into the context, but Skills are comparative cruise control.

The trinity of agents, MCPs and structured context like skills represent the foundation for accomplishing anything useful with LLMs, whether you’re a single developer or you’re building something for developers to use at scale.

This is just enough fencing to get both flexibility and predictable work.