I tried the same thing in @figma AI Agent, and it actually gave me an editable path too.
I think the difference is partly in how you prompt it. If you don't specify what the output should be and what it shouldn't be, the agent has to guess. AI is a bit like a naive intern. It’ll try to get the job done, but if you don’t explain what you actually want, it’ll make its own assumptions
Claude might also lean towards code because of how it works with Figma MCP, while Figma's agent handles a much wider range of creative tasks.
Tell it what the final output should look like, what to avoid, and what constraints matter.
Hope this helps someone get better results from their AI tools!
