Building agent skills by demonstration instead of hand-writing them: a record-and-compile approach

I was automating a repetitive desktop task and got tired of hand-writing agent skill files, so I went looking for a better approach and found a tool that takes an interesting angle - sharing it here for discussion. Instead of scripting the automation step by step, you demonstrate the task once on screen. It records the session, then an LLM compiles the event log plus screen context into a structured SKILL.json and a readable SKILL.md the agent can replay later. It runs as an MCP server exposing record, stop, compile, and list, so it plugs into any MCP client. What I found worth discussing is the "show, don't script" idea: reading native UI events plus a screen recording makes the resulting skill far less brittle than a rigid macro, and it lowers the barrier to turning everyday workflows into reusable agent capabilities. Open source, and I'm not affiliated with it - just a useful find from a fellow builder. I'll drop the link in the comments. Do you think demonstration-based skill creation is a viable path for agents, or does it break down on complex, branching tasks?