Microsoft's skill-recorder turns screen demos into agent skills

Microsoft's skill-recorder records a screen session, then rebuilds it as an intent plus ordered steps for an AI agent, per the project's GitHub repo.

That matters for anyone building on Copilot Cowork or Copilot Studio. Instead of hand-coding a skill definition, you record yourself doing the task once, and the GitHub Copilot CLI writes the reusable version.

The app is Electron-based and captures clicks, window switches and optional narration during a session. It hands the recording to the GitHub Copilot CLI, which reconstructs it as a named intent with ordered steps rather than a literal replay script.

The project's stated design goal: generated Skills "prefer the agent's native tools over replaying UI clicks, and generalize from your one example." That's the pitch against traditional RPA, where a macro breaks the moment a button moves. Here the agent is supposed to reason about what you were doing, not where you clicked.

Skill-recorder targets Microsoft Scout, Copilot Cowork and Copilot Studio. It needs a Copilot-enabled GitHub account, Node.js 24, and macOS or Windows 11 to run, the repo says.

The repo has picked up 2,233 stars since its July 29 launch, pushing commits daily through August 7.

The riskier bet is in that phrase "generalize from your one example." One demonstration usually leaves out what's supposed to vary between runs and what's fixed. Until skill-recorder shows it can tell the difference, expect generated Skills to need hand-editing before anyone trusts them in a Cowork or Studio flow.