Microsoft Skill Recorder turns recorded work sessions into agent skills

Writing an agent skill from scratch asks you to remember every click you make.

Microsoft open-sourced Skill Recorder on GitHub. It is now trending at rank 10 on Trendshift's daily list on 25 October. The repository has 4,685 stars under the MIT license. It was created on 29 July 2026. The project is a desktop app that records your work session. GitHub Copilot then reads the recording. It writes back an overall intent and an ordered list of steps.

What it records

The app captures clicks and app switches. It captures window switches and visited pages. An optional recorded terminal captures commands and their output. If you turn it on, it captures spoken narration too. A control bar during recording holds the microphone and the readiness checks. All capture is local. The hotkey is Command-Shift-R on macOS and Control-Shift-R on Windows.

What it builds afterwards

After you stop recording, you press Analyze. Microsoft states that the GitHub Copilot CLI reads the capture. It returns one overall intent and an ordered list of steps. You can edit the output until it reads right. From there, one step turns the run into something an agent can reuse.

Two kinds of output are possible. The first is a SKILL.md file. An agent reads this and runs the procedure on demand. The second is an Automation. It runs the same procedure on a schedule or on a trigger. Microsoft states that both prefer the agent's native tools over replaying UI clicks. The README example is a form. Record yourself submitting one form, and the output can teach the agent to submit many.

What it needs

The app runs on macOS. It also runs on Windows 11 x64 and ARM64. It needs a GitHub account with Copilot access. Microsoft lists three target agents: Microsoft Scout, Microsoft Copilot Cowork and Copilot Studio. The project pins Copilot CLI version 1.0.78. Microsoft states this version supports web OAuth without a terminal device code.

Where it fits

A familiar failure in agent work is a skill document that drifts away from the real task. The person wrote it from memory. Nobody updates it when the UI changes. A recording is harder to get wrong. The output still has to be reviewed and the agent still has to be tested. That is where the real work sits. But the first pass is cheaper. The second pass is a diff against a real capture.

One caveat is worth stating in plain words. Microsoft calls the Review SKILL.md step a visibility tool. The company states that it is not a security guarantee. A skill written from a screen capture can still contain a mistake. Running it against real data on the first try is a risk a careful team avoids.

Source

microsoft/skill-recorder on GitHub.