What is Microsoft's Skill Recorder?

Posted: (EET/GMT+2)

 

Microsoft's Skill Recorder is a new desktop tool for turning a task you perform once into a reusable AI agent skill. Microsoft made the GitHub repo available on July 24th.

The idea of this tool is very practical: record yourself doing a work task, let the tool analyze what happened, then turn the result into a reusable procedure.

Skill Recorder captures things such as screen activity, app and window switches, browser pages, clicks, and optional spoken narration. After recording, GitHub Copilot CLI is used to reconstruct the session into an intent and an ordered list of steps. From there, Skill Recorder can generate:

  • a SKILL.md file that an AI agent can run on demand
  • an automation that runs the same procedure on a schedule or trigger.

This is interesting because it moves agent instructions closer to real work. So instead of writing a long prompt from memory, you can demonstrate the task once. The tool can then help convert that demonstration into something reusable.

Microsoft's examples mention targets such as Microsoft Scout, Microsoft Copilot Cowork, and Copilot Studio. The basic workflow is like this:

  • record yourself doing the task
  • optionally narrate what you are doing
  • analyze the recording
  • review and edit the generated steps
  • create a reusable skill or automation.

Installation is done from the GitHub release instructions. The project is published as a source release, and the install command builds the pinned release commit locally. On Windows, the install command is PowerShell-based:

$commit="<40-character-release-commit>";
$env:SKILL_RECORDER_COMMIT=$commit;
irm "https://raw.githubusercontent.com/microsoft/skill-recorder/$commit/install.ps1" | iex

On macOS or Ubuntu, the install command is shell-based:

commit="<40-character-release-commit>";
curl -fsSL "https://raw.githubusercontent.com/microsoft/skill-recorder/$commit/install.sh" |
    SKILL_RECORDER_COMMIT="$commit" bash

Follow the current command from the latest GitHub release instead of copying an old command from a blog post. The project currently targets macOS primarily, but Windows 11 is supported too. There is one very important warning: do not record secrets.

A recording may include screen content, window titles, URLs, clipboard previews, and optional narration. According to the project documentation, recording happens locally first, but choosing Analyze sends recording data to GitHub's cloud for Copilot to process. Do not record yourself typing, pasting, copying, showing, or narrating things like:

  • passwords
  • API keys
  • access tokens
  • connection strings
  • customer data
  • other confidential information.

For developers, the most interesting use case is probably documenting and automating boring repeatable work:

  • opening a tool and exporting a report
  • following a release checklist
  • creating an issue or pull request from a standard pattern
  • checking a dashboard and summarizing the result
  • performing a support or operations task.

Still, the generated skill should be reviewed. A recording is a good starting point, not a replacement for engineering judgment. Good questions to ask before creating a skill:

  • is this task safe for an agent to repeat?
  • does the task need human approval before making changes?
  • can the task use APIs or command-line tools instead of UI clicks?
  • does the recording contain secrets or private data?
  • are the generated steps general enough for future runs?

Skill Recorder is a good example of where AI tooling seems to be going. Not only asking an agent to do something, but teaching the agent a reusable way of working.