Microsoft's Skill Recorder: record yourself once and hand the task to an agent, with GitHub's cloud as the fine print

By OneDigital · October 24, 2026.

Microsoft has released Skill Recorder, an open-source tool (MIT license) that records a real on-screen work session and turns it into a procedure an AI agent can repeat. According to the project description, it captures clicks, window switches, pages visited, activity in an optional terminal and the user's spoken narration.

GitHub Copilot handles the analysis. It reconstructs the overall intent and an ordered list of steps that the user can review and edit. Once the user approves, it generates either a "skill" (a SKILL.md file with the procedure the agent runs when asked) or an automation triggered by a schedule or an event.

The most interesting technical detail is what the tool does not do: it does not replay clicks like a macro. According to the project, it prefers the agent to use its native tools, such as GitHub's gh command-line tool or web queries, and it generalizes from a single example. Recording yourself submitting one form should teach it to submit every form of that type.

On privacy, the tool splits the process into two phases. During recording, everything stays on the device: low-frequency screenshots, window tracking and local transcription with Whisper (99 languages, a one-time download of about 252 MB). Only when the user clicks Analyze are the event timeline, screen images, narrated text and limited terminal excerpts sent to GitHub's cloud. The project warns users not to record or narrate passwords, tokens or API keys, and says it runs a local scan for secrets and personal data when protection is turned on.

The project lists several limitations. macOS is the main platform (Windows 11 is supported on x64 and ARM64), and a GitHub account with Copilot access is required. Reviewing the SKILL.md file provides transparency but, by the project's own account, does not guarantee security. The first installation compiles the code locally and can take several minutes.

Our take: the recorder itself matters less than the paradigm shift. Until now, automation required writing code or prompts; here, the unit of work is a human example. That greatly lowers the barrier to delegating repetitive tasks and fits the view we hold on employment: AI doesn't so much "take" your job as turn every routine task you demonstrate once into a candidate for delegation. In general, administrative and back-office work, precisely the kind described in repeatable steps, is the most exposed.

There is also a strategic angle. Skill Recorder feeds the Copilot and GitHub ecosystem: whoever captures how you work controls the format of "skills" and, with it, much of the plumbing behind agents. The open license helps, but relying on Copilot and GitHub's cloud for analysis limits users' real control over their data. For now, workflows involving sensitive information are off the table.

The short-term risk is twofold. A skill generated from a single example may fail on edge cases that nobody recorded, and a poorly generalized procedure run autonomously may take unintended actions. Human review of the step-by-step plan is not a formality: it is the safeguard that separates a useful assistant from an opaque automation. As agents become more autonomous, what will matter most is being able to audit what each skill actually did, not just what the file said.

In the long run, the direction is encouraging: if showing how a task is done is enough to delegate it, tacit knowledge no longer has to be tied up in hours of repetitive work. It is one small but real step toward an economy where people spend their time on work that calls for judgment and passion. The transition, however, will not be painless for those who make a living from those tasks today.