Skill Recorder: record a gesture, get a skill, the Coeurdar verdict
A Microsoft desktop application (MIT, TypeScript) that captures a working session on screen: clicks, window switches, pages visited, plus optional spoken narration while you work. The GitHub Copilot CLI then replays the recording, reconstructs the intent behind the manipulation and returns it as an ordered sequence of steps. What you get out is a SKILL.md file, the procedure an agent runs on demand. The order matters: between analysis and generation, the app shows you the intent it understood and the list of steps, and that is where you correct it. Fixing a six-line plan costs far less than fixing a SKILL.md written around a misreading. Narration is transcribed on your machine (with Whisper), and the recording only leaves for GitHub once you start the analysis. The same approved plan can also become a scheduled automation instead of a skill. Runs on macOS, Ubuntu and Windows 11.
Strengths
- Starts from the real gesture rather than the memory of it, so tacit steps land in the text
- Narrating out loud during capture carries the why, not just the click sequence
- Outputs a standard SKILL.md, readable and editable like any file in your repo
Limitations
- Generalises from a single run, so the procedure ignores the cases your session never hit
- Depends on the GitHub Copilot CLI for the replay, hence on the account and quota behind it
- Recent and fast-moving project: the shape of the generated file can shift between versions
Best for
- Turning a repetitive internal-tool manipulation into a procedure your agent replays
- Documenting a business gesture held by one person before they move teams
- Getting a first draft of a skill out when the blank page has blocked you for weeks