Skill Recorder turns one screen session into a reusable agent skill

The gap Skill Recorder fills between doing a task and teaching it

Most agent tooling assumes you can already describe the procedure. Skill Recorder starts from the opposite end: you perform the task once, on screen, and the tool derives the description. The README frames the promise as recording yourself submitting one form and teaching the agent to submit all of them, which is a generalization claim rather than a macro replay.

The audience is narrow and identifiable. You need a GitHub account with Copilot access, because the Copilot CLI ships with the app and does the reconstruction. You need to be comfortable running a shell command that builds a release commit locally. And you need a task that is genuinely repetitive, because a one-off procedure is not worth the recording and review cycle. The output targets three consumers named in the repository description: Microsoft Scout, Microsoft Copilot Cowork, and Copilot Studio.

Record, Analyze, Create: where the data actually goes

The pipeline has four stages and one sharp boundary. Recording, storage, frame extraction, and optional narration transcription all happen on your computer, and the README states that nothing leaves while you record. The boundary is the Analyze button: that is when the event timeline, extracted screen images, and narration text go to GitHub's cloud for Copilot to process.

What lands in that timeline is specific. Window tracking records active-app and window switches. Browser URLs are captured on macOS. Screen video is recorded by Chromium, but the app keeps low-rate snapshots only when the screen changes or a heartbeat is due, which is a deliberate size reduction rather than a continuous frame dump. Clipboard capture stores short previews of copied text that tie steps together. Narration is optional and transcribed on-device, with a one-time download of roughly 252 MB for a Whisper model covering 99 languages.

The reconstruction produces one overall intent plus an ordered list of steps, which you review and edit before generating anything. That human checkpoint matters, because the generalization from a single example is the part most likely to be wrong. The repository also carries a sensitive-content scanner, frame redaction, and OCR modules in the source tree, and the README still tells you plainly not to record, type, paste, or narrate passwords, tokens, or API keys. Treat the scanner as a backstop, not a guarantee.

Installing Skill Recorder from a source release

There is no downloadable binary. Skill Recorder is published as a source release: one command downloads a pinned Node.js runtime, builds the exact release commit on your machine, and registers a Skill Recorder (Source) app you can relaunch later. Nothing is installed globally. Open the latest release page and copy the command for your platform, because each release pins a specific 40-character commit that you substitute into the pattern below.

On macOS or Ubuntu, the install script is fetched at the pinned commit and the same commit is passed through the environment so the built source matches the script:

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

If you want the app to survive closing the terminal, add the detached flag after the pipe:

commit="<40-character-release-commit>"; curl -fsSL "https://raw.githubusercontent.com/microsoft/skill-recorder/$commit/install.sh" | SKILL_RECORDER_COMMIT="$commit" SKILL_RECORDER_DETACHED=1 bash

On macOS this places a Skill Recorder (Source) app in ~/Applications. On Windows 11, the PowerShell equivalent sets the commit in the environment and pipes the script into iex, which adds desktop and Start Menu shortcuts:

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

The README points to INSTALL.md for inspecting the script before running it, setting install options, updating, and uninstalling. That file is the right place to look, because the README itself does not document rollback.

Your first recording, and the permission you have to grant first

On first launch, macOS asks for Screen Recording permission. Grant it, then use the record control or the global shortcut, which is Command-Shift-R on macOS and Ctrl+Shift-R on Windows. The shortcut toggles recording from anywhere, so you do not need the window in front.

While recording, an always-on-top bar shows capture and microphone state. You can mute, unmute, or switch microphones mid-session, finish the take, or discard it with a confirmation if it went wrong. Then click Analyze. The first time you do this, the app offers Sign in to Copilot if you are not signed in yet. Review the reconstructed intent and steps, edit them until they read correctly, and only then generate the Skill or the Automation.

If you would rather build from source than use the installer, the README requires Node.js 24 and gives this sequence after checking out a release revision:

npm ci
npm run compliance:licenses
npm run dev

The dev script starts Vite and launches the Electron app with hot reload. Package.json pins the engine tighter than the README does: node >=24.19.0 <25 and npm >=11.17.0.

Where Skill Recorder is the wrong tool

The source-release model is the first constraint. Building a release commit on your machine is fine for an engineer and awkward for anyone who expects a signed installer, an auto-update channel, or an IT department that can approve a package. Package.json also marks the project private, which is consistent with a build-locally distribution model rather than a published npm artifact.

The second constraint is the cloud boundary. If your task touches regulated data, internal dashboards behind authentication, or anything you would not paste into a third-party service, Analyze is the point where that material leaves your machine. The README's own warning is unambiguous about keeping secrets out of recordings, and a warning is not a technical control.

The third is platform. macOS is described as the primary target and Windows 11 as supported, with a separate WINDOWS-VALIDATION.md in the repository. Ubuntu appears only in the install command pairing, so treating Linux as a first-class desktop target is not supported by what the README says. And if your task varies too much between runs, generalizing from a single recorded example will produce steps that look plausible and fail on the second execution.

How this differs from scripted automation and from writing skills by hand

The obvious alternative is a conventional UI automation tool that records clicks and replays them. The difference is in the output. Skill Recorder does not emit a replay script; it emits a procedure that prefers the agent's native tools, naming the gh CLI and web_fetch as examples, and generalizes from your single example. A click-replay macro breaks the moment a button moves. A skill that says to fetch a page or call a CLI tool survives layout changes, at the cost of depending on the model to interpret the steps correctly.

The second alternative is writing the SKILL.md yourself. That gives you exact control and no recording step, no screen capture permission, and no data leaving your machine. What you lose is the derivation: the ordered steps come from watching what you actually did, including the parts you would forget to write down. For a procedure you already understand well enough to document, handwriting is faster. Skill Recorder earns its place when the task is easier to perform than to describe.

Licence, maintenance, and what an upgrade actually costs

The project is MIT licensed, and package.json confirms the same identifier. The README points to THIRD-PARTY-NOTICES.md for the licences of bundled dependencies, and the build scripts include npm run compliance:licenses plus a check that asserts a catalogue bundle boundary, which suggests the maintainers treat the dependency surface as something to police. That is not legal advice; if you redistribute anything built from this source, read THIRD-PARTY-NOTICES.md and INSTALL.md, which the README says covers the licensing boundary between local source builds and redistributable packages.

On maintenance: the last push to main was on 2026-09-09, and the most recent release, v0.5.0, was published on 2026-08-12, following v0.4.2 and v0.4.1 earlier in August. The repository is not archived.

Upgrades are not a background download. Because each release pins an exact commit and the installer builds that commit, moving to a new version means running the install command again with the new commit hash. INSTALL.md is the documented place for update and uninstall steps. Budget for that as a deliberate action rather than something that happens to you.

Editorial conclusion

Adopt Skill Recorder if you already have GitHub Copilot access, work primarily on macOS, and have a repetitive task you can demonstrate once on screen without touching credentials. Skip it if you need a signed installer, run Linux as your daily desktop, or cannot send screen images and clipboard previews to GitHub's cloud. Before committing, run npm run eval and npm run eval:builder against your own fixture and confirm the generated steps prefer native tools such as gh or web_fetch instead of replaying clicks.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. microsoft/skill-recorder on GitHub
4. README
5. Releases