Record yourself doing a task once, then turn it into a skill your AI agent can repeat.
Tip
Want to install Skill Recorder? Go to the latest release. Scroll to Commit-pinned installation and copy the complete command for your operating system. It is ready to paste: you do not need to edit it or download the source code yourself. Step-by-step instructions below.
Skill Recorder captures a real work session on your screen: the clicks, the app and window switches, the pages you visit, commands/output from its optional recorded terminal, and (if you want) your spoken narration. It then uses the GitHub Copilot CLI to reconstruct what you actually did as a clear intent plus an ordered list of steps. From there, one step turns that single run into something an agent can reuse:
- a Skill: a
SKILL.mdprocedure an agent runs on demand, or - an Automation: the same procedure on a schedule or trigger.
Both prefer the agent's native tools (like the gh CLI or web_fetch) over replaying
UI clicks, and generalize from your one example, so recording yourself submitting one
form can teach the agent to submit all of them.
- 🔴 Record. Hit record (or
⌘⇧R/Ctrl+Shift+Rfrom anywhere) and just do your task. Skill Recorder captures your screen and activity locally, in the background. - 🎛️ Control. While recording, a small always-on-top bar shows capture, microphone, and recorded-terminal state. Open the session-scoped terminal, mute, unmute, or switch mics, then finish or discard with confirmation.
- 🧠 Analyze. Click Analyze and GitHub Copilot reconstructs one overall intent and an ordered list of steps. Review and edit until it reads right.
- ✨ Create. From an approved analysis, generate a reusable Skill and/or a scheduled Automation.
For skills, Review SKILL.md optionally shows the complete, read-only file text (including frontmatter) before installation or export. Preparing a preview writes no skill files. Close and reopen it without regenerating; Add/Export from either surface uses that exact preview. Editing the plan discards the preview. You can still choose Add/Export directly without reviewing or adding a confirmation step. Review provides visibility into the generated instructions, not a security guarantee.
Instead of recording your screen, use the smaller folder icon beside Record to choose a local MP4 or WebM file, such as a downloaded Teams demonstration. Chromium checks the actual video and audio before the details screen opens. Unsupported codecs, encrypted or unsupported track layouts, corrupt media, and files outside the import limits are rejected with an explanation. No converter, standalone FFmpeg installation, or additional codec download is required.
The initial limits are 512 MiB, 15 minutes, and 3840 x 2160 pixels (or an equivalent pixel count). Audio must be mono or stereo at no more than 48 kHz. Import supports one video track and at most one audio track. MP4 track types recognized are AVC, HEVC, VP9, or AV1 video with AAC/Opus audio; WebM recognizes VP8, VP9, or AV1 with Opus/Vorbis. Recognized track metadata is not a codec-support guarantee: the shipped Chromium runtime must actually decode it.
Import memory use is higher than the compressed file size. A Windows ARM64 stress check with a 512 MiB input, 15-minute stereo audio, and the existing Whisper model resident peaked at approximately 4.6 GB (4.25 GiB) across Electron processes. This is a measurement, not a guarantee for every file or platform, and excludes other apps and the external Copilot CLI process. The import decoder has a 2 GiB working-set guard; recordings that exceed it are rejected even when their file size is within the limit.
After validation, you can enter Additional details, or leave the field empty. Explain the task, supplement what is visible, or indicate where the demo happens. Timestamps are guidance, not trimming: other parts of the recording may still be processed. Choose the recording's spoken language when it contains audio. Analyze transcribes audio locally using the existing Whisper model (with the first-use download disclosed before proceeding); transcription failures stop the import's analysis so you can retry. Videos without audio and recordings with no detected speech are supported.
The resulting activity list uses the normal review flow: edit details, reorder, add or remove activities, then create and review your skill as usual. Imported analysis uses frames, speech, and your details, not captured OS events; it cannot recover hidden clipboard contents or commands that were never visible or spoken.
The original file is untouched. A local copy and derived artifacts are stored in Sessions, and deleting the session removes only that copy. Selecting a file does not send it to the cloud. Choosing Analyze sends selected frames, transcribed speech, and your details through the existing privacy-protection pipeline to GitHub Copilot. Debug exports include the imported media and details, just as they include private data from live recordings. Only import recordings you are authorized to use, and keep sensitive/confidential content out of them.
Live recording, its permissions and settings, and its event-first analysis are unchanged. Imports do not require microphone or screen-capture permission.
For development, npm test includes the import contract/privacy tests.
npm run test:import:runtime uses the reviewed Electron runtime to generate
synthetic MP4/WebM files with Chromium and exercise decoding, seeking, silent and
absent audio, unsupported codecs, cancellation, and source-file preservation.
Run that check on each release target; it does not replace checking representative
Teams files. Add -- --large to also import an exact 512 MiB synthetic MP4 and
check rejection of a file one byte over the limit; this check needs additional
memory and temporary disk space. The initial local measurements and runtime check were performed on
Windows ARM64; other platforms and real meeting recordings require their own
verification.
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 adds a Skill Recorder (Source) app you can relaunch anytime. Nothing is installed globally. You'll need a GitHub account with Copilot access; the Copilot CLI ships with the app.
macOS is the primary target. Windows 11 (x64 and ARM64) is supported too (see
WINDOWS-VALIDATION.md).
Open the latest release and follow these steps:
- On the release page, scroll to Commit-pinned installation. Find Windows 11 x64 or ARM64 or macOS or Ubuntu, then copy the entire command in that section using the copy button at the top-right of the code block.
- Open PowerShell from the Windows Start menu, or Terminal on macOS or Ubuntu. Paste the command and press Enter.
- Keep that window open while installation completes. The installer downloads what it needs, builds the app on your computer, creates shortcuts, and launches Skill Recorder. The first installation can take several minutes.
Important
Copy the command from the release page, not a template from the technical docs.
The release command already contains the correct version identifier; there is
nothing to replace. Do not paste commands containing
<40-character-release-commit> literally.
You do not need Git, a separate Node.js installation, or the Manual developer
setup instructions. The Source code (zip) and Source code (tar.gz) links
under Assets are not app installers.
To review the script before running it, follow the release page's inspect-first instructions instead. Installation remains subject to your organization's policies.
Open it again later: use Skill Recorder (Source) on your Windows desktop or
Start Menu; on macOS, find it in ~/Applications, Spotlight, or Launchpad; on
Ubuntu, use the matching application entry. You do not need to reinstall each time.
Advanced options, including running in the background on macOS or Ubuntu, are in
INSTALL.md.
- Grant Screen Recording. On first launch, macOS asks for Screen Recording permission; grant it and you're ready to record.
- Record, Analyze, Create. Do your task, then Analyze. The first time you Analyze, Skill Recorder offers Sign in to Copilot if you aren't signed in yet. Finish authorization in your browser; the app verifies sign-in and automatically retries the interrupted analysis once. There is no login terminal window to close. Canceling sign-in or leaving that analysis prevents the automatic retry.
On Windows, a local Microsoft Entra tenant hint can offer a choice between
Microsoft Enterprise SSO and a Personal GitHub account. This is only a
routing hint, not proof of employment. Microsoft sign-in has two separate steps:
finish SSO in the browser, then authorize the GitHub Copilot app. When SSO lands on
github.com/enterprises/microsoft, close that tab, return to Skill Recorder, and
choose Authorize Copilot. SSO alone does not authorize the app. Use the same
browser profile and Microsoft-linked GitHub account for both steps.
Without a hint, and on macOS/Linux,
sign-in goes directly to the standard GitHub flow. No work-email prompt is required.
The authorization step opens the default browser using the CLI's web OAuth flow, not a device code. Keeping the same browser profile lets GitHub reuse the session from SSO, so you can approve the GitHub Copilot app without entering a code. If GitHub asks you to sign in again, check that you are using the same browser profile. After authorization, close that tab and return to Skill Recorder; the app verifies access before continuing.
Skill Recorder pins Copilot CLI 1.0.78, which supports web OAuth without a terminal,
and runs --no-auto-update login --web-flow so sign-in does not depend on a
different CLI version in the user's auto-update cache. Older bundled versions
such as 1.0.71 do not support --web-flow.
Sign-in can be canceled and times out after five minutes. If browser login fails, the app shows an error and a manual command for the bundled CLI (use PowerShell on Windows). After manual sign-in, retry the analysis yourself. Signing in from a skill or automation builder does not automatically repeat a build or installation.
To inspect the script before running it, set install options, update, or uninstall, see
INSTALL.md.
⚠️ Keep secrets out of your recordings. Don't record, type, paste, or narrate passwords, tokens, API keys, or other confidential info. Choosing Analyze sends recording data to GitHub's cloud. Skill Recorder reminds you before every recording. Details in What gets captured.
Everything below is for people who want the details, or want to hack on the code.
Recording, storage, frame extraction, and optional narration transcription all happen on your computer; nothing leaves while you record. Only when you choose Analyze does Skill Recorder send the event timeline (window/document titles, URLs, clipboard previews, and terminal commands), extracted screen images, narration text, and bounded terminal-output excerpts requested during analysis to GitHub's cloud for Copilot to process. When protection is enabled, terminal commands and the complete local terminal transcript participate in the same on-device secret/PII scan before any requested excerpt is returned.
The in-app "Records your screen and activity" panel spells out exactly what's collected:
- Window tracking: active-app / window switches.
- Browser URLs: the page you're on (macOS).
- Screen video: recorded by Chromium; low-rate snapshots are kept only when the screen changes or a heartbeat is due.
- Clipboard: short previews of copied text that tie steps together.
- Narration (optional): spoken commentary, transcribed on-device in any of Whisper's 99 supported languages (a one-time ~252 MB model download on first use).
- Recorded terminal (only when opened from the floating recording bar): command, cwd, shell, exit status, duration, and the complete terminal output. The transcript has no app-imposed size cap and counts toward the saved session's storage. It captures only the app-owned terminal for that recording—there are no global hooks, profile edits, background command monitors, or observation of existing terminals.
⚠️ Please don't capture secrets. Passwords, access tokens, API keys, credentials, and other confidential information should never be recorded, typed, pasted, shown, copied, or narrated during a session.
For developers changing the code, not for installing the app. To use Skill Recorder, follow Install it instead.
Requires Node.js 24. After checking out a release revision:
npm ci
npm run compliance:licenses
npm run devnpm run dev starts Vite and launches the Electron app with hot-reload; ⌘⇧R (macOS) /
Ctrl+Shift+R (Windows) toggles recording from anywhere. Full manual setup, the build and
dist scripts, and the licensing boundary between local source builds and redistributable
packages are in INSTALL.md. Maintainers changing versions, dependencies,
assets, or releases must follow RELEASING.md.
The Copilot describer and builders have a fixture-based eval suite; see
evals/README.md.
npm run eval # score the describer against synthetic recordings
npm run eval:terminal # score indexed terminal-output use and redaction
npm run eval:builder # score the skill/automation generalization- INSTALL.md: install options, inspect-first install, updating, uninstalling, and manual developer setup.
- RELEASING.md: maintainer release runbook.
- evals/README.md: the describer / builder eval harness.
- THIRD-PARTY-NOTICES.md: licenses for bundled dependencies.
Please don't report security vulnerabilities through public GitHub issues. See
SECURITY.md for Microsoft's coordinated-disclosure process and reporting
channels.
File bugs and feature requests through
GitHub Issues (search existing issues
first to avoid duplicates). Support is limited to the resources described in
SUPPORT.md.
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit Contributor License Agreements.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.