Teach Recorder

What it is

A built-in, off-by-default feature for recording what you do on Windows — clicks, keystrokes, and screen state — and turning that recording into a reusable AI skill that agents can replay on your behalf.

What it does

Record a demonstration of a repetitive task. When you stop, Omniscio captures the ordered sequence of actions (with app/window context and element names). Tap Generate skill to send the recording to the AI, which produces a parameterized GeneratedSkill — a title, summary, named parameters, and numbered replay steps — that an agent can run.

Where to find it

Enable it in Settings → Lab, then open the Teach Recorder entry in the Automation group of the sidebar.

How it behaves

How to use it

1. Enable via Settings → Lab → Teach-by-Demonstration Recorder (or set AMC_SHOW_TEACH_RECORDER=1).
2. Open the Teach Recorder entry in the sidebar (Automation group).
3. Click Start recording — a small always-on-top indicator pill appears over your other apps. Perform your task, then Pause/Resume or Stop right from the pill (or Stop from the panel).
4. Select the recording from the list and click Generate skill.
5. Review the generated skill (title, params, steps) — the skill is stored with the recording and can be referenced or exported.

Notes

• Always-visible indicator with pause/stop: while recording, a floating pill (a recording/paused dot + elapsed timer + Pause/Resume + Stop) stays on top of whatever app you're in, so you never forget a recording is running and can control it without returning to Omniscio. Pausing fully stops all capture — clicks, keystrokes, and screenshots — until you resume, a real privacy pause. The pill is hidden from screen capture, so it never appears in your recording. (Global pause/stop keyboard shortcuts are a planned follow-up.)
• Desktop-only (Windows): uses WinHook and UI Automation for element capture; not available on mobile or macOS.
• Generate sends screenshots to the cloud AI: each step with a captured frame is included in the vision call. A clear disclosure is shown before generation starts.
• Click screenshots are marked at the click point: before a click frame is sent, a high-contrast marker (ring + crosshair, plus the element's box when Windows named it) is drawn on a copy at the exact click location, and the click coordinates + element bounds are added to the text the AI reads — so it knows which element was clicked even on web/canvas content Windows UI Automation can't name. Your stored screenshots are never modified, and no typed text is ever included. If marking fails for any reason, the plain screenshot is sent.
• API key required: generation uses the Anthropic API key configured in Settings → API Keys. If no key is configured, a prompt to add one is shown.
• Deleting a recording deletes its screenshots — but not instantly. Each recorded action keeps a full-screen JPEG of your desktop on disk, so the screenshots behind a recording are the same thing as a recording of whatever was on your screen at the time. Deleting a recording marks it deleted and hides it at once; the frames themselves are erased later by the retention sweep, once the recording is older than your data-retention window (30 days by default, set at Settings → Data & Storage) — the sweep removes the recording's whole directory, not just its database row. So a deleted recording stops being visible immediately and stops existing on disk within your retention window. To have frames gone sooner, shorten the retention window; erase-all clears them too.
• In-development — reveal via Settings → Lab or AMC_SHOW_TEACH_RECORDER=1.

Related

• use-skills.md — where a recorded skill is then run.
• skills-bloat-alert.md — what happens when too many skills load into every conversation at once.