live-tutorial

Live Tutorial

Demonstrate the requested workflow, then return a verified MP4 made from its real screenshots.

Write captions and annotation text in the language of the user's tutorial request, keeping app control names as displayed.

The request already authorizes completing and rendering the tutorial. Do not ask whether to continue or offer screenshots/text instead of the requested video merely because execution paused. Resume with the same recording, captures, and manifest; ask only for genuinely missing information or authorization.

Use PNG paths returned in each observation's artifacts array. The recording output_dir is the destination for the video, not the location of those screenshots; do not search it for frames. The renderer is ready to run with the manifest below: no source inspection, image-processing code, or dependency setup is needed.

Use the live tool results for recording directories, screenshot paths, and current target tokens. After a pause, read a specific saved artifact only if the checkpoint omits a required value; do not scan directories or reread evidence already available. Each new window observation invalidates earlier tokens, so make one observation with any needed query and screenshot, then act immediately from that same result. Use the shell tool only for the final renderer command.

1. Select the target application and one short route from its current view to the requested result. Demonstrate only the requested workflow, not every related option. Do not reset the application or visit unrelated settings merely to create a starting frame. If already at the result, show only the navigation needed to teach the requested step; do not claim a pre-existing state was caused by an action you did not perform.
2. Start computer_use.cua.start_recording with record_video: false; retain output_dir.
3. Use the shortest supported route: a known keyboard shortcut, an accessibility-backed action, or computer_use.cua.invoke_menu for native menus; never shell automation. Capture the starting state before acting. During recording, supported input actions automatically return a fresh screenshot and element queries include screenshots. Reuse the returned observation for verification and the next step instead of taking another identical screenshot. Observe again only if the result lacks needed evidence. Captions must describe what you actually did, including any shortcut; retain the meaningful intermediate states.
4. Always stop recording. Select 2–6 distinct captures from the successful route, normally one per action plus the result; omit setup, failed attempts, retries, and polling. Keep valid captures across retries rather than re-recording the workflow. The final result must agree with the demonstrated actions. Use a short instruction for each action and a confirmation caption for the result.
5. Write the manifest below with the file tool, then run the bundled renderer through the shell tool. Copy the returned artifact paths verbatim, including profile and session IDs; do not reconstruct them. Satellite results provide server-local copies for rendering. Do not replace this with a raw screen recording or a custom automation/rendering script. Do not inspect or rewrite the renderer unless it reports an error.
6. Check the renderer's resolvedTargets against the intended controls and the source observations. Correct any mismatch before returning fileMarker verbatim. The returned previewPaths identify the rendered step frames. If unfinished, explain the obstacle.