live-tutorial

Live Tutorial

Create a short annotated MP4 from a workflow Leon actually completes in the owner's application and workspace. It is an instructional slideshow, not a continuous screen recording or a reconstruction of steps Leon did not perform.

Outcome

Return one playable MP4 with 3–6 meaningful steps. Each step should show:

• the interface before the action;
• a numbered, concise instruction;
• a rectangle around the target and an arrow pointing to it;
• enough surrounding context to recognize the location.

Keep the original captures with the tutorial artifacts. Add annotations with deterministic local rendering; never recreate the interface with a generative image model.

Every frame must be evidence from the demonstrated journey. Do not reuse one screenshot to illustrate different states, invent missing steps, or claim success when the requested result was not reached.

Workflow

1. Identify the target application, task, and successful result. Ask only if one is genuinely ambiguous.
2. Call computer_use.cua.start_recording with record_video: false and retain its output_dir.
3. Complete the real task as a sequence of consequential owner-facing actions. Retain the initial observation, then after each selected action observe once to verify its result and ground the next action. That post-action observation becomes the next step's pre-action capture; do not capture the same state twice.
4. Retry or stop when an action cannot be visually verified. Do not turn an unchanged capture, an unperformed instruction, or a guessed future state into a tutorial step.
5. Once the requested result is visible, stop interacting and render immediately. Always call stop_recording, including after a failure. Completing the original tutorial request includes assembling the MP4; do not ask for separate confirmation to finish it.
6. Annotate each successful action's own pre-action screenshot and assemble the frames into an MP4 inside the retained output_dir. Reuse that exact path rather than reconstructing it from profile or session identifiers. If a render continues in a later turn, use only artifacts from the current LEON_PROFILE and LEON_SESSION_ID; never scan other profiles or substitute a pre-existing video.
7. Make the encoded frame dimensions even by padding at most one pixel on each axis. For FFmpeg, use pad=ceil(iw/2)*2:ceil(ih/2)*2; do not resize the tutorial or alter annotation coordinates for encoder compatibility.
8. Verify that the MP4 is non-empty and readable before reporting success. Return the MP4 only if the requested workflow was completed; otherwise explain where the demonstration stopped and preserve the verified captures.