Skip to content
Docs0.9GitHubInstall

GuidesReview an agent's video

Review an agent's video from your phone

Play an agent's cut in the chat, mark frames, segments, wording, and sounds, and send the notes back in one reply.

This guide is for reviewing a video an agent made without sitting at the computer. You watch the cut in the chat, point at what is wrong with marksMark. A note you attach to a place: a box on a photo, a frame or a segment of a video, a moment in an audio file, a line of a change., and the agent gets numbered notes it can act on.

The examples follow Lena, who has an agent cut a 40-second film in her project launch-film. Everything here works the same in the iPhone app and the web client.

Open the agent’s chat, hold the microphone, and say what you want; release to send. Slide left to cancel, or slide up to lock and keep talking. The transcript shows under the waveform once the agent has the note.

A 40-second cut from the screen recordings in footage/, with a voice-over.

While the agent works, the working lineWorking line. The line above the message box while an agent works: how long the turn has run, the tool call running now, and the plan's current step.Working line. The line above the message box while an agent works: how long the turn has run, the tool call running now, and the plan's current step. above the message box shows how long it has been at it, the tool call running now, and its todo list, such as todo 4/6 with the current step. Tap the line for the whole plan.

The video arrives as a message with its first frame and length. Alone in its message, it plays in place. The full-screen button opens the viewer, which has:

  • a filmstrip to scrub;
  • Previous frame and Next frame, which repeat while held;
  • Speed: 0.25×, 0.5×, 1×, 1.5×, 2×;
  • Captions, when the agent sent text tracks.

In the viewer, tap Mark. The player pauses and the Marks editor opens on the frame you were on.

Choose Box or Highlight and draw on the frame, here around a wrong logo. Then write the note under it, or hold the microphone and say it: “wrong logo here”.

Scrub to where the slow part starts and tap In. Scrub to where it ends and tap Out. Give the segment its note: “cut this in half”.

Frame marks a single moment with no drawing on it.

When the agent sends its narration and on-screen text as tracksTrack. A timed text file an agent attaches to a video: its narration, its on-screen text, or its subtitles.Track. A timed text file an agent attaches to a video: its narration, its on-screen text, or its subtitles., the editor lists them under Cues, each line with a small frame from where it starts. With more than one track, Track switches between them.

Open a cue’s menu to:

  • Change it to your exact wording;
  • Remove it;
  • Add after, or Add at play head, for a new line;
  • leave a Note on it.

The editor shows how long your new wording takes to say against the cue’s length.

An audio file in the chat has its own Mark button. On its waveform, Time marks one moment and In and Out mark a stretch. A note on a sound effect names the exact file it is about.

Tap Send. Add a message if you like, typed or spoken, and send. One reply carries up to 20 marks, numbered across everything you marked.

The agent gets each mark’s time or segment, where on the frame it is, and your note, with a picture per marked frame showing the numbers. For a cue it gets the current text and your suggested text. If the send fails, the draft stays and Retry sends it whole.

The agent sends a new render as a new message. The earlier cut stays in the chat as you saw it, so you can compare the two.

  • To render: whatever it uses on your computer, such as ffmpeg. Vendetta plays files; it doesn’t make them.
  • A format the apps play: MP4 or MOV with H.264 video (or HEVC, which not every browser plays) and AAC audio, linked in its final reply.
  • Text tracks for the Cues list: SRT or WebVTT files named after the video, such as intro.narration.srt and intro.onscreen.srt, linked next to it. A video takes up to 6 tracks.

The vendetta skill, which install.sh installs for your agents, tells them all of this.

A video holds up to 256 MiB, and one reply carries up to 20 marks. Marks lists the limits on what you send, and Agent skills the limits on what an agent sends.

A video you have played once is kept on the device, so it plays again without loading.

Spoken notes need transcription on the computer. Without it the microphone is unavailable and you type the notes.