I asked an AI agent to edit my video. Here are all three versions.

1. Give the agent the job, the footage, and the brand

I recorded a short video asking whether AI could edit as well as I can. I wanted the raw take and the edited version beside each other so people could judge the result. I also gave some instructions out loud while recording, including a diagram, a zoom on my coffee cup, and quiet retro music.

I was working in Codex with Tesseract by Mirage, which lets an agent build and render an editable video project locally. Linear supplied the task brief, and GitHub supplied my video brand files.

There were three different video ideas on the card. I had to tell the agent which one we were editing. Here is my first request, with the original spelling and app mentions preserved.

Eight clips were copied from the card and checked against the originals. The selected complete take ran 57.375 seconds. The other concepts stayed separate. I then pointed the agent at my existing video assets instead of describing a new look from scratch.

The videos use my navy, blue, light blue, white, and teal palette, with Syne and DM Sans. The font files mattered as much as the color names. The first font import needed a compatibility fix before the rendered text actually used the intended letterforms.

2. Watch the first cut before fixing it

The first version gave me a playable comparison with captions and graphics. It also left the things I would immediately change in an edit: a caption rectangle sitting too low, explainers that felt static, and too much footage after I finished talking.

Watch the opening captions, then jump to the coffee and the ending. Each player shows the original on the left and that version's edit on the right. You hear one edited mix.

The file passed its technical checks, but I still didn't like the treatment. I described the changes in terms of things I could see and hear.

That gave the agent specific work: move the captions onto my torso, remove their background, animate the explainers, make the coffee cue noticeable, and end the clip after the final thought. Asking for a more engaging edit had left too many of those decisions unspecified.

3. Move the captions and give the graphics a purpose

Version 2 moved the words up onto my torso and removed the caption boxes. Words appeared at their spoken timing, while short phrases stayed together on screen. The graphic elements gained entrances and exits, with the camera, ChatGPT Work, and Tesseract explanation appearing in sequence.

The coffee zoom got a stronger warm sound cue, and the music dipped around it. The ending lost 2.625 seconds of recording-stop footage. A closing push-in carried the final question, and the clip stopped at 54.75 seconds.

I said, "that version works for now." Seeing the result at the final comparison size helped. A caption can look fine in a full-height portrait preview and become too small when the video shares the frame with the original.

I kept the complete take in the editable project. Both sides of the revised comparison use the same source timing, and both end together. Trimming the end of the delivery did not delete the original recording.

4. Give the captions more range and stop repeating the ding

The next request was smaller. I wanted stronger captions and more variety in the sound effects.

Version 3 kept the same 167 verified words and their timing. It regrouped them into 42 natural phrases, down from 70 shorter groups in v2. Supporting words made a small entrance. Selected words appeared larger, rose a little, and settled once. Five blue underlines drew attention to particular phrases.

The captions also stayed white while the explainer used teal. I could read the words without having two competing teal elements asking for attention.

The nine sound cues now have different jobs. The camera graphic gets a double-click, the chat action gets short typing and send ticks, and the Tesseract handoff gets a snap. Air sounds handle transitions. The coffee moment keeps its warm accent, and the ending gets a gentle impact. The words themselves do not each make a noise.

My response was "banger, well done." I asked for v3 to be uploaded and its location recorded on the task.

Where the audio actually came from

I didn't need to supply a sound folder for these effects. Tesseract's local helper can make simple procedural taps, clicks, pops, pings, whooshes, risers, downlifters, and impacts. Combining those sparingly produced the different accents.

The music was a separate job. This edit used an instrumental bed synthesized locally with Python and imported into Tesseract. Calling that Tesseract music generation would be inaccurate. If I wanted a particular licensed track or recorded sound, I would supply it or ask the agent to use another authorized source.

The captions also depended on local transcription outside Tesseract and a verified word map. Tesseract handled the editable text and animation; it did not transcribe the recording for us.

5. Keep my eyes level when the camera moves

After approving v3, I added another instruction. Long stretches of narration need subtle camera movement, especially when there is no graphic on screen or when a framing change can help cover a cut. During a zoom, my eyes should stay at the same height in the frame.

That became a rule in /video-edit-tesseract. A typical narration move starts with a slow 2 to 5 percent push or release over 3 to 6 seconds. The agent pairs the zoom with a position adjustment so the camera movement does not pull the eyes toward the top of the frame. It checks the position after cropping, including inside the smaller comparison panel.

This broader narration movement has not been rendered into v3. The cup close-up and closing push-in were already there. The new rule is saved for the next edit, with a calculator and instructions to inspect the start, middle, and end of each move. Ordinary slow camera movement stays silent.

6. Start the next edit with the instructions already saved

The skill now tells the agent to verify the words before captions, plan graphics around spoken cues, keep captions on the torso, vary sound by purpose, and trim the recording-stop pause. It points to the approved edit for reference. I should not have to discover the same low caption boxes again on every first cut.

It also keeps the work editable. I get an MP4 to watch and a native .tsrct project that retains the footage, text, graphics, and sound layers. The agent checks the encoded video as well as its previews, including caption overlap, cue timing, the ending, and whether the file plays all the way through.

This is a new prompt for the next job, built from the saved instructions. It is not the prompt that produced v3 in one pass.

To try the underlying tool, start with Mirage's Tesseract repository and installation instructions. My style settings sit on top of its video-editing and motion-graphics skills. The separate video about tracking coding agents in Linear and the coffee edit are still waiting. I can use this skill when I pick those up.