Can an AI agent edit the videos you film on your phone?
The short answer
Mostly, yes: on Oct 17, 2026 I filmed 8 clips on my phone, AirDropped them to my Mac and typed one message, and my agent on a Linux PC transcribed, cut and captioned them into a first full cut 1 h 43 min after the clips reached my Mac.
Before agents, a 5–6 minute video took me about 4 hours to edit; it's still not hands-off, because I type the message that starts it and nothing is published without my OK.
Chapters
Receipts (Oct 17, 2026, Korea time)
| What | Number | Source |
|---|---|---|
| Filmed | 8 clips, 18 min 31 s of talking, 13:38–14:04 | in the video |
| Phone → Mac → Linux PC | 14:24 on my Mac (AirDrop); 14:26 my Linux PC pulled all 8 (1.2 GB); it checks every minute | in the video |
| What I typed | 14:29 one message (17:34 a second one: keep the real background) | in the video |
| Transcribed, then Gemini watched | 14:30–14:36 transcribed; 14:43–14:45 Gemini watched 19 one-minute pieces | in the video |
| Fact-check | 14:46: 14 claims, 1 false (cut) | in the video |
| First full cut | 16:07 rendered; phone to first cut: 1 h 43 min | in the video |
| The edit | 39 pieces, 28 pauses cut (11.2 s); cuts start 0.10 s before the first word and end 0.18 s after the last | in the video |
| What broke | the Gemini call answered in 65 seconds, then didn't exit for 37 min 38 s; every model call now has a hard timeout | in the video |
Try it on your own computer
- Get clips onto your computer with a small job that runs every minute, and only move a file when its size is the same on two checks in a row, so you never grab a half-copied video.
- Use the transcript only to pick lines. Force-align the exact words you kept (I use WhisperX), then cut 0.10 s before the first word and 0.18 s after the last, and cut any pause over 0.4 s inside a sentence.
- Have a model that can watch video review the render in pieces of 70 s or less, with no tools and timestamps, and check its claims against the frames before you fix anything.
- Put a hard timeout on every model call in your own wrapper (mine is 360 s) and keep whatever it printed before it hung.
Talk to your phone, let your agent edit: the recipe and the exact prompt Free, on GitHub
Transcript
0:00 Right now, I'm shooting it. I'm probably going to send this to my system, and it's going to automatically edit it and do all the captions and all the animations for me.
0:12 I'm starting this channel because editing video was just super hard. There's so much stuff that you need to do for video editing. At least one decent video would take like 4 hours. 5 to 6 minute video would probably take 4 hours, before agents.
0:26 But now I just film it whenever I have thoughts, and then I have a system to edit it automatically for me. I just send it over there to my Linux box, and it has a certain workflow: certain things to do, certain things not to do. I think it takes about an hour or two.
0:41 It'll slice all the frames, put it all together, and then an hour later there is an alarm and, ding dong, there's a new video here. I look at it, it's like, whoa, it's way better than what I would edit it. Now, just yap whatever you want. It'll just slice it and make it look good and make it sound professional.
1:02 I even tried voice cloning. Honestly, voice cloning is not there yet. Just the emotion is the hardest part, I guess. You can't really fake it. It's almost there, but it's not. If you actually have tried the latest one, I think it's Fish Drama 3, and it's not there. I have tried it on my local computer.
1:25 A lot of the SaaS these days, I just build it myself. Most of them are pretty easy. I think SaaS is moving towards custom workflow. Honestly, the low-average SaaS is all gone. I'm not sure why you pay for it at this point. You can just prompt it to create one that fits your needs.
