WEBVTT 00:00:00.000 --> 00:00:04.929 So a client asked me this one recently — can I edit video with AI? And this one's 00:00:04.929 --> 00:00:09.920 a toss up. Some of it you can do. Some of it you can't. And the line between those 00:00:09.920 --> 00:00:12.446 two is not where most people think it is. 00:00:12.446 --> 00:00:16.554 The client's a video editor. Does all kinds of work, all kinds of clients. So 00:00:16.554 --> 00:00:20.662 this wasn't somebody trying to skip learning the craft. This was somebody who 00:00:20.662 --> 00:00:24.986 already had a process, and wanted to know how much of it a machine could take off 00:00:24.986 --> 00:00:25.635 their plate. 00:00:25.635 --> 00:00:29.917 We tried the ambitious version first. It didn't work. Then we tried something 00:00:29.917 --> 00:00:34.369 smaller... and it took an edit that was a two day job down to about an hour. The 00:00:34.369 --> 00:00:37.524 difference between those two things is the whole episode. 00:00:37.524 --> 00:00:41.580 This is AI-Honestly. I'm Drew... Each week I take one real question from a 00:00:41.580 --> 00:00:45.802 founder or business owner and tell you what's real, what's not, and what it'd 00:00:45.802 --> 00:00:46.580 take to build. 00:00:49.580 --> 00:00:53.496 I get this question a lot, and I get it, because there's a ton of noise in this 00:00:53.496 --> 00:00:55.756 space and most of it is about the wrong thing. 00:00:55.756 --> 00:01:00.141 When people say AI video, they're mostly talking about generating video. Making 00:01:00.141 --> 00:01:04.694 footage that didn't exist. That's its own world, it moves fast, and it is not what 00:01:04.694 --> 00:01:06.437 I'm talking about today at all. 00:01:06.437 --> 00:01:10.290 I'm talking about editing. You've got real footage. A real shoot, a real 00:01:10.290 --> 00:01:14.415 company, real people on camera. It has to be accurate, it has to be on brand. 00:01:14.415 --> 00:01:18.431 Nobody is generating anything. The question is whether AI can take that raw 00:01:18.431 --> 00:01:20.276 footage and cut it into something. 00:01:20.276 --> 00:01:24.412 And there's a second thing hiding inside the question. There are editing programs 00:01:24.412 --> 00:01:28.341 out there with AI features built in. That's a real category, and some of them 00:01:28.341 --> 00:01:32.321 are fine. But that means learning a new piece of software and moving your work 00:01:32.321 --> 00:01:36.095 into it. That's not what this client wanted. They wanted to keep their own 00:01:36.095 --> 00:01:39.920 process. Their tools, their brand graphics, their way of doing things. They 00:01:39.920 --> 00:01:42.195 just wanted a machine to do the tedious part. 00:01:42.195 --> 00:01:45.929 So the version of the question I'm answering is this. I have a process. Can 00:01:45.929 --> 00:01:49.814 I point Claude Code, or Codex, or one of those harnesses at it, and have it do 00:01:49.814 --> 00:01:52.690 some of my editing for me. Without moving to new software. 00:01:52.690 --> 00:01:55.476 Yes. But maybe not the version you're hoping for. 00:01:55.476 --> 00:01:59.684 Here's what we tried first. Take a big pile of raw footage. Have the AI watch it 00:01:59.684 --> 00:02:03.839 — look at the visuals, figure out what's happening in each shot — pick the good 00:02:03.839 --> 00:02:08.154 stuff, and cut its own version together. End to end. Footage in, edited video out. 00:02:08.154 --> 00:02:09.036 That did not work. 00:02:09.036 --> 00:02:13.431 Couple of ways it didn't work. The picking was okay-ish. It could tell you 00:02:13.431 --> 00:02:18.368 roughly what was happening in a shot. But the cutting was bad. And not bad like the 00:02:18.368 --> 00:02:23.124 artistic choices were wrong. Bad like the timing was just weird. Cuts landing in 00:02:23.124 --> 00:02:25.893 strange places. Nothing you'd hand to a client. 00:02:25.893 --> 00:02:29.993 And then I made it worse, because I wanted more. I wanted B-roll laid in. I 00:02:29.993 --> 00:02:34.204 wanted lower thirds, the little name graphics. Every one of those was another 00:02:34.204 --> 00:02:38.471 layer of complication stacked on a thing that wasn't working right at the base 00:02:38.471 --> 00:02:38.803 level. 00:02:38.803 --> 00:02:41.854 So we backed off, and tried something different. And this is the part that 00:02:41.854 --> 00:02:42.147 worked. 00:02:42.147 --> 00:02:45.458 Instead of asking it to look... we asked it to read. 00:02:45.480 --> 00:02:49.398 First step is transcripts. Whisper — that's the open source speech-to-text 00:02:49.398 --> 00:02:53.746 model, it's free, you run it on your own machine — pulls a transcript out of every 00:02:53.746 --> 00:02:57.503 video file. And now the footage isn't footage anymore. It's words, with 00:02:57.503 --> 00:02:58.576 timestamps attached. 00:02:58.576 --> 00:03:01.872 And words, AI is very good at. So now you can say: here's everything anybody said 00:03:01.872 --> 00:03:04.962 across all this footage, find me the pieces that make a story. And it can do 00:03:04.962 --> 00:03:06.610 that. It's a text problem at that point. 00:03:06.610 --> 00:03:10.957 Then it built us a little website. Just a local page where you could scroll through 00:03:10.957 --> 00:03:15.145 all the proposed clips, see the script those clips would add up to, and read the 00:03:15.145 --> 00:03:19.015 whole thing before anybody touched a timeline. No visuals. Just words. And 00:03:19.015 --> 00:03:23.097 that's where we made changes. Cut this bit. Move that one up. This part drags. 00:03:23.097 --> 00:03:27.576 Once the script was right, it saved out every clip as its own file. With room on 00:03:27.576 --> 00:03:32.055 the front and room on the back of each one. Handles. So there's material to trim 00:03:32.055 --> 00:03:32.338 into. 00:03:32.338 --> 00:03:35.922 And then those clips go into Final Cut Pro. The client's Final Cut Pro. Their 00:03:35.922 --> 00:03:38.468 project, their process, the way they've always done it. 00:03:38.468 --> 00:03:41.064 That's the whole thing. So here's the before and after, because this is the 00:03:41.064 --> 00:03:41.766 number that matters. 00:03:41.766 --> 00:03:46.361 Before, a job like that is one to two days. That's a person sitting there 00:03:46.361 --> 00:03:51.338 watching all the footage. Finding the usable clips. Figuring out what the story 00:03:51.338 --> 00:03:56.507 even is. Getting it into an order that's compelling. One to two days, and honestly 00:03:56.507 --> 00:04:00.017 most of that is not the skilled part. It's just looking. 00:04:00.017 --> 00:04:03.128 After, it's about an hour. An hour in Final Cut, making the cuts land. 00:04:03.128 --> 00:04:07.241 Now I want to be fair about that number, because there's time in there that 00:04:07.241 --> 00:04:11.632 doesn't show up. The AI takes a while to go get the clips and build the script — 00:04:11.632 --> 00:04:16.134 that's real time. But you're not sitting there watching it happen. Go do something 00:04:16.134 --> 00:04:20.692 else. And then reading through the script and making your changes, that takes a bit 00:04:20.692 --> 00:04:24.583 too. But that's reading. That's not scrubbing a timeline for six hours. 00:04:24.583 --> 00:04:28.234 And not because the machine edited anything. Because nobody had to sit there 00:04:28.234 --> 00:04:30.620 and hunt through hours of footage for the moments. 00:04:30.620 --> 00:04:34.527 But be clear about what came out the other end. We did not finish with an 00:04:34.527 --> 00:04:38.651 edited video. We finished with clips. Sorted, in order, with a story attached 00:04:38.651 --> 00:04:42.776 to them. Somebody still had to make the cuts land where they wanted. Somebody 00:04:42.776 --> 00:04:47.009 still had to drop in the brand graphics they've been using for years. The taste 00:04:47.009 --> 00:04:49.939 part stayed with the human. The tedious part went away. 00:04:49.939 --> 00:04:51.007 So what would it take. 00:04:51.007 --> 00:04:55.980 Tools first... Whisper, which is free and open source. F.F.mpeg, which is the 00:04:55.980 --> 00:05:01.018 command line tool that does the chopping of the actual video files. Also free. 00:05:01.018 --> 00:05:05.990 It's not sitting on your machine by default, but it's one command to install, 00:05:05.990 --> 00:05:11.093 and it's been the thing under the hood of basically every video tool for twenty 00:05:11.093 --> 00:05:11.486 years. 00:05:11.486 --> 00:05:15.481 And then a coding harness. Claude Code, Codex, whichever one you're already 00:05:15.481 --> 00:05:15.805 using. 00:05:15.805 --> 00:05:18.545 That's it. No new subscription. Nothing added to the monthly. 00:05:18.545 --> 00:05:23.253 And here's the part people get wrong about cost. You'd think this would burn a 00:05:23.253 --> 00:05:27.777 huge amount of AI usage, because it's video, and video sounds expensive. It 00:05:27.777 --> 00:05:32.545 doesn't. Most of the work is happening down in the terminal, through those free 00:05:32.545 --> 00:05:37.070 packages. The AI is doing the thinking part — reading transcripts, choosing 00:05:37.070 --> 00:05:40.371 clips, writing the commands. That's cheap. That's text. 00:05:40.371 --> 00:05:43.701 So even on the twenty dollar a month plan — the entry level one — you can do this. 00:05:43.701 --> 00:05:45.633 That surprised me a little. 00:05:45.655 --> 00:05:49.972 Who can do it. I'll be straight with you. This is terminal work. You are going to 00:05:49.972 --> 00:05:53.857 be looking at a command line. If you've never done that, this is not your 00:05:53.857 --> 00:05:58.012 Saturday afternoon project. But the good news is it's a process — you build it 00:05:58.012 --> 00:06:00.980 once, and after that you're just feeding it new footage. 00:06:00.980 --> 00:06:02.466 Now. What to watch for. 00:06:02.466 --> 00:06:07.250 The big one, and I want to say this plainly. You cannot dump raw footage in 00:06:07.250 --> 00:06:12.033 one end and get a polished, on-brand, finished video out the other end. Not 00:06:12.033 --> 00:06:17.076 yet. Anybody showing you that is either showing you a demo with unusually clean 00:06:17.076 --> 00:06:21.924 footage, or they're showing you generated video, which is a different thing. 00:06:21.924 --> 00:06:26.914 Second one. The split I'd draw is between words and pictures. If your video has 00:06:26.914 --> 00:06:30.880 people talking — interviews, testimonials, a founder explaining 00:06:30.880 --> 00:06:35.550 something, a webinar — you can go a long way, because it turns into a text 00:06:35.550 --> 00:06:40.668 problem, and text is the strongest thing these models do. If your ask is visual — 00:06:40.668 --> 00:06:45.786 find me every shot where the product is on screen, find me the good drone stuff — 00:06:45.786 --> 00:06:50.903 it can do that. It's just less reliable. And less reliable is a real problem when 00:06:50.903 --> 00:06:53.782 the person watching the result is your client. 00:06:53.782 --> 00:06:57.841 Third. Watch what you pile on top. Every extra thing I wanted, the B-roll, the 00:06:57.841 --> 00:07:01.689 lower thirds, made the whole thing shakier. There's a version of this that 00:07:01.689 --> 00:07:05.485 works because it's narrow. The more you add, the faster it stops working. 00:07:05.485 --> 00:07:09.334 And fourth. What you're buying here is time. Not a finished product. If you're 00:07:09.334 --> 00:07:13.283 going into this expecting to remove the editor, you're going to be disappointed. 00:07:13.283 --> 00:07:17.131 If you're going into it expecting to hand the editor a running start... that's 00:07:17.131 --> 00:07:18.581 real, and that's worth a lot. 00:07:18.581 --> 00:07:23.511 That's it for this episode. If you've got a question you've been wondering about — 00:07:23.511 --> 00:07:28.137 can AI do this thing? — send it to me. The address is in the show notes. I'll 00:07:28.137 --> 00:07:32.945 pick the good ones and we'll figure out together how real they actually are. I'm 00:07:32.945 --> 00:07:34.649 Drew. This was AI, Honestly.