I’m terrible at video editing. I don’t have the creative instinct a lot of people have of knowing which clip goes where, which caption style fits which video style, when a cut feels snappy versus jarring, and what type of video effects would effectively hook people to keep watching. I also don’t have the patience to sit at a timeline and figure it out by trial and error. Editing, to me, is where a good idea goes to die in a scrubbing bar.
There aren’t a lot of tasks I do that make me feel this way, and in 2026, every task that makes me feel this way feels like one I shouldn’t have to do myself anymore. That’s the itch, right? Everything else I dread has quietly been handed off to some agent or another. So, instead of opening an editor, I opened my terminal and asked Claude Code to edit the video for me.
An Anthropic engineer gave me the idea
Plus a bit of editing anger
I’m not ashamed to admit that I was an iPad kid and I grew up watching YouTubers, and I’ve always loved the idea of creating content. Between my passion for journalism, tech, and content, and the fact that I know people nowadays are more receptive to short-form content than long-form (this hurts to write), I decided to create a tech-focused Instagram account a while ago and begin posting tech-related reels. The script-writing, hunting for topics, figuring out what works for my audience and what doesn’t — none of that was a problem. What I struggled with was editing, and I admittedly gave it the least effort I possibly could.
My reels went up rough, under-edited, and I made peace with the fact that the editing would always be the weakest part. Even then, I spent a good hour trying to make the video the best I could. A few weeks ago, Thariq Shihipar, a Member of Technical Staff at Anthropic, shared that he used Fable 5 to edit Fable 5’s own launch video.
Now, I use Claude Code for practically everything I can. I’m a big fan of using it for non-coding uses, and while I’ve used it to create slide decks and edit images, editing a full-fledged video from the command line without a GUI was a leap I hadn’t even thought to make.
Seeing someone do it (and do it well enough to ship) was just the nudge I needed. If Claude Code could edit a launch video, surely it could survive one of my reels. So, I used a lot of his advice as a starting point, opened my terminal in an empty folder, and asked Claude Code to do the one part of content creation I’d never willingly do myself.
The whole edit ran on a transcript
The cutting was where it struggled
Before I jump into the prompts I used, it’s important for you to understand my filming process, since the prompts are extremely specific to it.
When I film, I don’t do takes. I don’t stop the recording, reset, and start again. I just keep talking. I say a sentence, don’t like how it landed, and say it again. And again. Until one of them sounds like a person and not someone reading off a script. Then I move to the next line and repeat the whole ritual. What comes out the other end is a single nine-minute take that’s maybe two minutes of usable video buried under five repetitions of every sentence. The keeper is almost always the last attempt since I’ve warmed up and stopped fumbling by then.
So the first and most tedious job in any of my edits isn’t captions or effects. It’s sitting there, listening to myself say the same thing five times, and keeping only the last clean one. That’s the work I wanted to hand off to begin with. My opening prompt laid the whole thing out before letting it touch anything:
I want to edit a 9-minute talking-head video for Instagram in a casual “yap” style (jump cuts, word-by-word captions, sound effects, B-roll images). I filmed it as one continuous take. My editing style: I repeat each sentence a few times until I like a delivery, then move on — so the raw footage has lots of duplicate/false-start takes. When I repeat a line, the LAST repetition is the keeper. Cut the earlier attempts.
I also told it, explicitly, not to charge ahead: check my environment first, propose a folder structure and a step-by-step plan for me to approve, and never cut anything destructively. I wanted a cut list I could review before a single frame was deleted, and my original file left untouched no matter what.
It respected all of that. It checked what I had installed (I was missing ffmpeg and a transcription tool), told me what to install, and laid out an eight-step roadmap: transcribe, detect repeated takes, rough cut, tighten, captions, B-roll, sound effects, export. Every step wrote to a new file, so the original never moved.
The plan hinged on transcription. Claude Code used WhisperX to turn my audio into text with a timestamp on every single word, and that word-level timing became the backbone of the entire edit. Once it could see my speech as timestamped text, it could find the runs of near-identical sentences, keep the last one, and build a cut list for me to sign off on. I approved it, it assembled a rough cut, and then showed me the result.
When I didn’t like the way a couple of takes had been picked, I told it which ones felt off, and it re-cut them and showed me the result again. I then noticed there was still a lot of repetition in the video, so I flagged it and Claude Code got to work again. This cycle did continue for a while, since the repetition seemed to keep surviving every pass. I’d watch a new cut, catch the same stammering doubling back on itself, flag it again, and wait while it re-rendered.
Claude Code then stopped patching the cut and figured out the actual cause: the transcription software was collapsing my rapid repeats into a single timestamp, so every cut was quietly dragging the stutters back in. It re-transcribed with settings that forced each repeat into its own chunk, rebuilt the cut at the word level, and finally most of the stammering was gone.
There was still a tiny bit of repetition left at the end, but that was more an issue with how I’d recorded than anything Claude Code did wrong. The point here is that Claude Code managed to take a 9-minute video of me rambling the same lines over and over and turned it into a tight 1 minute 22 seconds video, and all I had to do is hit Enter on my keyboard. This was the toughest part of the process, and once it was behind us, the rest moved a lot faster.
With the cut locked, Claude Code moved on to captions
Then it did the parts I was sure it’d fail
I wanted the trendy, all-caps, one-word-at-a-time style where each word pops as you say it. This is the part I was most sure it would fumble, because animated captions are fiddly even in a real editor! It hit a wall almost immediately: the version of ffmpeg I’d installed had its text-rendering filters stripped out, so it simply couldn’t draw text onto video the normal way. Instead of giving up, Claude Code drew the captions itself by rendering each word as its own image frame, big and bold with a thick outline and the active word popping yellow, then layering those frames over the video. It timed every word to my voice using the same transcript from earlier, and it even autocorrected the brand names the transcriber had misheard along the way (it kept hearing “SadaPay” as “Sadabay,” among other things).
It then added B-roll. I instructed it to find its own images rather than supplying them myself, so it went and pulled real ones off the web (logos and photos relevant to what I was talking about) then built them into clean, consistent branded cards and cut to them on the exact word they related to. When I listed off a bunch of apps, the matching logos popped in one after another, timed to each name.
The sound effects were the part that made me laugh, because it couldn’t download any. So, it generated a soft whoosh and a light pop from scratch in code, then mixed them in low and subtle: a whoosh on the bigger cutaways, a pop on the rapid-fire logo snaps. It even caught that the pops were sitting too quietly under my voice to be heard and bumped them up until they actually registered. Finally, it exported the whole thing as a MP4 file in the folder.
With a little editing, the video was better than my usual
It took an absurd amount of time though
The captions, images, and even the sound effects Claude Code added to the video were all perfect. They didn’t require any editing from my end whatsoever. However, as I mentioned above, I noticed imperfections with the cut itself. There were a few instances where the cut wasn’t smooth and some where two takes hadn’t been fully separated. So, I did need to head to clean those up, and it took me a total of two minutes and thirty-three seconds (yes, I kept a timer).
Now, while the final video ended up being something I’d totally use, the biggest drawback here is the initial part of the edit took an absurd amount of time. The cutting stage (the transcribing, the re-transcribing, the diagnosing, the re-rendering) ate up a chunk of my afternoon. Most of it was spent watching a timer tick and typing “where are we at” while my poor MacBook heated up. A human editor (and even me) who knew what they were doing would’ve made those cuts in a fraction of the time.
The best video editor I’ve used doesn’t know it’s a video editor
This was a super fun experiment, and it goes to show just how much you can do within a tool like Claude Code. In no way is it limited to just writing and debugging code! I’ve used it to build slide decks, edit images, and now cut, caption, and score an entire video. None of these tasks involve it doing the one thing it’s ostensibly built for. Give it a terminal, the right tools to install, and a clear enough description of what you want, and it’ll happily wander far outside its lane and figure the rest out as it goes.
Would I reach for it every time I need to edit a reel? Probably not, at least not right now. However, as a way to hand off the single most tedious part of making content, it did the job, and it did it better than I would have. That’s more than enough to make me do it again!
