I’m not good at video editing. I don’t have the creative instincts that many people have, such as which clips go where, which caption styles go with which video styles, at what points do cuts feel snappy or jarring, and what kinds of video effects effectively keep people watching. I also don’t have the patience to sit on a timeline and figure it out through trial and error. For me, editing is where good ideas go under the scrubber.
There aren’t many jobs that make me feel this way. In 2026, all the tasks that make you feel this way will feel like things you no longer need to do yourself. That’s what’s itchy, isn’t it? Everything else I’m afraid is being secretly handed over to some agent. So instead of opening an editor, I opened a terminal and asked Claude Code to edit the video.
Anthropic engineers gave me the idea
Plus a little bit of editing anger.
I’m not ashamed to admit that I was a kid with an iPad, grew up watching YouTubers, and have always loved the idea of creating content. Between my passion for journalism, technology, and content, and the fact that people these days are more receptive to short-form content than long-form content (which is painful to write), I decided a while ago to create a tech-focused Instagram account and start posting tech-related Reels. Writing the script, finding the topic, figuring out what works and what doesn’t for the audience, none of that was a problem. What I struggled with was the editing, but I’m sure I put in the least amount of effort possible.
My reel was poorly edited and shoddy, but I accepted the fact that editing is always the weakest part. Still, I spent a lot of time making the video the best it could be. A few weeks ago, Thariq Shihipar, a member of Anthropic’s technical staff, revealed that he used Fable 5 to edit the Fable 5 launch video.
Now I use Claude Code for everything I can. I love using it for non-coding purposes, and have used it to create slide decks and edit images, but editing full-fledged videos from the command line without a GUI was a leap I hadn’t considered.
Seeing someone doing it (and doing it well enough to ship) was exactly the boost I needed. If Claude Cord can edit the launch video, he can definitely live on one of my reels. So I used much of his advice as a starting point, opened a terminal in an empty folder, and asked Claude Code to do some of the content creation I’d never want to do myself.
The entire editing was carried out based on the transcript
The difficult part was cutting.
Before I get into the prompts I used, it’s important to understand my filming process because the prompts are very specific.
When I shoot, I don’t do takes. It does not stop recording, reset it, and start it again. Just keep talking. Say a sentence, don’t like how it lands, and say it again. And again. Until one of them sounded like a human being and not the person reading the script. Then move to the next line and repeat the entire ritual. What comes out the other side is one nine-minute take, with perhaps two minutes of usable video buried in five repetitions of each sentence. The keeper will almost always be the last try, as by then you’ve warmed up and stopped fumbling.
So the first and most tedious part of my editing process isn’t captions or effects. It’s sitting there, listening to yourself say the same thing five times, and only holding onto the last clean word. That’s the job I wanted to take over from the beginning. My opening prompt explained the whole thing before letting me touch on 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, and B-roll images). It was shot in continuous takes. My editing style: I repeat each sentence a few times until I like the delivery, then move on. As a result, the raw footage contains many duplicate takes and false starts. When I repeat a line, the last repeat is the keeper. Cut previous attempts.
I also clearly told him not to rush in first. Look at your environment first, suggest a folder structure and step-by-step plan for approval, and never cut anything destructive. I needed a cut list that I could review before a single frame was removed and the original file left intact no matter what.
He respected all of that. It checked what I had installed (ffmpeg and transcription tools were missing), told me what to install, and gave me an eight-step roadmap: transcription, repeat take detection, rough cuts, closing, captions, B-roll, sound effects, and export. Every step writes to a new file, so the original file is not moved.
The plan depended on transcription. Claude Code used WhisperX to convert my audio to text with timestamps on every word. That word-level timing became the backbone of the entire edit. Once it was able to recognize my speech as time-stamped text, it was able to find sequences of nearly identical sentences, keep the last one, and create a cut list for me to approve. I approved it, assembled a rough cut, and showed me the results.
When I didn’t like the way some takes were selected, I told them which takes I didn’t like and they recut the takes and showed me the results again. Then I noticed that there was still a lot of repetition in the video, so I flagged it and Claude Code started working again. This cycle continued for a while as this repetition seemed to continue to persist on every pass. I looked at the new cut, saw the same stutter doubled, flagged it again, and waited until it was re-rendered.
Claude Code then stopped patching the cut and discovered the actual cause. The transcription software had combined the fast repeats into a single timestamp, so every cut was silently unstuttering. It was retranscribed with a setting that forced each repeat into its own chunk, reconstructing the cuts at the word level, and ultimately eliminating most of the stuttering.
There was still some repetition at the end, but that was more a problem with the way I recorded it than with Claude’s code being wrong. The important thing here is that Claude Code managed to turn a nine minute video of me rambling the same lines over and over into a tight 1 minute 22 second video. All I had to do was press the Enter key on my keyboard. This was the most difficult part of the process, but once this was done the rest went much faster.
With the cut locked, Claude Code has moved on to the captions
Then the part I was sure would fail ran
I wanted a trendy, all-caps, one-word-at-a-time style where each word pops out as you say it. This is the part I was sure I would fail the most, as animated captions are tricky even in real editors. I hit a wall almost immediately. The version of ffmpeg I installed had the text rendering filter removed, so I couldn’t draw text on the video in the normal way. Claude Code didn’t give up, rendering each word as its own image frame, rendering the active word large and bold with a thick outline and the active word popping out in yellow, and overlaying those frames on top of the video to draw the caption itself. It used the same transcription as before, timing every word to my voice, and even auto-corrected brand names that the transcriber misheard along the way (in particular, “SadaPay” kept sounding like “Sadabay”).
Next we added B-roll. I told them to find their own images rather than provide them themselves, so I grabbed real images from the web (logos and photos related to what I was talking about), incorporated them into a clean, consistent branded card, and cut out the relevant words exactly. When I listed a large number of apps, matching logos were displayed one after another for each name.
I couldn’t download the sound effects, so they made me laugh. So I created a soft hiss and a light pop in the chord from the beginning and mixed them together low and subtly. A whooshing sound on a large cutaway, a popping sound on a rapid-fire logo snap, etc. I realized the pops were too quiet under my voice to be heard, so I turned them up until they actually registered. Finally, I exported the whole thing as an MP4 file to a folder.
With a little editing, the video turned out better than usual.
It took an extremely long time, but
The captions, images, and even the sound effects that Claude Cord added to the video were all perfect. There was no need for any editing on my part. However, as mentioned above, I found the cut itself to be flawed. There were several instances where the cuts were not smooth or the two takes were not completely separated. So I had to head over to put them away, which took me a total of 2 minutes and 33 seconds (yes, I had set a timer).
Now, the final finished video is something I absolutely want to use, but the biggest drawback here is that the first part of the editing took an inordinate amount of time. The cutting stage (transfer, re-transfer, diagnosis, re-rendering) ended up taking up a good chunk of the afternoon. Most of it was spent watching the timer tick and typing “Where am I?” while my poor MacBook heated up. Any human editor who knew what they were doing (and even me) would have made these cuts in a split second.
Best video editor I used, didn’t know it was a video editor
This is a really fun experiment and shows how much you can do within a tool like Claude Code. It’s not just limited to writing and debugging code. I’ve used it to create slide decks, edit images, and now cut, caption, and score entire videos. None of these tasks fulfill the single purpose they were ostensibly built for. If you give it a terminal, the right tools to install, and a clear enough explanation of what it needs, the car will happily roam far outside its lane and figure out the rest as it goes.
Will you reach for it every time you need to edit a reel? Probably not, at least not right now. But as a way to take over the most boring parts of content creation, it worked well enough and worked better than I would have done it. More than enough to do it all over again!
