A 3-minute interview video cut entirely by Claude: the transcript-first workflow behind it
Ubon, a business analyst, turned 16 raw interview clips into a 3-minute church Youth Week video with captions, title cards, a logo, music and cleaned-up audio, all without opening an editing app. The thread explaining how passed 1,300 bookmarks in a day. Here's the method in our own words.
The setup. Claude has no timeline. It edits by writing code and running video tools in the background, so your job changes from editor to director: say what you want, check what comes back, give notes. Ubon used a Claude chat that can run code and create files, uploaded the clips (or connected a folder through the desktop app or Google Drive when they were too big), and renamed bad takes with SKIP so Claude left them out.
Step one is the important one: transcribe first. Claude can't watch video the way you do, so the first request was a transcript of every clip with word timings, plus still frames from each. After that it knew who said what, when, and what every shot looked like. Because it cut from the transcript, edits landed between words rather than in the middle of them.
The brief and first cut. Instead of 'edit this', the brief said what the piece was (everyone answers the same question), to use each person's best take, skip repeats, stay under three minutes and use the church's fonts and colours. Claude removed pauses and false starts, punched in slightly where two parts of one answer were joined so the cut wouldn't jump, put the question on screen as a title card, and added an opening and closing logo card.
Captions and sound. With word timings already in hand, word-by-word captions were easy: just name the font, colour, size and position. The sound pass surprised Ubon most. Per the thread, Claude ducked the music whenever someone spoke, separated a song's vocals from its instruments, cleaned background noise behind one speaker and set the overall loudness to a normal level for social.
Notes with timestamps. The most useful habit is giving exact notes, such as 'lips and voice drift out of sync from 1:40', 'cut the mistake at 2:18', 'noise from 2:29 to 2:35'. Claude found the cause, fixed it and re-checked. Two more tricks from the thread: send a screen recording of a reel you like and ask it to copy the timing (Ubon did this for a 35-photo slideshow, with a flash and shutter sound on each change), and before you watch, ask it to check frames around every cut, lip sync and volume.
Limits, per the author. Each version takes a few minutes to render. It only fixes what it finds or what you point out. Colour grading and fancy effects are still easier in CapCut or Premiere, and you still need the rights to any music you use. The starter prompt Ubon shared, roughly: transcribe my clips first, then make a [length] video for [platform] about [topic], use each person's best take, cut long pauses, add word-by-word captions in [font and colour], add opening and closing cards with my logo, and keep the music low under speech.
Why it matters. This is the same probe, edit, verify loop that agent toolkits like ffmpeg-skill are built around, done by a non-editor in a chat window. If you cut interview or testimonial content, the transcript-first step and timestamped notes carry over to any agent setup.
Signals: the author's own thread took off (396 likes, 1,313 bookmarks and 55k views within a day of 11:50 AM MST Nov 16), and tracked curator @RoundtableSpace picked it up at 10:15 PM MST. That's HOT by our rule.
Related tools on ReelStack
More from the Lab
- Mariko Mori's Claude-coded animation installation, ElevenCreative Dub & Sync, and HyperFrames 0.9.36 to 0.9.39
- Concat: an open-source, local-first CapCut replacement that Claude Code can drive through a built-in MCP server
- Kapwing's MCP lets Claude or Codex caption, translate and voice-clone-dub a video, and the team published three slash-command skills
- Don't start an Opus 5.5 film from a blank prompt: the open repos that show their prompts and code
Topics: Audio, voice & music