You commented EDIT. Here's the skill that cut that reel.
The NERF reel started as a 2:33 take with two restarts and a couple of lines I gave up on halfway. It came out at 42 seconds with captions, 20 sound effects and broadcast-loud audio. I typed one line to start it: "continue reels, nerf reel, raw take in downloads". I gave it two notes along the way: speed up the list, add SFX.
What it is
A Claude Code skill: one SKILL.md file that tells Claude which stages to run, in what order, and where to stop and show you the file. Claude does the work with ffmpeg and a few small scripts. Captions and graphics are Remotion code. No editor opens at any point.
The six stages
Find the take. It grabs the newest video in Downloads unless you name one.
Cut the mistakes. It splits the audio at silences into chunks of speech and transcribes each chunk alone. A restart shows up as its own line, so Claude keeps the clean take and drops the rest. Then it transcribes the cut and fails if any line repeats.
Speed up the middle. Hook and call to action stay at 1.0x. The middle runs at 1.25x. On NERF that took the cut from 49.5 to 41.8 seconds.
Captions and graphics. Word-level timestamps drive the captions, so each word lights up as you say it. Claude renders a still at each beat and checks nothing covers your face.
Sound effects. Claude writes a cue sheet: time, sound, and the thing on screen it hits. A cue with no on-screen event gets cut. No sound gets used more than 4 times.
Master the audio. Two passes to -14 LUFS. Before mastering, my mixes sat around -21, about 7 dB quieter.
Stages 2, 4 and 5 stop and wait for you. That's where my two notes went in.
Paste this into Claude Code
Save it as ~/.claude/skills/reel/SKILL.md, then type /reel.
---
name: reel
description: Turn a raw talking-head take into a finished vertical reel. Trigger: "/reel", "cut this take".
---
Run these stages in order. At each STOP, show me the file and wait.
1. Take: use the newest .mov or .mp4 in ~/Downloads unless I name one.
2. Cut: split the audio at silences into speech islands. Transcribe each island separately with word timestamps. Show me the island list, keep the last clean take of each line, cut the rest. Transcribe the output and fail if any line repeats or a dead gap remains. STOP.
3. Speed: hook (first sentence) and CTA at 1.0x, everything between at 1.25x. Re-transcribe afterwards, because every timing moved.
4. Edit: build the reel in Remotion at 1080x1920. Word-by-word captions from the transcript. Render a still at each beat and check that nothing covers my face. Render the full video. STOP.
5. SFX: write a cue sheet as JSON [ms, sound, on-screen event]. No cue without an event. Max 4 uses of any one sound. Mix it under the voice. STOP.
6. Master: two-pass ffmpeg loudnorm to -14 LUFS, true peak -1 dB, audio at 48 kHz. Measure the result with ebur128 and show me the numbers.Then start it with one line, like I did:
/reel raw take in Downloads, cut it tightThe mastering step, copy-paste
Pass 1 measures:
ffmpeg -i in.mp4 -af loudnorm=I=-14:TP=-1.0:LRA=11:print_format=json -f null -Pass 2 applies. Put the four numbers pass 1 printed (input_i, input_tp, input_lra, input_thresh) into the measured_ values:
ffmpeg -i in.mp4 -c:v copy -af "loudnorm=I=-14:TP=-1.0:LRA=11:measured_I=X:measured_TP=X:measured_LRA=X:measured_thresh=X:linear=true,aresample=48000" -c:a aac -b:a 256k -ar 48000 out.mp4Check it:
ffmpeg -i out.mp4 -af ebur128=peak=true -f null - 2>&1 | tail -12My NERF final read -14.3 LUFS, peak -0.9 dBFS.
Before you run it
Keep aresample=48000 and -ar 48000. Drop them and the encoder can pick 96 kHz, which Instagram may play silent. I had to re-send a reel because of it.
You need a transcriber with word timestamps. I use ElevenLabs Scribe (paid). Local Whisper is free, but on one of my takes it smoothed over four stutters Scribe caught.
It's not zero input. It does the hours. You still make the taste calls at the three STOPs.
Run it on your next take and DM me the before and after on Instagram. I read every one.
I send one of these a week: an AI tool, what it does, and how to use it without writing code. Subscribe and the next one lands in your inbox.
PS. Want this done for your business site? That's what we do at optimax-ai.com.