My 1:53 raw intro was full of fumbles, spoken retakes and an 11-second pause. Claude Code turned it into 46 seconds with the names spelled right, captions on and the sound at -14 LUFS for YouTube. I never opened an editing app. This page walks through the whole flow I use for every long-form video on my channel, step by step.
Why one prompt does not edit a video
The first thing everyone types is one line: "edit this video for YouTube". Claude will try, and it will tell you it is done. However, when I tried it, the loudness was about 8 decibels too quiet, there was no list of what was cut, and it had misheard my own name.
The fix is to split the job into steps. Each step reads what the step before it left in the folder, does one thing, and writes a report you can check. That is the whole method. The prompts are plain English, and Claude Code writes and runs the scripts behind them.
What you need before you start
- Claude Code on a paid Claude plan. Pro is the minimum.
- ffmpeg, ffprobe and Python 3. Step 0 checks for them and installs only what is missing.
- A Whisper model that times every word. mlx-whisper on Apple silicon, faster-whisper anywhere else, model large-v3-turbo.
- HyperFrames, free and open source from HeyGen. It turns a web page into video, so Claude can design title cards, callouts and captions. It needs Node.js 22 or newer.
- One folder per video, with your raw take in it as
raw.mp4.
One habit makes the edit far easier: say "retake" out loud whenever you fumble a line while recording, then say the line again. Step 4 looks for that word.
Steps 0 to 6: from raw take to a clean cut
These are the seven prompts I build live in the video. Each one is a single plain-English request.
- Step 0, check the tools. Claude checks ffmpeg, ffprobe and Python, installs what is missing, sets up Whisper and HyperFrames, and tests each one in a small check folder.
- Step 1, write your edit rules. A
rules.mdin plain words: shrink any pause over 1 second to a short breath, keep only your last take after "retake", drop false starts and fillers. Anames.mdlists the names your transcriber gets wrong. - Step 2, a time on every word. Whisper transcribes the take with a start and end time on every word. Whisper can skip or invent words, so Claude checks every stretch of speech has words.
- Step 3, cut the dead air. Pauses are found from the sound, never clipping a word, and picture and sound are cut at the same frames. Claude tells you how many pauses it cut and the seconds saved.
- Step 4, cut the retakes. Only your last take of each line stays. Claude writes
cuts.md, every cut with its time, the words removed and the reason, then transcribes the result again to check it. - Step 5, fix names and write captions. Every name is spelled the way
names.mdsays, andcaptions.srtgets short lines that break where a phrase ends. Claude lists any word that still looks misheard. - Step 6, set the loudness. The finished file is at -14 LUFS with the true peak at or under -1.5 dBTP, which is where YouTube plays it, and Claude measures the raw take against the result.
Read cuts.md before you move on. It lets you check the whole edit without scrubbing through the video.
Steps 7 to 13: the rest of my pipeline
The clean cut is where most tutorials stop. A long video people finish watching needs more than that. These are the stages my own pipeline runs after the cut:
- Step 7, your brand style, once. Colours, fonts and caption style written down one time, so every video looks like yours.
- Step 8, pacing. A review of the clean transcript that flags slow stretches and proposes cuts. You answer yes or no.
- Step 9, the storyboard. Each section of the video gets a frame plan before anything is designed.
- Step 10, designed frames and music. HyperFrames renders title cards, diagrams and word pop-ups over your recording, with a music bed under it.
- Step 11, checking. A separate check for sync, leftover retakes, loudness, frozen frames and private details on screen.
- Step 12, make it learn. Your answers and hand edits are written back as rules, so the next video needs fewer of them.
- Step 13, finish and ship. Chapters, title, thumbnail, description and the upload.
My last long-form take ran 38 minutes 24 seconds raw and became a 17-minute video, with 107 long pauses and 15 mistakes cut, in about 4.5 hours of machine time. Before this pipeline, one edit took me about a full day.
Start with the free kit
The prompts for steps 0 to 6, exactly as I typed them, are in my free Video Editing Starter Kit. It also has rules.md, names.md with my sample, hints for steps 7 to 13 and the fourteen-step map. Make one folder, drop your take in it as raw.mp4, and paste the Step 0 prompt.
Build the whole pipeline with me
Steps 7 to 13 took me days to work out, and they are where a video goes from clean to watchable. That is what my long-form editing cohort covers.
It is four weekly live sessions, starting Sunday 10 January, with office hours every Thursday. Each week I build one stage live on my own recording, you build it on yours before the next session, and the clinic fixes what broke. By the end you have shipped one long-form video of your own, and the pipeline is yours to run on every video after it. Join the waitlist on Maven.
The cohort assumes you already use Claude Code. New to it? Start with Claude Code Builder, my six-week live cohort that starts from a blank folder. If you mostly want Shorts and Reels, read how I make YouTube Shorts with Claude Code. Not sure which fits? See the whole path.
A note on the numbers
Every number on this page comes from my own takes and my own reports. Claude writes its own scripts each run, so your times and loudness will move a little between runs. Read each step's report, then check the file yourself.